NVIDIA Predicts 570x GPU Performance Boost

Gianna Borgnine writes "NVIDIA is predicting that GPU performance is going to increase a whopping 570-fold in the next six years. According to TG Daily, NVIDIA CEO Jen-Hsun Huang made the prediction at this year's Hot Chips symposium. Huang claimed that while the performance of GPU silicon is heading for a monumental increase in the next six years — making it 570 times faster than the products available today — CPU technology will find itself lagging behind, increasing to a mere 3 times current performance levels. 'Huang also discussed a number of "real-world" GPU applications, including energy exploration, interactive ray tracing and CGI simulations.'"

Comments

anonymous coward predicts first post

[Anonymous Coward]

suck my asshole, homos

Goody!

[Anonymous Coward]

Then we can use our GPUs as our CPUs!

Re:Goody!

[MarkRose]

Yo dawg, I herd you like CPUs, so I put a CPU in your GPU so you can still get work done in vi while emacs is loading!

haha yeah right

[Anonymous Coward]

just like how intel said they'd have 4nm chips? haha, wee dont believe that shit. we know far more than the companies about fabrication. QUEUEUE a 1000 posts about how 560x is unrealistic.

Re:haha yeah right

[4D6963]

Intel said 4 nm for 2022, that's in 13 years. What precisely allows you to doubt that claims, except maybe the fact that deadlines are often missed? Let me rephrase that, what allows you to think that it'll be reached much later than anything else?

Also, queue a dozen+ posts explaining to the armchair pundits how 560x is possible.

Re:haha yeah right

[Anonymous Coward]

WHOOPAWW. It's the new woosh.

Re:haha yeah right

[Martin Blank]

The IEEE figures that semiconductor tech will be at the 11nm level around 2022. Intel and Nvidia both claim that they'll be significantly further along the path than the IEEE's roadmap. Maybe they're right, and I hope they are, but there are some very significant problems that appear as the process shrinks to that level.

Re:haha yeah right

[4D6963]

Undoubtedly, however this is too early to call, for all we know they might just make it.

Re:haha yeah right

[PIBM]

Stupid I know, but I would have had more confidence in a 500x increase, just because there's less significant digits and a wider error margin.

Re:haha yeah right

[vertinox]

Intel said 4 nm for 2022, that's in 13 years. What precisely allows you to doubt that claims, except maybe the fact that deadlines are often missed? Let me rephrase that, what allows you to think that it'll be reached much later than anything else?

I'm dunno. Most CEOs don't make claims unless their business plan includes said claims else they look like a fool at the next shareholder meeting. That doesn't stop them from making claims that don't come through.

Remember Steve Jobs saying they would break the 3.0ghz with the IBM by the next WWDC... And then they didn't... Coincidentally they dropped IBM shortly after and went with Intel.

Anyways... Intel seriously uses Moore's Law as their road map so its a self predicting prophecy.

Also, queue a dozen+ posts explaining to the armchair pundits how 560x is possible.

Simple. Move the goal posts.

Re:haha yeah right

[RightSaidFred99]

Yeah, and Intel's kind of new in the field so I doubt they know what they're talking about.

Re:haha yeah right

[RightSaidFred99]

Anyways... Intel seriously uses Moore's Law as their road map so its a self predicting prophecy.

No, they don't. It's descriptive, not something that ties your hands or, conversely, guarantees anything.

Re:haha yeah right

[Anonymous Coward]

I'd be more inclined to believe they might get their current GPUs working properly within 6 years.
They cant do bumps right and have had real problems with 45nm GPUs.

But a 500x increase sounds line an estimate pulled out of his head/ass/PR department but 570x is more like some geek calculated figure.
Besides, he meant 5.70x faster but the Nvidia GPU burned out and miscalculated.

Re:haha yeah right

[Anonymous Coward]

560x performance of current GPUs is not likely to occur in the next six years for the following reasons:

It does not fit with the plainly understood pace of improvement of performance of current generation GPUs. It just sounds absurd to most people and to expect them to not think its absurd you need to put up or shut up. Extraordinary claims demand extraordinary evidence. TFA didn't even try and deliver.

ATI and Nvidia routinly tape out marginal improvements and play games with core counts/SLI to obtain cost effective performance scaling rather than making significant architectural R&D heavy changes to obtain improvements.

There are no sigificant underlying fabrication or technological capabilities on the horizion with the capability of being mass produced on the time scale in question that would account for the improvement. They must be at least known or immediatly forseeable in available liteature today if they are to have any hope of production in the next six years. Technology does *NOT* move that fast.

Specifically the power requirements alone of a high-end modern GPU are in excess of 200 watts. There is no known technology in existance that can produce 560x performance figures while keeping the power budget in line with anything resembling reality. In fact such a technology if it were to exist in six years would exceed the energy effeciency of the human brain.

Re:haha yeah right

[Anonymous Coward]

you are wrong.

Re:haha yeah right

[4D6963]

Agreed with the other poster, you're wrong, they've been actively pursuing the goal of fulfilling the prophecy. They make it a primary goal to increase the number of transistors by all means. If it wasn't for the law they wouldn't have done things like they have. The law itself won't fulfil itself that simply, you need to throw billions at it to keep up.

Re:haha yeah right

[Martin Blank]

I'm not shorting Intel's capabilities, but the IEEE has some solid people in it, too -- many of whom work at Intel -- and they're very capable of recognizing the potential problems with process shrinks. The issues that come about at the sizes they're discussing involve quantum tunneling effects that would (as I understand it) interfere in accurate computing. There is also doubt that transistors can be made to work at all at sizes below 16nm because the mechanisms that might deal with quantum tunneling may bring about other deleterious effects that may be even more difficult to solve.

I'm not saying that it's impossible, or that Intel is too optimistic. They know a lot more about it than I do. But these kinds of things do slip, and it's hard to predict advances of this sort so many years down the road.

Re:haha yeah right

[fractoid]

Intel said 4 nm for 2022, that's in 13 years. What precisely allows you to doubt that claims, except maybe the fact that deadlines are often missed? Let me rephrase that, what allows you to think that it'll be reached much later than anything else?

Don't you start getting funky quantum effects happening around that size? Not saying it's not possible to build a CPU, but I thought there was a lot of talk about CPUs basically stagnating in clock speed and component density due to quantum tunneling or somesuch. The reason GPUs are outstripping CPUs these days is that everything they do is inherently massively, massively parallelizable. I think in the future, (and this is a fairly safe bet because it's already happening) computers will have the CPU basically as a coordinator and the grunt work will be done by massively parallel programmable vector processing units.

Re:haha yeah right

[w0mprat]

I can't think how many press releases i've seen about advancements paving the way towards single-molecule switches. Is there really any grounds for anyone doubting moores law just not stopping for at least the next 15 years?

Re:haha yeah right

[Anonymous Coward]

Bear in mind that Intel have repeatedly beaten the IEEE predictions for process shrink timelines over the last 6 years or so.
So, if anyone can get to 4nm by 2022, Intel will.

I do share your skepticism over the lack of accommodation for the presumably increased difficulty of making transistors work once we get into the sub-16nm regime, but we should bear in mind the historical trackrecord involved.

Re:haha yeah right

[svirre]

We are already seeing unreliability issues appearing. As trasistors get down to sizes so small they actually only hold a handful of electrons on their gate this problem will magnify significantly. This does not however mean these device can't be used, just that you have to compensate.

We are allready now seeing self correcting or error tolerant circuits that either do their calculations with some redundancy and correct, or detect errors and replay the calculation.

Re:haha yeah right

[exploder]

Are you suggesting there are other, better, possibly cheaper ways to exponentially increase processing power over time, and that Intel has chosen to stick with increasing the transistor number solely to avoid falling under Moore's curve?

That sounds stupid. So I'm probably misunderstanding what you're saying.

Re:haha yeah right

[HiThere]

I have a feeling that there are people at Intell and Nvidia that know more about those problems than the IEEE prognosticators. This doesn't mean they won't hit unexpected roadblocks, of course.

OTOH, when a company makes a public announcement, one must always wonder whether the tech people even had any input into what marketing wanted to say.

Re:haha yeah right

[HiThere]

He didn't say that there were other, better methods. He said they used the law to set their goals...and they spent lots of extra money to meet their goals on target.

It does sound silly. Also bureaucratic. Also like an ad campaign. So it's probably true.

N.B.: How much worse off would Intel have been if it had increased it's speed at half the pace? It's hard to be sure. They might not be ANY worse off. They'd have more competition from AMD, though, and IBM might not have decided to drop out of the race, so they might be significantly worse off. But they would have saved a LOT of money. So they might be better off.

Just consider. By pushing the pace now they may keep China from expanding their fab plants. They might keep Japan in specialized markets. That could be worth more than the cost of an extra generation of fab plants.

Re:haha yeah right

[cheesybagel]

Intel is usually fairly optimistic about long term predictions. They bet on EUV being available early and isn't for example.

Re:haha yeah right

[RightSaidFred99]

You guys are...special. Intel spends all that money because it reduces the cost, to them, of producing chips. AMD is reasonably close to Intel, are they spending all that money too? If Intel slacked off their pace AMD would pull even or ahead like they did in the P4 days.

Intel moves as fast as they can based on business decisions about the cost/benefit of spending money on R&D and fabs. Not based on Moore's "law".

If Intel could move faster and it would make them more money, they would speed ahead of Moore's. If they determined they didn't need to, they would fall behind it.

More importantly, Moore's law isn't accurate for shit. Do the math. It's a good high level description and if you pick specific periods of time. It has also been changed to suit people's needs...12 months, 24 months, 18 months. It's a load of shit as anything other than a very rough estimate.

Re:haha yeah right

[Anonymous Coward]

wait... do we have 4 nm? or do the terminators?

Re:haha yeah right

[renoX]

>What precisely allows you to doubt that claims, except maybe the fact that deadlines are often missed?

I remember seing Intel's roadmap with P4 at frequency which were never produced..

Predictions of the future

[mollog]

I see a few tags that cast doubt on the prediction. Why? I'll bet there were skeptics of Moore's Law when that became widely disseminated.

What troubles me is that this sort of cell GPU is not more widely used in everyday applications. We who program for a living are feeling like we have been engaging in 'self stimulation' for years and wish there were some new target platform/market that we could so some interesting work in.

Re:Predictions of the future

[Anonymous Coward]

Panties Stink!
They really, really stink!
Sometimes they're red, sometimes they're green,
Sometimes they're white or black or pink
Sometimes they're satin, sometimes they're lace
Sometimes they're cotton and soak up stains
But at the end of the day, it really makes you think
Wooooooo-wheeeee! Panties stink!

Sometimes they're on the bathroom floor
Your girlfriend- what a whore!
Sometimes they're warm and wet and raw
From beneath the skirt of your mother-in-law
Brownish stains from daily wear
A gusset full of pubic hair
Just make sure your nose is ready
For the tang of a sweat-soaked wedgie
In your hand a pair of drawers
With a funky feminine discharge
Give your nose a rest, fix yourself a drink
cause wooooooo-wheeeeeee! panties stink!

Re:Predictions of the future

[TheRealMindChild]

Well, it comes down to simple math. For the performance to get to 570-fold more than what it is now, in the same style package, either:

1. The GPU has to become 570-fold more efficient
2. The GPU has to become ~570-fold smaller so they can fit 570 of the things onto a card

Both seem highly unlikely.

Re:Predictions of the future

[eln]

I don't doubt the prediction at all, I just have concerns about the vat of liquid nitrogen I'm going to have to immerse my computer in to keep that thing from overheating, and the power substation I'm going to need to build in my backyard to power it.

Re:Predictions of the future

[javaman235]

Its easy to get a 570x increase with parallel cores. You will just have a GPU that is 570 times bigger, costs 570 times more and consumes 570 times more energy. As far as any kind of real break through though, I'm not seeing it from the information at hand.

There is something worthy of note in all this though, which is that the new way of doing business is through massive parallelism. We've all known this was coming for a long time, but its officially here.

Re:Predictions of the future

[Anonymous Coward]

The claim is further undermined by the fact that nvidia's last two "generations" of GPUs were just rebrands and die shrinks.

Re:Predictions of the future

[Anonymous Coward]

The prediction is complete nonsense. It assumes that CPU processors only get 20% faster per year (compounded). That would only be true if they did not add more cores to the CPU. And finally GPUs are hitting the same thermal/power leakage wall that CPUs hit several years ago - they will at best get faster in lock step with CPUs.

A GPU is not a general purpose processor, as is a CPU. It is only good at performing a large number of repetitive single precision (32 bit) floating point calculations without branching. Double precision (64 bit) calculations - double in C speak - is 4 times slower than single precision on a GPU. And the second you have an "if" in GPU code, everything grinds to a halt. Conditions effectively break the GPU SIMD (single instruction multiple data) model and bring the pipeline to a halt.

Re:Predictions of the future

[LoudMusic]

I don't doubt the prediction at all, I just have concerns about the vat of liquid nitrogen I'm going to have to immerse my computer in to keep that thing from overheating, and the power substation I'm going to need to build in my backyard to power it.

But GPUs today are somewhat more than 570x more powerful than they were several years ago and we haven't had to submerge them in a vat of liquid nitrogen yet, so what makes you think that's going to be the case in the next 570x power increase? (whenever that happens ...)

Re:Predictions of the future

[LoudMusic]

Well, it comes down to simple math. For the performance to get to 570-fold more than what it is now, in the same style package, either:

1. The GPU has to become 570-fold more efficient
2. The GPU has to become ~570-fold smaller so they can fit 570 of the things onto a card

Both seem highly unlikely.

You don't feel it could be a combination of both? Kind of like they did with multi-core CPUs? Make a single unit more powerful, then use more units ... wow!

There is more than one way to skin a cat.

Re:Predictions of the future

[eln]

Maybe not, but they do require a lot more cooling and power than they did before.

Re:Predictions of the future

[fuzzyfuzzyfungus]

Even more curious, he claims that GPUs will see 570x improvement, with CPUs only getting 3x.

One wonders what night-miraculous improvement in process, packaging, logic design, etc. will improve GPUs by hundreds of times, while somehow being virtually useless for CPUs...

Re:Predictions of the future

[Anonymous Coward]

Moore didn't predict that processors would be 736x more powerful in 240 months.

Re:Predictions of the future

[Anonymous Coward]

Even generously assuming they'd achieve an 8x die shrink, that'd need to be producing chips with a 41000mm^2 die area. (Their current chips are already the biggest at 576mm^2.)

Re:Predictions of the future

[oldhack]

That's where your backyard dirt patch growing watermelon comes handy.

Re:Predictions of the future

[negRo_slim]

Perhaps for the top end models that holds true, but as the market for roll your own HTPC's has shown (at least in terms of cooling), there are plenty of passive heat sink options available.

Re:Predictions of the future

[Anonymous Coward]

Because using more units is cheating? C'mon, someone can probably find a way to glue 570 GPUs together now. That doesn't mean that GPUs saw a 570-fold increase in performance in the time it took me to come up with my brilliant plan.

Re:Predictions of the future

[Twinbee]

Or 3d-erize the chip?

Re:Predictions of the future

[BikeHelmet]

Or... not.

Currently CPUs and GPUs are stamped together. Basically, they take a bunch of pre-made blocks of transistors(millions of blocks, billions of transistors in a GPU), and etch those into the silicon, and out comes a working GPU.

It's easy - relatively speaking - and doesn't require a huge amount of redesign between generations. When you get a certain combination working, you improve (shrink) your nanometre process and add more blocks.

However, compiler technology has advanced a lot recently, and with the vast amounts of processing power now available, it should be simpler getting more complex blocks fully utilized. A vastly more complex block, with interconnects to many other blocks, could perform better at a swath of different tasks. This is evident when comparing the performance hit from Anti-Aliasing. Previously even 2xAA had a huge performance hit, but nVidia altered their designs, and now Multisampling AA is basically free.

I recall seeing an article about a new kind of shadowing that was going to be used in DX11 games. The card used for the review got almost 200fps at high settings - with AA enabled that dropped to about 60fps, and with the new shadowing enabled, it dropped to about 20fps. It appears the hardware needs a redesign to be more optimized for whatever algorithm it uses!

Two other factors you're forgetting...

1) 3D CPU/GPU designs are coming slowly, where the transistors aren't just on a 2D plane... that would allow vastly denser CPUs and GPUs. If a processor had minimal leakage, and low power consumption, 500x more transistors wouldn't be a stretch.

2) Performance claims are merely claims. Intel claims a quad-core gives 4x more performance, but in many cases it's slower than a faster dual-core.

570x faster for every game? Doubtful. 570x faster at the most advanced rendering techniques being designed today, with AA and other memory-bandwidth hammering features ramped to the max? Might be accurate. A high end GPU from 6 years ago probably won't get 1fps on a modern game, so this estimate might even be low.

A claim of 250x the framerate in Crysis, with everything ramped to the absolute maximum, might be even accurate.

But general performance claims are almost never true.

Re:Predictions of the future

[volsung]

The GeForce 9 series was a rebrand/die shrink of GeForce 8, but the GTX 200 series has some major improvements under the hood:

* Vastly smarter memory controller including better batching of reads, and the ability to map host memory into the GPU memory space
* Double the number of registers
* Hardware double precision support (not as fast as single, but way faster than emulating it)

These sorts of things probably don't matter to people playing games, but they are huge wins for people doing GPU computing. The GTX 200 series has also seen a minor die shrink during the generation, so I don't know if the next generation will be more of a die shrink or actually include improved performance. (Hopefully the latter to keep up with Larrabee.)

Re:Predictions of the future

[Anonymous Coward]

Simple. GPUs are already massively parallell in a way that is actually usable. While the gigahertz-race has pretty much stopped, the transistor-race is still on, but getting the most out of a multi core CPU for a single application is non-trivial. GPUs however are a completely different ballgame, where the performance of the card pretty much scales with the number of shader cores.

Re:Predictions of the future

[Anonymous Coward]

Doing multiple layers either via lamination or deposition would make sense. But then there's this problem: How do you get the heat out of it? Those things aren't exactly running cool as they are now.

But then again, maybe they figured something out that we don't know.

Re:Predictions of the future

[Anonymous Coward]

I stand corrected, but the lower end models are still rebranded ninth series.

Re:Predictions of the future

[Twinbee]

Perhaps have them running at lower clock speeds. Slightly slower clock speeds means much cooler chips afaik.

Re:Predictions of the future

[macshit]

I see a few tags that cast doubt on the prediction. Why? I'll bet there were skeptics of Moore's Law when that became widely disseminated.

It may have something to do with the fact that Jen-Hsun Huang is a notorious blowhard, whose foot spends far more time in his mouth than on the ground.

Combine that with the extremity of the claim (570-fold over 6 years is a huge rate of improvement), and the tags don't seem so surprising.

Re:Predictions of the future

[Loomismeister]

Why do people keep saying 570-fold like it means the same thing as 570 times greater. 570 fold would be equivalent to 2^570ith, and thats not what Nvidia predicted.

Re:Predictions of the future

[C18H27NO3]

We'll obviously need a Turbo button on it set in the off position until such time that the CPUs catch up.

Re:Predictions of the future

[Ant P.]

1. The GPU has to become 570-fold more efficient
      2. The GPU has to become ~570-fold smaller so they can fit 570 of the things onto a card

Both seem highly unlikely.

If graphics card development in the last 10 years is anything to go by, nVidia's plan is that the GPU will become 570 times larger, draw 570 times more power and the fan will spin 570 times faster

Re:Predictions of the future

[Lord Bitman]

From reading slashdot, I know there are several technologies which have been "a couple of years away" for a while now which could (if people bothered with the expense) turn the most common problem in computer heat dissipation into "how do I prevent it from getting too cold and forming condensation on everything?"

Re:Predictions of the future

[Poorcku]

If I were the NVIDIA chairman I would make the GPU 570 fold bigger and smack you in the face with it you all knowing bastard ! :)

Re:Predictions of the future

[Anonymous Coward]

You don't feel it could be a combination of both? Kind of like they did with multi-core CPUs? Make a single unit more powerful, then use more units ... wow!

So, does it come with a 570-fold price increase?

Re:Predictions of the future

[Tomfrh]

I know where you're coming from, but "570-fold" does mean 570 times greater. It doesn't mean 570 doublings.

The analogy is zig-zag folds, not doubling folds.

Re:Predictions of the future

[Anonymous Coward]

And those budget cards with heatsinks (8600, 9500 and below) have a hard time competing with higher end series 7.
I'm still waiting for something worth replacing my 7600 with, that doesn't need additional power connectors. Seems I either have to buy an oven or enjoy a very modest performance increase.

Re:Predictions of the future

[SocratesJedi]

Indeed, if both were improved at the same rate you would only need a SQRT(570x) ~= 24x improvement in both technologies to get a combined increase of 570x (also ignoring Amdahl's law). Nothwithstanding the (possibly serious) limitations of Moore's law there are four 18-month periods in 6 years predicting a 16x reduction in chip size over that time period. I have no experience in the increase in efficiency increases, so I won't comment on that, but to my mind the combination at least doesn't seem horrifically farfetched.

Regardless though, you are right to be skeptical of a long range prediction like that.

Re:Predictions of the future

[melf-san]

Maybe the high-end ones, but the low-end GPUs are mostly passively cooled and still much more powerful than old GPUs.

Re:Predictions of the future

[tyrione]

Well, it comes down to simple math. For the performance to get to 570-fold more than what it is now, in the same style package, either:

1. The GPU has to become 570-fold more efficient
2. The GPU has to become ~570-fold smaller so they can fit 570 of the things onto a card

Both seem highly unlikely.

It's not a linear relationship.

Re:Predictions of the future

[geekoid]

Moore's law is ending. The fab issues at the scale can be costly. Moore's law is about cost, not speed.

Re:Predictions of the future

[TheRaven64]

It's also worth noting that a 570x improvement over six years means doubling performance ever 9 months. Just from die shrinkages from process improvements, you can fit twice as many pipelines on the same amount of silicon in a slightly longer period than this. It isn't hard to imagine that improved designs will fill up the gap.

Re:Predictions of the future

[jonbryce]

People don't seem to be buying these faster CPUs anymore, as there is a recession on, and the existing ones are fast enough. The machines I see in the shops still have pretty much the same Intel models they had two years ago. I don't think that has ever happened before.

I think the main change we will see in the next three years or so is faster Atoms rather than blazingly fast Core i9s or whatever Intel chooses to call them.

Re:Predictions of the future

[raftpeople]

"GPUs however are a completely different ballgame, where the performance of the card pretty much scales with the number of shader cores."

But only to the degree that your problem maps to that level of parallelization. There are many problems that do not perform well on the GPU.

Re:Predictions of the future

[Anonymous Coward]

in many cases it's slower than a faster dual-core

Kinda like how in many cases car is slower than a faster van?

Re:Predictions of the future

[Anonymous Coward]

But that's why there's a specific problem right in the name.

Re:Predictions of the future

[Starayo]

The day when my computer also doubles as a mini fridge will be awesome indeed.

Re:Predictions of the future

[Lord Ender]

Either

• your post contains a false dichotomy, or
• Angelina Jolie is giving me a blowjob right now

Neither seem highly unlikely.

Re:Predictions of the future

[Connie_Lingus]

more like a mini-hotplate

Re:Predictions of the future

[Antique Geekmeister]

No, no, it runs on _cold_ fusion. That's only 2 years away, right? Peter Hagelstein has been writing that kind of letter, article, and review of scientific papers for over a decade now. But they keep accidentally publishing his work in the "science fact" section, rather than the "short fiction" categories.

Re:Predictions of the future

[Anonymous Coward]

Well GPU does suggest Graphics doesn't it? perhaps Duke Nukem Forever at 800 FPS?

Re:Predictions of the future

[shutdown -p now]

Performance claims are merely claims. Intel claims a quad-core gives 4x more performance, but in many cases it's slower than a faster dual-core.

570x faster for every game? Doubtful.

Why? Rendering is specifically a task that is easily parallelizable, and thus using more cores will scale well even in existing games (unlike what happened with quad-core on desktops) - just split the output area in chunks, and render in parallel...

Re:Predictions of the future

[shutdown -p now]

If graphics card development in the last 10 years is anything to go by, nVidia's plan is that the GPU will become 570 times larger, draw 570 times more power and the fan will spin 570 times faster

At that point, it would effectively become a helicopter, no?

Re:Predictions of the future

[fractoid]

Well, it comes down to simple math. For the performance to get to 570-fold more than what it is now, in the same style package, either:

1. The GPU has to become 570-fold more efficient
2. The GPU has to become ~570-fold smaller so they can fit 570 of the things onto a card

Both seem highly unlikely.

Either in isolation seems highly unlikely, true. But a 23x improvement in efficiency coupled with a 23x improvement in the number of cores on a GPU? In six years? That sounds a bit more feasible.

If there were, just coincidentally, to be some theoretical progression rate that involved doubling every 18 months, you'll notice that after ~6.87 years you're at ~570. Which is close enough to their estimate that it could be how they arrived at this figure.

Re:Predictions of the future

[mgblst]

Why? I'll bet there were skeptics of Moore's Law when that became widely disseminated.

Ok, so since there were skeptics at Moore's Law in the beginning, then anything goes from now on? Somehow you use this as a justification to never listen to skeptics anymore. That is just idiotic. You know there will people skeptical about the P4 as well, and Itanium.

Re:Predictions of the future

[russotto]

What troubles me is that this sort of cell GPU is not more widely used in everyday applications.

I've looked at it for a few things (encryption, mostly) and there always seem to be a few "gotchas" which end up becoming a bottleneck. It's not clear if this is due to the programming model or the chip, however... access to the accursed carry flag or equivalent keeps coming up, and I know that one is in the programming model.

Re:Predictions of the future

[fractoid]

Embed a network of microscopic cooling tubes, with coolant flowing through via either capillary action or forced by a pump.

Re:Predictions of the future

[fractoid]

Wouldn't that be more like "in many cases the car is slower than two faster cars"?

Re:Predictions of the future

[LBt1st]

Yes a helicopter.. with Really Sweet Graphics!!

Re:Predictions of the future

[Klintus Fang]

You do need to keep in mind one of the primary reasons CPUs are designed the way that GPUs are designed. GPUs are very good at doing massively parallel repetitive computations like what you need to do when processing an image for a display or like some small segments of the HPC market regularly do. But GPU performance grinds to a halt as soon as your code needs to take a branch.

Unless you write the kind of code in which if statements are nearly non-existent, you are likely to find programming for a GPU-style architecture to be unhelpful.

Re:Predictions of the future

[Klintus Fang]

You need to keep in mind the reason that CPUs are not designed the way that GPUs are designed. GPUs are very good at easily parallelizable repetitive computations. GPU performance grinds to a snails pace as soon as you have to take a branch.

Unless you are writing the kind of code in which if statements are nearly non-existent, you are not likely to find programming for a GPU-style pipeline to be very palatable.

Re:Predictions of the future

[wall0159]

A 570-fold increase over 6 years is less than a 3-fold increase per year. This is fast (towards double Moore's Law), but then it's possible that an Nvidia engineer knows something about GPU research that we don't...

I'd say this is news, and it's remarkable, but it's not totally implausible.

Re:Predictions of the future

[Anonymous Coward]

I'm pretty sure it's the first one. Even Angelina Jolie can't be in two places at the same time.

Re:Predictions of the future

[BikeHelmet]

Why? Rendering is specifically a task that is easily parallelizable, and thus using more cores will scale well even in existing games (unlike what happened with quad-core on desktops) - just split the output area in chunks, and render in parallel...

I think I explained my position fairly well. But if, by freak chance, we end up playing Crysis in 6 years at 34200fps, I will happily eat crow.

Re:Predictions of the future

[Chousuke]

Imagine the flight simulators!

Re:Predictions of the future

[91degrees]

And the second you have an "if" in GPU code, everything grinds to a halt. Conditions effectively break the GPU SIMD (single instruction multiple data) model and bring the pipeline to a halt.

This isn't totally accurate. Generally conditions are handled by conditional writeback. You simply ignore the result if the test fails. You effectively have to perform both branches of a condition so there's a performance hit over a CPU there but "if(x 0){x = -x)" isn't going to hurt your performance.

Re:Predictions of the future

[Starayo]

It already does that. My room is warm in winter without anything heating it besides my computer.

Re:Predictions of the future

[Briareos]

If graphics card development in the last 10 years is anything to go by, nVidia's plan is that the GPU will become 570 times larger, draw 570 times more power and the fan will spin 570 times faster

At that point, it would effectively become a helicopter, no?

A flaming, almost flat helicopter that makes the horns of Jericho sound like a kazoo?

Could be.

np: SCSI-9 - Together (Easy As Down)

Re:Predictions of the future

[ShakaUVM]

>>Conditions effectively break the GPU SIMD (single instruction multiple data) model and bring the pipeline to a halt.

They're getting better. Hell, IIRC I think that when shaders first came out you couldn't do conditionals at all.

Re:Predictions of the future

[Anonymous Coward]

You do know that the new GPUs have double-precision support, and that people have shown that raytracing (loaded with "if" decision points) works well on a GPU? I'd say the thing that grinds everything to a halt is keeping 10,000 threads busy.

Re:Predictions of the future

[Ost99]

The GTX260 and up are new chips.
GTX250 is a die shrunk 9 series (9800gtx+)

Re:Predictions of the future

[Schnoogs]

I guess because you used big words and acted real confident in your dismissal people modded you informative. Next time read the article and don't spout of about GPU's performing branching as if that's even on the table and thus somehow dismisses this guys prediction.

Re:Predictions of the future

[denoir]

Perhaps not all that unreasonable. The Geforce 3 was launched in 2001 and was the first real GPU that could do floating point operations.

2001: Geforce 3, 2.6 MFLOPs
2009: Geforce GX295: 1.78 GFLOPs

That's a factor of 680 in six years. Ok, the 295 is a dual GPU solution, but Nvidia is set to release the nv300 series this year that will have twice the number of cores. So he's actually predicting that the increase in computing power will be smaller than in the previous seven year period.

Re:Predictions of the future

[Anonymous Coward]

(1+5.70)^(1/6)-1 = 37% per year

Hasn't this rate of growth been the norm for the last 15 years?

Re:Predictions of the future

[Anonymous Coward]

Except that rumours suggest that the next series of nVidia processors will be MIMD, rather than SIMD, due to pressure from Intel's Larrabee.

Re:Predictions of the future

[ceoyoyo]

"It assumes that CPU processors only get 20% faster per year (compounded). That would only be true if they did not add more cores to the CPU."

"It is only good at performing a large number of repetitive single precision (32 bit) floating point calculations without branching."

If we wanted a 64-bit GPU it would be easy enough to make. GPUs used to do weird mixes of integer and floating point math until the manufacturers made an effort to guarantee 32-bit precision throughout. That leaves the branching part of your statement, which is the same for CPUs with multiple cores. A modern general purpose GPU (that is, one that CAN branch) is pretty similar to a many-cores CPU in those terms.

The two are converging. CPUs are getting more general purpose and CPUs are getting more parallel.

Re:Predictions of the future

[mdwh2]

One presumes that predictions like these are "for a given cost", or at the least, on a single chip - just as is the case of Moore's law. No one claims that Moore's law was beaten when they started putting hundreds of CPUs together in a supercomputer.

I don't see how it's cheating if you can do it on a single chip, or at the same price - modern day GPUs already have tens or hundreds of cores, and CPUs have multiple cores too. Do you disable all but one of them on your computer, because it's "cheating"?

Re:Predictions of the future

[mdwh2]

No, you can't just magically have get both extra efficiency and doing smaller size, out of the same increase.

Moore's law has meant you can get double the number of transisters per 18 months - and it's this which has given the extra efficiency we've been seeing. There isn't a second "increase of efficiency" trend! You may note that chips generally haven't actually been getting smaller, because people would generally prefer the increased performance. You either choose between same number of transisters in a smaller area, or twice as many transisters in the same area. You can't count it twice!

In fact, it's only been in recent years that increasing transister counts as been used to add multiple cores, but that comes at the expense of increased efficiency of any given core. Don't think that just because we've been going to dual and quad cores, that somehow means we're increasing at a faster rate than before...

Re:Predictions of the future

[mdwh2]

Indeed, and given they're pushing GPUs for GPGPU, it seems rather misleading to claim it will outdo CPUs. If an algorithm is suitable for GPGPU, then it should also benefit from adding multiple CPU cores. If an algorithm is hard to parallelise on a CPU, I don't see how a GPU will help.

Re:Predictions of the future

[mdwh2]

Nope, I make that just under 24x improvement.

As I said in my reply to this poster, you can't just arbitrarily claim there are two such trends!

I mean, why stop there? "CPU/GPUs can get faster by either having more transisters, making them more efficient, or making the algorithms more efficient. I'm going to claim we have a 2x increase every 18 months, independently for all three of these, therefore in 6 years we'll have an increase of 4096 times!!!"

It's nonsense - there's only a 2x increase for number of transisters, and it's that increase which is used for "more efficiency". And increase in efficiency of algorithms isn't anywhere near 2x every 18 months. But if you're allowed to make stuff up, I guess anything's possible!

Which is close enough to their estimate that it could be how they arrived at this figure.

I seriously hope not.

Re:Predictions of the future

[mdwh2]

Yes, I'll look forward to my NVIDIA card with 6.7x performance in 6 years' time.

For everyone else hoping for 570x improvement, they'll be needing more like 570^(1/6)-1 = 188% per year. Way ahead of the Moore's Law trend of about 59% per year.

Re:Predictions of the future

[mdwh2]

That would only get you 256x improvement, surely (2^(6*12/9))? I make it more like just under 8 months.

And that "slightly longer period" is more like 18 months, which only leads to 16x improvements in 6 years. If it turns out that the doubling time itself more than halves, then great, but it's quite a change...

Re:Predictions of the future

[mdwh2]

About cost of the individual transistors.

For cost of the fab plants, you may be thinking of Rock's law, although yes, you are right that this does may put an economic limit on Moore's law.

Re:Predictions of the future

[Anonymous Coward]

Single cores were made faster with instruction level parallelism. That requires larger cores and on gpu context it's dangerous since the cores are very small atm.

Re:Predictions of the future

[lsatenstein]

What about opening up to the idea the todays intel architecture is holding back performance. That 64 bit computers in a 256bit world are arcane, and wasteful of electrical energy. With 256 or even 128 bit standard architectures, many more efficient instructions (one step instructions instead of multiple load register instructions) would lead to increased efficiency. Then the GPU could really be an adjunct to a desktop or mobile phone. Thems my 2 cents

Re:Predictions of the future

[SocratesJedi]

Sure you can! It's definitely possible to get efficiency increases unrelated miniaturization. There is significant research into identifying materials that can be cycled at higher frequencies which would give a speedup unrelated to simply putting more transistors in the same space. Alternatively, you could optimize the architecture of the microprocessors so as to require less clock cycles per opcode which would translate into an efficiency increase independent of either transistor size or clock frequency.

Re:Predictions of the future

[Anonymous Coward]

Actually. You're wrong.

Your basic premise; That you get half the die-size OR double the transisters per area is correct, however you're missing one simple fact. At the sizes being discussed, the speed of electricity comes into effect. Halve the distance between transistors and you're getting a -significant- boost in efficiency. Not only does the cost in -time- come down (shorter trip is quicker), there's also the fact that each trip 'spills' less power into it's neighbours. This is one of the main reasons why it used to take -200.C temperatures to break the 1GHz mark, and these days my -phone- is coming ever closer to breaking that without burning me.

Re:Predictions of the future

[mdwh2]

Sure it's hypothetically possible, but you need to show evidence rather than assuming it.

Have CPUs/GPUs been increasing in speed by 4x every 18 months rather than 2x? If not, what will change for it to happen in future? (In fact, for 570x in 6 years, you need a doubling of speed every less than 8 months...)

Alternatively, you could optimize the architecture of the microprocessors so as to require less clock cycles per opcode which would translate into an efficiency increase independent of either transistor size or clock frequency.

Less clock cycles per instruction has been a trend, but I believe that this has just been one way of using "extra" transisters. Indeed, as I say, using extra transisters for more cores is only a recent trend, AIUI because getting more speed from a single core using extra transisters is becoming harder. But no one's been able to magic greater than 2x performance in speed from 2x the number of transisters, at least not over any kind of period - do you have any counter-examples?

Re:Predictions of the future

[mdwh2]

So this is why CPUs have been quadrupling in speed every 18 months, whilst the number of transisters double.

Oh wait, they haven't.

This is my point - you can make wild assumptions and claims about how you think it should be possible to gain extra kinds of performance, but unless you post actual evidence, there is no way you can assume you can get a 2x factor out of each one independently (as opposed to smaller speed ups which may add up to, at most, a 2x factor in total).

Re:Predictions of the future

[Schnoogs]

HAHAHAHAHAHH!!

I got modded troll for this!!!

HAHAHAHAHH!!!

This place continues to be a joke!!!! Thanks for proving my point! You mod the guy with the baseless claim as informative but then mod me a troll for pointing that out. Guess we don't want anyone exposing the silliness that goes on here. We don't want anyone challenging the status quo.

What a joke this place is.

Re:Predictions of the future

[Schnoogs]

Go figure!! Modded troll again! MS IS TEH EVIL!!!! Informative?

In other news...

[Hadlock]

In other news, ATI is selling their 4870 series cards for $130 on newegg, which are twice as fast as an Nvidia 9800GTS which is the same price (at least on Left 4 Dead, Call of Duty, and any other game that matters). ATI is blowing Nvidia out of the water in terms of performance per dollar and will continue to do so through at least the middle of next year. See here:

http://www.tomshardware.com/charts/gaming-graphics-cards-charts-2009-high-quality/benchmarks,62.html

Yeah, I'd be making outrageous statements too if I were Nvidia.

Re:In other news...

[Hadlock]

Here's the L4D comparo, sorry for the wrong link:

http://www.tomshardware.com/charts/gaming-graphics-cards-charts-2009-high-quality/Left4Dead,1455.html

The 9800GT and 8800GT are both in the 40-60fps while the 4870 (single processor) is in the 106fps range. It's a pretty staggering difference.

Re:In other news...

[Anonymous Coward]

I don't understand what they are measuring?

I use it to play videos in all sort of ways, and nvidia is there even for older nvidia cards using directshow etc, atleast 2-3x faster than the fastest ATI. I would never recommend an ATI to my customers, the ati- drivers are even more buggy than NVIDIA...

Re:In other news...

[JustNiz]

Yeah and the 280 is in the 120fps range. Whats your point?

Re:In other news...

[NervousNerd]

The GTX 280 goes for $240 on NewEgg. Around a 15 FPS difference for $100.

Re:In other news...

[Spatial]

Troll mod? No, this is mostly true.

While his example is wrong (Nvidia's competitor to the HD4870 is the GTX 260 c216), AMD do have better value for money on their side. The HD4870 is evenly matched but a good bit cheaper.

The situation is similar in the CPU domain. The Phenom IIs are slightly slower per-clock than the Core 2s they compete with, but are considerably cheaper.

Re:In other news...

[clarkn0va]

Agreed. My primary use for the nvidia gpu is watching HD. Let's do some math.

1080 * sqrt570 = 25784

I like. Considering even the most basic of today's gpus, the ion and tegra, for example, are capable of 1080p, Mr Nvidia is predicting that my handheld 6 years hence will be able to smoothly decode mkvs and output them real-time to my new UltraMegaFullHD(TM) 25784p tv? Bring on the future!

Re:In other news...

[TeXMaster]

In other news, ATI is selling their 4870 series cards for $130 on newegg, which are twice as fast as an Nvidia 9800GTS which is the same price (at least on Left 4 Dead, Call of Duty, and any other game that matters). ATI is blowing Nvidia out of the water in terms of performance per dollar and will continue to do so through at least the middle of next year. See here:

http://www.tomshardware.com/charts/gaming-graphics-cards-charts-2009-high-quality/benchmarks,62.html

Yeah, I'd be making outrageous statements too if I were Nvidia.

Even when it comes to GPGPU (General Purpose computing on the GPU), ATI's hardware is much better than NVIDIA's. However, the programming interfaces for ATI suck big times, whereas NVIDIA's CUDA is much more comfortable to code for, and it has an extensive range of documentation and examples that provide developers with all they need to improve their NVIDIA GPGPU programming. It also has much more aggressive marketing.

As a sad result, NVIDIA is often the platform of choice for GPU usage for HPC, despite it having inferior hardware. And I doubt OpenCL is going to fix this, since it basically standardizes the low-level API, keeping NVIDIA with its superior high-level API.

Re:In other news...

[MrBandersnatch]

Depending on vendor it is now possible to get a 275 less than a 4890 and a 260 for only slightly more than a 4870; at lower prices its very competitive too. My point is that both NV and ATI are on pretty level ground again and the ONLY reason I now choose NV over ATI is because of the superior NV drivers (both Linux and Windows side)...oh and the fact that ATI pulled a fast one on me with their AVIVO performance claims. Shame on you ATI!

Re:In other news...

[0100010001010011]

In addition to VDPAU enabled mplayer, I can actually FIND CUDA enabled apps. There's CUDA enabled md5 crackers, cuda enabled BOINC, Matlab has a CUDA plugin. I'm considering buying CUDA compatible card so I can install it at work just to play with it in Matlab.

Re:In other news...

[Hadlock]

It's almost double the price and provides features you can only appreciate on a 120hz display. The 9800GT and 8800GT are the same price and the ATI card blows it out of the water. If you're buying cards in the $140 range you may or may not have a 120hz display. I don't know anyone who owns one though.

Re:In other news...

[JustNiz]

>> provides features you can only appreciate on a 120hz display

well thats a new one. There's not even slight technical merit to that statement but its certainly demonstrates the amusing creativity of ATi fanbois.

>> The 9800GT and 8800GT are the same price and the ATI card blows it out of the water

I have no argument that you should go with Ati if you're windows only and looking at cheaper-end cards.

Its totally irrelevant to me though as I go for best overall performance, decent drivers, and only consider cards that have drivers that work well with Linux. ATI suck on all counts in my areas of interest.

Re:In other news...

[7-Vodka]

And their linux drivers still SUCK.

Re:In other news...

[Anonymous Coward]

While ATI is currently the better buy in the midrange the price/performance difference isn't nearly as stark as you make it out to be. Comparing the HD4870 to a 9800GT is curious at best. The lowest price for a 9800GT on newegg right now is a full $42 less than an HD4870(both prices after rebate.) Nvidia has no card to directly compete against the HD4870 512mb but the GTX 260 is slightly better than the 1 gig version of the HD4870 and also slightly more expensive. If you go a step lower in performance you have the HD4850 vs the GTS 250(aka 9800GTX+) and they are roughly comparable both in performance and price.

Re:In other news...

[Anonymous Coward]

Difference between a 4870 and a GTX260. GTX 260 performs better at 1080p and higher. 4870 performs out performs at anything less than 1080p.

They're pretty evenly matched performance wise, so pick your card by whats on sale at the time and the native resolution of your monitor. If you're playing at 1280x1024, the 4870 will give you the best performance. 1680x1050 is anyones game. Bigger than that and you should think about the GTX260.

But then you have the 4870 1GB version which out performs the GTX260 even at high resolutions, but also costs more.

Really, the GTX260 edged it's way into my PC because it was on sale for $35 less than the 4870 1GB and I have a high resolution display. If I would've waited a few weeks for ATIs sales there would be a 4870 1GB in my case. Do your research and shop around.

Re:In other news...

[geekoid]

"ATI's hardware is much better than NVIDIA's. "
no. it's not.

Re:In other news...

[raftpeople]

>> provides features you can only appreciate on a 120hz display

I enjoy the following features of my GTX280 (used for calcs not games):
CUDA (I compile C code, throw in a couple of lines of stuff for the GPU and it runs on the GPU, easy)
Hardware optimizes my memory accesses and at times branchy code so the GPU is doing as much work as possible (makes it easy to get good results on the GPU)

Re:In other news...

[Lord Ender]

The 120Hz display bit is here:

http://www.nvidia.com/object/3D_Vision_Overview.html

This requires an NVIDIA graphics card and a 120Hz display. Furthermore, it's damn cool!

Re:In other news...

[fractoid]

WTF Mods. He's just saying that at this price point you can get nearly double the performance from ATI than from nVidia. I love nVidia too, I run a 9800GT, but I'm not going to mod someone troll for pointing out that something else is now faster and cheaper.

Re:In other news...

[MemoryDragon]

Exactly, ATI is the best to go if you never leave the windows world. However even there NVidia has made major inroads with PhysX and CUDA.
I recently dropped my perfectly fine ATI 4850 and bought a inferior NVidia 9600 GT the reason the ATI Linux driver simply after 10 years still where shoddy no bsd drivers.
The lesser speed I could mostly compensate buy buying a card with more ram, and additionally I got CUDA support. While badaboom is not as fast as Turbo 264 HD on my mac it is a welcome addition.
Being able to encode h264 in dual speed is a really nice addition to the non working ATI me too solution ATI provides.

Re:In other news...

[Anonymous Coward]

you can only appreciate on a 120hz display

You mean like on all standard HDTV displays?

Re:In other news...

[Anonymous Coward]

For about $160 you can get a Geforce GTX 260, which is significantly faster than the 4870 and the better bang for the buck.

Re:In other news...

[TeXMaster]

In addition to VDPAU enabled mplayer, I can actually FIND CUDA enabled apps. There's CUDA enabled md5 crackers, cuda enabled BOINC, Matlab has a CUDA plugin. I'm considering buying CUDA compatible card so I can install it at work just to play with it in Matlab.

Precisely. At my current job we're developing with CUDA, too. I just think it's really sad thing ATI missed the 'developer friendly' train despite its superior hardware.

As I said, OpenCL is not going to fill the gap. _Maybe_ an ATI CUDA compiler would, though.

Re:In other news...

[Hadlock]

Toms Hardware says it's nearly identical

http://www.tomshardware.com/reviews/radeon-geforce-comparison,2007-7.html
 
Biggest difference is that the $160 Nvidia card has 120mb less ram for nextgen games with lots of high res textures; whereas a 1GB 4870 is $20 cheaper, and runs at 250 watts, whereas the Nvidia card consumes 360 watts, which is a lot of heat to exhaust out of a midsize case.

Re:In other news...

[TeXMaster]

"ATI's hardware is much better than NVIDIA's. " no. it's not.

"Is too".

Now that we've left the infantile trolling behind, a brief expsition of some of the reasons why ATI > NVIDIA hardware-wise:

• hardware double support was introduced in ATI way before it was on NVIDIA (a couple of generations earlier)
• ATI has actual vector operation support
• ATI has more die dedicated to shading (the part which is used for GPGPU)
• ATI has simpler scheduling (meaning less die dedicated to it, and thus more die dedicated to actual ALUs)
• ATI runs at a lower frequency, and still managed to achieve _actual_ TFLOP performance before NVIDIA, and cheaper

Re:In other news...

[FooBarWidget]

I agree. I recently bought a laptop with an ATI card and the biggest reason why I did that is because I heard they went open source. I was disappointed by the fact that their latest Catalyst driver doesn't work well on Ubuntu 9.04. The one recommended by Ubuntu works but it's VERY slow when restoring a window in Compiz. All in all it feels like a downgrade compared to my Intel integrated graphics card. Sigh. :(

Re:In other news...

[Anonymous Coward]

http://www.nvidia.com/object/3D_Vision_Overview.html

Re:In other news...

[Anonymous Coward]

I'm no guru when it comes to video cards, but I changed my video card to an ATI following advice from multi-media forums - and I run it under Ubuntu 9.04. I have NEVER seen my movies look so damn good before. The picture quality compared to my last video card is stunning.

You may criticise ATI under Linux for its drivers, but my experience has been completely different. This is one of the best PC related purchases i've made in a long time.

For your info, i'm running a Mitsubishi HC900 projector on a 3m x 2m screen, so the video quality is noticeable!
AC

Re:In other news...

[Anonymous Coward]

I blame the distros. I have an ATI 4650 in a system running Slackware 13.0 64-bit, and it was super easy to install the ATI drivers from ATI.

I've used ATI for years under Linux, and never had an issue installing their drivers or using the open source radeon/radeonhd on some of the newer cards.

Re:In other news...

[ceoyoyo]

Damn kids. We had to program in (different) assembly for ATI and NVIDIA and we thought we were lucky! ;)

Re:In other news...

[Anonymous Coward]

Yeah, you're not seeing the open source drivers yet - they didn't release their code as open source, they gave out their specs.

The drivers are being worked on, but it'll be another few months before you're likely to see any useful results. I keep up with this stuff by reading phoronix.org - they have a newsfeed with lots of stuff about graphics driver development. They have the new drivers being put into the NEXT version of the kernel for R500 and below (your card is probably R600 or R700) and they have written enough of the development drivers for R600 to barely have Compiz working (although it might be at least as useful as the Catalyst drivers at this point).

Re:In other news...

[Anonymous Coward]

while AMD graphic chips appear to be top, their driver department is the opposite. i mostly use OpenGL and evenly priced NVidia Cards ALWAYS beat ATI/AMD Cards in terms of performance and reliability (not only crashes, but also driver updates deploying a whole range of bugs)

Additionally AMD's Linux and Solaris support is abysmal. I don't see, why i should let myself being locked to one OS supplier by the supplier of a hardware component. That's entirely against the idea of Personal Computers.

Orrr...

[Anonymous Coward]

CPUs will start looking more like GPUs... with multiple stream processing units... kind of like the Cell processor...

FOUL!!

[Runaway1956]

No fair! That obvious response was mine! I want it back! But, the question remains: should we use our CPU's as GPU's?

But how?

[Anonymous Coward]

I read the article, but I don't see any explanation of how exactly that performance increase will come about. Nor is there any explanation of why GPUs will see the increase but CPUs will not. Anyone have a better article on the matter?

Re:But how?

[Anonymous Coward]

Parallelism... and creative accounting

A GPU with 48 Stream processing units count as one GPU...

A Die with 48 Cores counts as 48 CPUs.

Re:But how?

[Spatial]

It's Nvidia. Aren't they always saying things like this?

It'll come about because BUY NVIDIA GPUS THEY ARE THE FUTURE, CPU SUX

Re:But how?

[Anonymous Coward]

This "CPU SUX" - I'm unfamiliar with it. Is it the tenth SU meaning the older version is SUIX? Was the fifth version SUV? The third version SUIII?

Or does it mean, "System Unified eXtension" ?

I'm older and I can't keep up with all this new technology and acronyms.

Re:But how?

[santiago]

Are you crazy? Why would I buy a GPU now if they're going to be 570 times faster in just a few years? I'd just be wasting my money!

Re:But how?

[risk one]

I expect the real intention is more like BUY NVIDIA STOCKS THEY ARE THE FUTURE, CPU STOCKS SUCK.

Who needs profits from actual sales when you can manipulate the market?

Re:But how?

[Anonymous Coward]

check the drawing board in http://tinyurl.com/l7zxl8
fill in obscured equation 50*1.5^6 ~ 570x
and you get correct 11x GPU performance boost

Weather

[masmullin]

Can he predict when I get my f'ing flying car!

Arbitrary number?

[forceman130]

I'm just curious how they came up with the 570x number - as opposed to say, 565x or 593x.

Re:Arbitrary number?

[eln]

The marketing guys originally wanted to say 1000x, but when they ran it past the engineers, the engineers couldn't stop laughing at such a ridiculous assertion. The marketing guys kept lowering the number, but the engineers just couldn't stop laughing. 570x is how low they got before the engineers passed out from laughing so much, which the marketing guys interpreted as agreement.

Re:Arbitrary number?

[Twinbee]

I'm wondering how you came up with 565 and 593x instead of say 564.8 or 593.82745109200174822x

Re:Arbitrary number?

[Anonymous Coward]

Here's the math

It goes something like this:

2015 Projection

CPU-Alone: 1.2^6 = 3X

CPU+GPU: 50 * 1.????underpants??? = 570X

Re:Arbitrary number?

[gmermnstinsmermwords]

the perfect amount of coffee and donuts keeps an angineers laught pass out point at just ||...

Re:Arbitrary number?

[Poorcku]

This is just about the funniest thing I've read in a month. the last sentence is gold. thx.

Re:Arbitrary number?

[Zen-Mind]

Everybody knows that 73% of people will believe anything you say when you state numbers in your claims and it rises to 91.7% when you add decimals... :P

Re:Arbitrary number?

[fractoid]

I want to buy your ??underpants??

Re:Arbitrary number?

[Anonymous Coward]

It's probably a Fermi estimate.

Re:Arbitrary number?

[Schnoogs]

Next time read the article...the guy making the prediction is the CEO of Nvidia.

Re:Arbitrary number?

[dkf]

CPU-Alone: 1.2^6 = 3X
CPU+GPU: 50 * 1.????underpants??? = 570X

At a guess, it's this:
50 * 1.5^6

Looks to me like the actual numbers aren't exactly short on underpants. Why the arbitrary x50 fudge factor? Why do they believe they'll get 50% better each year? (Why do they believe that everyone will have power supplies to support such heinous monsters?)

Re:Arbitrary number?

[Anonymous Coward]

They are assuming that a GPU has a 50 times performance advantage over CPUs right now. This is only true for a set of mathematical problems amenable to being parallelized - like graphics. Then they take the gains of the CPU in the next five years and include them in the CPU+GPU figure. But their assumption that CPUs do not add more cores over the next 6 years is dubious.

Re:Arbitrary number?

[Schnoogs]

Modded troll for suggesting someone read the article???

HAHAHAHAHAHAHHA!!!!!

Oh the irony!!! The guy who DIDN'T read the article is modded informative but the guy who DID read the article is modded troll.

HAHAHAHAHAH!!!!

Slashdot...where trolls read the articles and informative users don't!!!

HAHAHAHAHAHAH!!!!!

And you know what...

[JustNiz]

that will be minimum spec for Windows 2026, even though Windows 2026 won't have any more useful functionality than XP has.

Re:And you know what...

[hydromike2]

we cant know for sure, we will never see anything past windows 2012....

Re:And you know what...

[characterZer0]

It will at least have the same useful functionality over that Server 2008 has - non-admin users can schedule tasks, and Powershell.

Re:And you know what...

[Anonymous Coward]

Holding out for Windows 2600... for nostolgia reasons

Re:And you know what...

[Anonymous Coward]

Predictable, lame, FUD. You know damn well you're writing from your XP partition.

Re:And you know what...

[Zumbs]

But it will look soooo nice!

Re:And you know what...

[Anonymous Coward]

yet it will still be decades ahead of linux

Re:And you know what...

[Anonymous Coward]

that will be minimum spec for Windows 2026, even though Windows 2026 won't have any more useful functionality than XP has.

The ten-dimensional superstring fonts are kind of cool

Re:And you know what...

[Anonymous Coward]

that will be minimum spec for Windows 2026, even though Windows 2026 won't have any more useful functionality than XP has.

Meanwhile, it won't be supported on Linux until the next generation comes out, and even then it will take a year for new drivers with support for those cards to stabilize.

Good to know!

[CopaceticOpus]

Thanks for the heads up, Nvidia! I'll be sure to hold off for 6 years on buying anything with a GPU.

Re:Good to know!

[kramulous]

That was my immediate thought as well.

We're about to drop $250K on a GPU cluster and if performance increases to that amount in 6 years, why on earth would we buy now?

Dammit, there's just no win when you fork out for clusters (of any kind).

Should spend 50K now, stick the rest into stocks and buy 50K every year. Of course, the dudes up the tree don't like that kind of thinking.

Re:Good to know!

[Anonymous Coward]

CUDA rainbow tables
no hash is safe

GPU FTW

put an umbrella on it

[gmermnstinsmermwords]

This marks a Huge mile marker in our collective ambitions to compose all modules for the fullframe everlasting nature graphics art rendering systems

So...

[XPeter]

I have to wait six years to play Crysis?

Re:So...

[Spatial]

That Crysis joke is getting old. You can run it easily on a $100 GPU now.

Now Arma 2...

Re:So...

[XPeter]

That may be so. but I've yet to see a graphics setup that can run it on a 2560X1600 30" Monitor at Very High settings with 8X AA enabled.

I looked up the Arma sys req, and it's not as intense as Crysis was for the time.

Now Alan Wake...

Re:So...

[Spatial]

Hah! You call that a resolution? This is a resolution.

Re:So...

[SBrach]

Why doesn't anyone make a monitor that could support that resolution.

Re:So...

[fractoid]

Because you really can't see much difference? I'm running 1920x1080 on a 24" widescreen and that's "plenty" - I can still see individual pixels but they're not THAT far off my eyes' resolution if I'm sitting ~50cm away from the screen. Unless you're going to make a wraparound 'surround screen' (oh god please yes!) there's not much point in that much resolution. It's like encoding an MP3 at 512kbps - sure, there may be SOME perceptible difference but it's not enough to be worth the extra hassle.

Re:So...

[DMalic]

I'm pretty sure you can see the diff on an HDTV sized screen. After all, 2560x1600 for a 30", and TV screens are 50+.. even granting the fact you sit further, I'm pretty sure the difference would be visible and useful for say desktop work.

Re:So...

[Hatta]

All that resolution, and it still looks rendered. Instead of merely pushing more pixels, it would be nice if they did more to them, so it doesn't look so artificial.

OK

[Yurka]

But they'd better hurry up with "Mr. Fusion" which will be needed to power that thing, and finally buy the license for that demon of Mr. Maxwell's to cool it.

Re:OK

[ZonkerWilliam]

What and become a major source of global warming? Great!

Re:OK

[Spatial]

Well they did say "energy exploration" was one of its uses...

Re:OK

[fractoid]

If Mr. Maxwell would just let me borrow his demon for a while, I wouldn't NEED Mr. Fusion.

why so serious?

[ZonkerWilliam]

Seriously, it's so easy to give ambiguous figures then they can't be held to it.

Jen-Hsun Huang is full of shit

[Rix]

He constantly runs his mouth without any real thought to what he's saying. It's just attention whoring.

Re:Jen-Hsun Huang is full of shit

[vivek7006]

Mod parent up.

Jen-Hsun Huang is a certified clown who just a short while back was running around saying things like 'we will open a can of whoop-ass on Intel'.

What a dumbass ...

Re:Jen-Hsun Huang is full of shit

[im_thatoneguy]

It's even more bullshity than normal since he's also evidently predicting the end of Moore's law. CPUs only improving by 3x in 6 years?!

6 years/1.5 years = 4 Cycles of Moore's law.

2^4 = 16x performance increase.

So I guess Moore's law in the next year is going to go from a doubling every 18 months to a doubling every 4-5 years? When did that happen?

Re:Jen-Hsun Huang is full of shit

[Anonymous Coward]

without any real thought?
he is just saying "we hope to keep current trend for next six years" in a way, to get free press
in other words:
- there are some specific calculations, where current GPU is 50x faster than CPU
- suppose, that GPU computing speed will continue to grow by 50% annually
- then after 6 years, we can have GPU 50*1.5^6 = 570x faster than current CPU
see drawing board at http://blogs.nvidia.com/ntersect/2009/08/hot-chips-2009-keynote-by-jen-hsun-huang.html

And one more?

[mustafap]

> 'Huang also discussed a number of "real-world" GPU applications, including energy exploration, interactive ray tracing and CGI simulations.'"

Add to that 'MD5 collisions etc"

GPU coding really is going to separate the men from the boys. I sense a return to the old days, where people had to think about coding, and where brilliant discoveries were made.
( like this: http://en.wikipedia.org/wiki/HAKMEM )

Darn, pity I'm too old now. I'll have a play though...

Re:And one more?

[ceoyoyo]

Correction: GPU programming USED to separate the men from the boys. Now I hear you can practically program the things in C.

The math

[StreetStealth]

6 years = 72 months

Moore's Law states a doubling in transistors (but we'll call it performance) at every 18 month interval, so:

72/18 = 4 Moore cycles

2^4 = 32

So in six years, Gordon Moore says we should have 32x the performance we have now.

But it's indeed interesting... Silicon was a much easier-to-predict medium in the 20th Century. And yet here we have these two mature, opposing approaches to silicon-based computing, represented by the CPU and the GPU, with some predicting unprecedented growth for one and stagnation for the other. What will happen? How will hard material and process limitations affect the development of these? Will something exotic like artifical diamond-based ICs disrupt the market? An exciting time is in store for the sector, no doubt.

Re:The math

[Anonymous Coward]

2^4=16, actually.

And Moore's law doesn't say shit about performance. It only talks about the transistor density. So, if twice as many transistors means twice the performance, then Moore's law can also be applied to performance. But, if you need 10 times as many transistors to convert a single-cycle processor to a 5-stage pipeline processor, performance might only increase by a factor of 1.5. In general, performance won't double anywhere near as quickly as Moore's law might seem to suggest.

Re:The math

[McNihil]

Or what he is actually saying is that they (nVidia) will have more than 9 generations (~9.15) within 6 years... 1.5 generations/year... which I believe is fairly doable and actually slightly slower than the 6 month release cycle we have been accustomed to since 1998.

In other words "business as usual"

Re:The math

[BikeHelmet]

So in six years, Gordon Moore says we should have 32x the performance we have now.

No - 32x the transistors.

You fail to predict how using those transistors in a more optimized way(more suitable to modern rendering algorithms) will affect performance.

Just think about it - a plain old FPU and SSE4 might use the same number of transistors, but when the code needs to do a lot of fancy stuff at once, one is definitely faster.

(inaccurate example, but you get the idea)

Re:The math

[JJJK]

You're right, nothing about performance in that law. But with GPUs, things are a bit different... if you can squeeze twice as many shader units onto the die, you'll probably get almost twice the performance if you stick to the special class of "GPU-compatible" programs (those that need massive parallelism with little synchronization etc).

Though I would have expected the GPU people to use some of those extra transistors to implement double precision and generally make the GPU cores look more like CPU cores (more pipelining, branch prediction) to make them better suited for more complex problems like raytracing.

Re:The math

[geekoid]

No - 32x the transistors. at the same cost, in the same space.

GPUs need more RAM for us

[Entropius]

I do high-performance lattice QCD calculations as a grad student. At the moment I'm running code on 2048 Opteron cores, which is about typical for us -- I think the big jobs use 4096 sometimes. We soak up a *lot* of CPU time on some large machines -- hundreds of millions of core-hours -- so making this stuff run faster is something People Care About.

This sort of problem is very well suited to being put on GPU's, since the simulations are done on a four-dimensional lattice (say 40x40x40x96 -- for technical reasons the time direction is elongated) and since "do this to the whole lattice" is something that can be parallelized easily. The trouble is that the GPU's don't have enough RAM to fit everything into memory (which is understandable, they're huge) and communications between multiple GPU's are slow (since we have to go GPU -> PCI Express -> Infiniband).

If Nvidia were to make GPU's with extra RAM (could you stuff 16GB on a card?) or a way to connect them to each other by some faster method, they'd make a lot of scientists happy.

Re:GPUs need more RAM for us

[Anonymous Coward]

You mean like the Tesla?

Re:GPUs need more RAM for us

[Anonymous Coward]

nvidias quadro fx series goes up to 4 gb :P

Re:GPUs need more RAM for us

[avandesande]

Perhaps you know my brother? He has been doing a lot of this stuff.

http://www.google.com/search?client=firefox-a&rls=org.mozilla%3Aen-US%3Aofficial&channel=s&hl=en&source=hp&q=vandesande+qcd&btnG=Google+Search

Re:GPUs need more RAM for us

[Anonymous Coward]

Not only for scientist, but for game players.

I dream of the day I can max out GTA4, but I only have 1GB of video ram.

Re:GPUs need more RAM for us

[Chirs]

Could you grab some motherboards with multiple expansion slots and load them up with dual-gpu boards?

Re:GPUs need more RAM for us

[ae1294]

If Nvidia were to make GPU's with extra RAM (could you stuff 16GB on a card?) or a way to connect them to each other by some faster method, they'd make a lot of scientists happy.

Do you really need to ask them to do this for you? I'd think if you are a grad student you might be able to get together with some Electrical Engineering students and rig up something and turn a profit! The only thing you really need to know is how much memory the GPU can address, if you can get a hold of the source for the drivers, etc..

A video card isn't much more than a GPU with memory soldered on to it...

Re:GPUs need more RAM for us

[Anonymous Coward]

I do a number of molecular dynamics simulations myself, and while computational science on GPUs has been intriguing, for my purposes, it's been hampered by the lack of double precision. That may not happen, as it's not necessary for actual graphics, but if nVidia wants to market to a big cadre of computational scientists, that's what this community would need.

Re:GPUs need more RAM for us

[Anonymous Coward]

I find this is a common misconception amongst scientists (and some computer scientists). The most effective way to use GPUs are as streaming processors, think big vector computational units. I don't know much about lattice QCD calculations, but if it is truly easily parallelized sections of the volume can be segmented nicely, there is no reason that one card needs the full volume. Using a streaming model hides the latency from GPU to GPU because hopefully no GPUs will talk to each other, they would work in isolation, and then the results would be merged together in a separate pass. So in short extra RAM and lower latency should not be too exciting for scientists with the current 1GB+ GPGPUs when their problems have high computational costs.

You might want to check out http://portal.acm.org/citation.cfm?id=1412825

Re:GPUs need more RAM for us

[Entropius]

That product was actually specifically mentioned in the plenary talk at the 2009 Lattice Gauge Theory conference as the most likely contender for doing QCD on GPU's. It's still got the problem I mentioned, though -- not enough RAM to store everything, and not enough bandwidth to talk to the other units that are storing it.

Re:GPUs need more RAM for us

[Entropius]

You can -- that's what people are trying now. The issue is that in order for the GPU's to communicate, they've got to go over the PCI Express bus to the motherboard, and then via whatever interconnect you use from one motherboard to another.

I don't know all the details, but the people who have studied this say that PCI Express (or, more specifically, the PCI Express to Infiniband connection) is a serious bottleneck.

Re:GPUs need more RAM for us

[Entropius]

That too.

It turns out single precision is enough for lattice QCD. The step that requires the most CPU time doesn't need to generate an exact result; it only needs to get close. If the result is too far off then you wind up wasting time, but the result will still be valid.

(This is the Metropolis procedure, if you're familiar with it: the accept/reject step takes care of any computational errors that occur)

Re:GPUs need more RAM for us

[CannonballHead]

A bio botanical hair stylist?

Re:GPUs need more RAM for us

[CannonballHead]

(oops. forgot qcd bit.)

Re:GPUs need more RAM for us

[TeXMaster]

You mean like the Tesla?

No, that won't do. The NVIDIA architecture (which is shared between Tesla and graphic cards) is 32-bit, meaning that it can only flat-address 4GB of RAM tops. The more sophisticated Tesla solutions are essentially built from clusters of Tesla cards, each with its own 4GB of RAM tops. Separate memory spaces means expensive memory transfers to share data between the cards, which is not an issue if you can get good domain decomposition, but is a BIG issue if you cannot.

The revolution for HPC on GPUs would be a 64-bit GPU architecture.

Proper support for doubles and possibly even long doubles would be a plus, for applications that need it.

Re:GPUs need more RAM for us

[NP-Incomplete]

Sound like you could use one of these http://www.dinigroup.com/index.php?product=DNBFC_S12_PCIe

Re:GPUs need more RAM for us

[Anonymous Coward]

nVidia and ATI have both had cards with hardware double precision support commercially available for more than a year now.

Re:GPUs need more RAM for us

[Anonymous Coward]

Have you considered breaking the workload up into single time-slices? In other words, the first batch is all lattice points of the form (x,y,z,0), the second is (x,y,z,1), etc. I may be misremembering what little I know about lattice QCD -- I worked on an MPI lattice QCD simulation briefly as a grad student about a million years ago -- but I believe in doing so you can effectively "double buffer" it. In short, you allocate two time-slices worth of memory on the GPU, and alternate which one is the active time-slice and which one is the previous time-slice. And since GPUHost memory transactions can be done asynchronously, you can hide them with the kernel execution time.

Apologies if this seems too obvious - but I've been working with CUDA practically since day one and I've seen quite a few people make the mistake of assuming that a 3D problem should be broken into 3D, rather than 2D, sub-problems.

Re:GPUs need more RAM for us

[Elbows]

The newest Nvidia cards (compute capability 1.3) have double support. Memory is still a problem, though. The cards have a 32-bit address space with no paging/virtual memory, so I don't think they can go over 4GB, and in practice they aren't even up to that limit yet.

They are very fast if you can deal with the limitations, though. At my company we're seeing 2-3x speedup on various image processing tasks with a modest amount of optimization. And that's compared to highly-tuned code running on quad-core CPUs.

Re:GPUs need more RAM for us

[Anonymous Coward]

Erm, the Tesla 10 line and the Quadro FX 5800 both have (and can use) 4GB of memory.

Re:GPUs need more RAM for us

[tiks]

if you are memory bandwidth limited or do too much of interleaved parallel & serial processing then you might find these folks ( http://www.drccomputer.com/ ) interesting. they do a opteron socket based co-processor. In many applications i have seen being able to model the right type of operation while hiding access latency works out best. Now only if somebody will does the same co-processor with nvidia gpus that would be superb.

Re:GPUs need more RAM for us

[Anonymous Coward]

they do its called CUDA read about it

Re:GPUs need more RAM for us

[Pulzar]

I have a couple of questions... First, if your task is easily parallelized, then why do you need the full data stored on every card? Can't each GPU handle a piece of the full computation using a piece of the lattice?

Second, a 16x PCIe gen 2 slot is not much slower than the top-speed infiniband connection (8 vs 12GB/s) that you're using to interconnect your CPUs, so I can't see why PCIe would be such a bottleneck. Are you sure that that's the problem? Are you using a chipset that supports full 16x GPU slots?

Finally, is each calculation in this process so simple that an 8GB/s stream of data can't fully saturate a GPU? Isn't there more computation that can be done per one "chunk" of data?

Re:GPUs need more RAM for us

[shadowofwind]

GPUs are also both cache bandwidth and device memory bandwith limited. If you're working on a problem where your computational complexity is very high compared to your memory bandwidth requirements, then that doesn't matter. Otherwise, your performance will be more similar to an Intel multi-core CPU. Also, scientific problems that don't have high memory bandwidth requirements often require better than 32 bits of numerical precision, which is impractical on current GPUs.

Re:GPUs need more RAM for us

[Anonymous Coward]

"Impractical"? They run at 1/8th throughput (nVidia) or 1/4th throughput (ATI), but that is still fast enough to be very practical for many simulations/computations in scientific computing.

Re:GPUs need more RAM for us

[Anonymous Coward]

You've completely misinterpreted the problem. GPUs may be streamy, but the overall video card model is built to load a big set of data at the start of a game level and then keep re-using it for the whole level. The GPU is very fast, and the PCIe bus is comparatively very slow. (Just like the CPU is faster than its connection to main memory, but the GPU:PCIe ratio is even worse).

Thus the only effective way to use a GPU is to master using its video memory as a cache. Entropius' problem is that the cache won't fit the smallest possible chunk of data that can be parallelized. Thus, some parts of that chunk would have to go across the bus multiple times before the entire chunk's computations would be done, which means the entire thing would be limited by bus speed. It's exactly the same problem as the main loop of a general purpose program not fitting into CPU cache; your 3ghz CPU ends up only doing work as fast as it can talk to the 0.8ghz RAM. I don't know much about lattice QCD either, but it wouldn't take much cache thrashing for the GPU to end up being as slow or slower than the CPU.

The other potential problem is for tasks that are embarrassingly parallel but which don't require much calculation to be done for any particular byte... because if it takes less time for the GPU to do its calculations than it to took to load into video memory, once again, you're only computing at bus speed.

Re:GPUs need more RAM for us

[Chirs]

I was actually thinking about loading up as many video cards as possible on one motherboard, avoiding the inter-motherboard connection. There are motherboards with four slots that physically take x16 cards, though most of them don't give full bandwidth to all four slots.

Re:GPUs need more RAM for us

[Tolkien]

They already have (see also here). Also, check out OpenCL. OpenCL is the future, we've begun testing/working with it at my office for calculation engine. It's still in its' infancy, but it can only get better!

Sick

[Web Host Directory]

This card is sick on Crysis Wars

Re:Sick

[Joce640k]

How can a graphics card be ill?

Re:Sick

[Anonymous Coward]

Crysis has been known to give GPUs a fever...

The big question is.

[Dyinobal]

Will I need a separate power supply or two to run these new video cards? or will they include their own fission reactors?

Take it from the guy selling the GPU......

[Carbaholic]

that CPUs will only improve by a factor of 3.

570x is not that far

[gravos]

Keep in mind that is only ~3x per year because 3^6 = 729. If Moore's law holds with a 2x every 18 months that would be 16x in 6 years 570/16 = 35.652. The sixth root of 35 is 1.8. So they only have to improve the architecture by ~2x every year and ride Moore's law.

Re:570x is not that far

[4D6963]

Nice repost (it's a comment FTFA)

Re:570x is not that far

[CrashNBrn]

You mean increase an additional 100% (2x) every year on top of the normal 100% (2x) every 18 months. Quite frankly, almost every time I read interviews with Jen-Hsun Huang it sounds like he is talking completely out of his ass.

Re:570x is not that far

[osu-neko]

Almost any time I hear mention of Moore's Law, the speaker is engaging in posterior communications...

Think about that for a minute...

[dbet]

I currently run many shooter games at around 60-80 FPS at 1280x1024, and my GPU is hardly at the top end. So in 6 years, I can run a 12,800x10,240 resolution screen (if such a thing exists) at over 300 FPS. Um, sure.

We should not buy NVIDIA products until 2014

[Anonymous Coward]

Why bother buying one now when such better stuff is around the corner? ;)

Five years is a long time to wait with graphics cards, but the jump in speed will be worth it!

Nvidia's Boss: Kill Your Company a Bit Every Day

[Louis Savain]

In other news, Nvidia's Boss: Kill Your Company a Bit Every Day

Brilliant sales pitch

[Minwee]

"Did I mention that our next model is going to be SO amazing that you'll think that our current product is crap? The new model will make EVERYTHING obsolete and the entire world will need to upgrade to it when it comes out. People won't even be able to give away any older products. Sooooo... how many of this year's model will you be buying today?

"Hello? Are you still there?

"Hello?"

Re:Brilliant sales pitch

[Zen-Mind]

"Did I mention that our next model is going to be SO amazing that you'll think that our current product is crap?

Hey, I have good faith in NVidia, they already fulfilled half of that promise ... I think their current product is crap.

Thanks a lot

[Tenebrousedge]

My netbook tried to load that image and broke down in tears.

Nope!

[goombah99]

Then we can use our GPUs as our CPUs!

No No. GPU's only become CPU's when they are 570.34567 times faster. You will note that he precisely said only 570 times faster. That is he did not say an even 600 or 1000 or 500, but precisely 570, so we can assume he knew it was not 570.34567.

conversation right before he walked out

[Anonymous Coward]

"now look... I don't know why you aren't getting this... I've told you about 570 times now..."

% VS Times

[AmigaHeretic]

I'm sure this is just another case of some moron seeing 570% increase and going, WoW! my next GPU will be 570 TIMES faster!!

For the rest of us of course 570% increase is 5.7X faster.

So, CPUs increasing 3X in the next 6 years and GPUs increasing 5.7X I can maybe believe.

Re:% VS Times

[Anonymous Coward]

So a 100% increase would be 1.0X faster?

Re:% VS Times

[Anonymous Coward]

570% = 57x improvement, though maybe a exagerated, is possible. Remember, your talking about speed improvements in reference to the original item (current gpu) and the 6 year later time. If they manage to double the speed of gpu every year (which is reasonable to some degree if optimistic)

year: performance x improvement = final improvement
starting at 1 for inital speed
1: 1 x 2 = 2
2: 2 x 2 = 4
3: 4 x 2 = 8
4: 8 x 2 = 16
5: 16 x 2 = 32
6: 32 x 2 = 64

That's a 640% improvement in 6 years.

But you must consider eventually, gpu will hit a performance wall just like cpu did in recent years, it's just not there yet. Whether it hits one within 6 years, I don't know.

Re:% VS Times

[glwtta]

For the rest of us of course 570% increase is 5.7X faster.

It seems the rest of us don't understand what a "percent increase" means, either.

(hint: 570% increase == 6.7X)

Re:% VS Times

[Anonymous Coward]

I'm sure this is just another case of some moron seeing 570% increase and going, WoW! my next GPU will be 570 TIMES faster!!
For the rest of us of course 570% increase is 5.7X faster.
So, CPUs increasing 3X in the next 6 years and GPUs increasing 5.7X I can maybe believe.

100% increase = 2x as fast.
570% increase = 6.7x as fast.

Re:% VS Times

[AmigaHeretic]

Well, it runs at 5.7X the original speed right?

So a 1X CD-ROM runs at 100% speed.

A 2X CD-Rom runs at 200% the speed of an original 1x drive, no?



A 3X CD-Rom runs at 300% the speed of an original 1x drive, no?

So 5.7X CD-Rom drive runs at 570% the speed of the original drive.


So, again my point is the dumb CEO sees 570% the speed of the current chips and says, WoW! 570 "times" faster. Wrong it will run at 570% the originalor 5.7 times faster.

Where is my math wrong?

Re:% VS Times

[Anonymous Coward]

I'll feel I'm being petty here, but since I never miss an opportunity to elevate myself above my fellow man (albeit as an anon' coward)
surely a 570% increase would men 6.7x faster than the original?

Re:% VS Times

[WaZiX]

Well, it runs at 5.7X the original speed right?

So 5.7X CD-Rom drive runs at 570% the speed of the original drive.
  Where is my math wrong?

Indeed, 50% of the original performance means that performance was multiplied by 5.7. But it's a 470% increase....

In your original post you spoke of a 570% increase, which would give a multiplier of 6.7.

To make it a bit clearer, in your first post, had you used a 5% increase... you would have DIVIDED performance by 20 using your method!

Get it?

Re:% VS Times

[Wildclaw]

Where is my math wrong?

This isn't about math (well maybe a little) as much as it is about wording. Basically, the difference between "as fast/increasing/the speed" and "faster/increase". The first is a multiplicative action, while the second is an additive. So if you say, 100% as fast, you are basically saying a * 100% = a. While if you are saying 100% faster you are saying a + a*100% = 2*a. Now looking at the posts. You started by this assertion.

seeing 570% increase and going

OK. A 570% Increase. That would be a+a*570% = 6.7*a.

570% increase is 5.7X faster.

increase/faster. Good.

GPUs increasing 5.7X

Ouch. Here is the mistake. Using increasing. Suddenly you have downgraded the new anticipated speed from 6.7*a to a*5.7=5.7*a.

Now, in your second post you used the multiplicative action straight through. But that is pretty much the opposite of the first post that used the additive action except for a single time at the end.

Also, as a general advice, use X (times) only to represent a multiplicative action. That is what the general meaning of the word is, so it can be confusing when you here it describing an additive action. Not really wrong, but confusing.

Re:% VS Times

[Anonymous Coward]

570% = 57x improvement

Go back to school please (if you're in one, pay more attention). 570% = 5.7x improvement.

Re:% VS Times

[Anonymous Coward]

Crud. Mis-modded.

about 2.88x better each year for 6 years

[ZiggyM]

is another way to say it.

crysis!!

[Conditioner]

WTH, and no one made a "I can finally run crysis" or "i can finally run Vista" joke ?

Bitchin' fast! 3D

[Mursk]

I hope it looks like this:

http://www.russdraper.com/images/fullsize/bitchin_fast_3d.jpg

Re:Bitchin' fast! 3D

[geekoid]

Nice.

Re:Bitchin' fast! 3D

[Anonymous Coward]

Wow, that reminds me of the prototype Voodoo 5 6000 cards which were insanely long, needed so much power that it required a separate AC adaptor, but still didn't any better performance than the Geforce3. I for one hope that this 570x performance boost doesn't require an additional 500W of power.

Re:Bitchin' fast! 3D

[jamesshuang]

I am VERY entertained by how they used 256mb of ram as a "ridiculous" amount.... http://www.newegg.com/Product/Product.aspx?Item=N82E16814187054 hehehe

285 Trillion floating point operations per second?

[AmigaHeretic]

Current Nvidia GPUs have 500 gigaflops of performance in single precision. 20 teraflops would be 40 times faster. 570 times faster in 2016 would be 285 teraflops.

They are shitting bricks

[Latinhypercube]

They are shitting bricks because fast multicore (8 core) cpu's are around the corner. We are all going the multi-cpu route. Hence we have an extra cpu or two to do gpu calculations. So Nvidia needs to reassure everyone the gpu's are the way forward. Not multi -purpose multi-core cpu's.

Re:They are shitting bricks

[Rockoon]

And wasn't it only two years ago the NVIDIA repeatedly stated that RTRT was stupid?

..and then a year ago that they began saying that they could do it just as well as Intel?

Don't get me wrong. I love nVidia cards. I just don't believe anything they say.

press people don't listen

[Anonymous Coward]

I was at Hot Chips. The press guy totally got it wrong. It was a toy example of performance relative to 1 core.

Comparison

[Lord Lode]

Yeah, they also always say that batteries are 10 times better. But can I use my phone 50 days instead of 5 days without recharging? No, it's always the same amount of days as 10 years ago.

Re:Comparison

[geekoid]

because they are a lit smaller. Use the current batterys and make them the size of batteries 15 years ago and you will get 50 days...hell, probably 5 months.

Overestimating the short-term...

[marciot]

Someone once said that tech visionaries often overestimate what can be accomplished in the short-term, but underestimate what can be accomplished in the long-term, I believe this is what will happen here. Not in five years, but in ten years to twenty years, we might see improvements exceeding this estimate.

Re:Overestimating the short-term...

[geekoid]

"tech visionaries" and "futurists" are just people that don't really do anything, but want the attention anyways. Usually accredited to someone who did something kind of neat 20 years past.

Really?

[ChristianMc]

So does this mean in 6 years, the average core clock of a GPU in Megahertz will be... OVER 9000?!?!

570 times better?

[Anonymous Coward]

Wow, imagine the rendering and physics on the breasts of female characters in games!

I mean, seriously, what else is there to render? Stupid trees? Give me a break...

Boobies!

I can finally run DNF!

[galanom]

I can finally run DNF!

Random numbers

[Rix]

Like I said, he's just running his mouth and pulling numbers out of his ass.

GPU Designer

[dunc78]

Look, the GP has probably been designing GPUs for at least 7 minutes. Why would you not think he has all the bases covered?

Easy to compare to CPU's

[cbreaker]

Once the GPU can do general purpose computing (read: Everything) then we can talk about a paltry 3x improvement. The GPU is a specialized processor and therefore can do that one thing very efficiently.

They could make cars extremely fuel efficient if they they only didn't have to turn, accelerate and brake, and carry passengers. But they do, and the CPU also has to be able to process all kinds of math.

Just like how some Oil exec last year said we'd see $1 gas at the pump this year, we don't see a nearly six hundred fold increase in GPU peformance in the next six.

Mod that one up.

[Joce640k]

Desktop CPUs are now "fast enough" for most people (and have been for a few years now). The main bottleneck in desktops is normally lack of RAM (caused by all the crap people install) and hard disks.

The interesting PC developments in the next few years will be SSDs and low-consumption CPUs.

Most people don't want mega-GPUs either - remember that Intel has more GPU market share than ATI and NVIDIA combined.

nVidia already provide for HPC

[gnuman99]

Look at nVidia's website. They already provide HPC solutions,

http://www.nvidia.com/object/tesla_computing_solutions.html
http://www.nvidia.com/object/preconfigured_clusters.html

http://www.nvidia.com/object/product_tesla_s1070_us.html

# of Tesla GPUs 4
# of Streaming Processor Cores 960 (240 per processor)
Frequency of processor cores 1.296 to 1.44 GHz
Single Precision floating point performance (peak) 3.73 to 4.14 TFlops
Double Precision floating point performance (peak) 311 to 345 GFlops
Floating Point Precision IEEE 754 single & double
Total Dedicated Memory 16 GB
Memory Interface 512-bit
Memory Bandwidth 408 GB/sec
Max Power Consumption 800 W
System Interface PCIe x16 or x8
Software Development Tools C-based CUDA Toolkit

Maybe point your quantum mechanics simulations people to HPC specific nVidia solutions, not commodity hardware not designed for HPC (ie. no RAM, limited cores, etc.)

My question is...

[Anonymous Coward]

how much will my electric bill go up?

A GeForce GTX 260 that I just recently bought requires bare minimum of 650w power supply. With all the talk about going green lately, I'm wondering when or if we'll hit a ceiling for energy consumption in the home computing market.

Re:My question is...

[DMalic]

looks like it has a 236 watt TDP. so it needs a very hefty PSU, but nowhere near 650 watt, even with a fast quadcore.

The *real* math

[Anonymous Coward]

Ugh... Repeat after me:
2^0=1
2^1=2
2^2=4
2^3=8
2^4=16
570x perf with 16x transistors, in the same power envelope, those engineers must have done a really crappy job so far!

What does it mean in capabilities?

[Anonymous Coward]

We can, already, generate images which appear totally real to many.

With a high enough frame rate, and resolution, it can be made more real than real.

But once you reach the point that you exceed the human eye's ability to take in, and the brain to comprehend the images, you are beyond diminishing returns.

You don't actually see most of what you think you see anyway, for instance, you are probably not aware continuously of the fact that there are areas in your field of vision in which you cannot see anything at all...your brain fills in the blank by filling it in from previous images of the areas (as your eyes are never totally still.)

I'd rather see a graphics/display combination which would automatically adjust the gamma so that the images are visible on the screen...too many programs of all sorts assume that your monitor will be set for it's highest gamma, rather than assuming that it will be set in the mid-range, which is where you get the best combination of highlight and shadow detail.

But hey, with that kind of increase, I should be able to run surround video...assuming someone builds a display....

NVidia is panicking in the face of Larrabee!

[itsybitsy]

The coming Larrabee extensions to the X64-86 architecture will cream nvidia and they are frightened of this so they are panicking with silly predictions soothsaying the future bright for GPUs while pretending to portent a bleak future for CPUs.

I'll take a Larrabee style CPU capable Predicate+SIMD+4Tread Multicore processor over NVidia or ATI any day BECAUSE it is a general purpose cpu with TONS of software that can work on it already. Basically Larrabee is an open spec architecture extension while NVidia keeps it's assembly language secret. Give me access so my code can run free! Thanks Intel! Now share Larrabee with AMD please! Oh, and bring the Larrabee instruction enhancements to ALL Intel and AMD CPUs please! Thanks again.

Re: Not using free drivers, are you?

[xiando]

Yes, NVIDiAs propietary nvidia driver really is better than ATIs propietary fglrx driver - which sucks. However, it must be mentioned that both the free xorg radeon and radeonhd drivers perform better than ATIs propietary driver in 2D (while the 3D support is poor) and that the obfuscated xorg nv driver for nvidia is redicilous. If you are using GNU/Linux and you are fine with propietary drivers the nvidia is probably the best choice. Intel and ATI are way better choices if you are not.

What he actually means...

[XDirtypunkX]

What he actually means is that your 570 dollar graphics card will be worth 1 dollar. That's a 570x better performance/price ratio!

2003 CPU/Graphics Card list from pricewatch.com

[DotDotSlasher]

From the historical-perspective dept.
I keep track of CPU and graphics card prices from pricewatch.com, for no-good-reason.
slashdot does not like my long lists (too few characters per line) so here are the abbreviated lists which keep the most expensive options.
Here are my scrapes from ~6 years ago:

Sep 16, 2003 List of Graphics Cards
$384 - Fire GL Z1 128mb
$696 - Fire GL X1 256mb
$529 - Fire GL X1 128mb
$469 RADEON 9800 Pro Ultimate
$373 - RADEON 9800 Pro Ultimate 128
$417 - RADEON 9800 Pro 256MB
$294 - RADEON 9800 Pro 128MB
$368 - RADEON 9800 All-In-Wonder Pro
$192 - RADEON 9800 128MB
$299 - RADEON 9700 Pro Ultimate
$266 - RADEON 9700 Pro
$389 - GeForce FX 5900 256MB
$235 - GeForce FX 5900 128MB
$251 - GeForce FX 5800 128MB


6/16/2003 List of CPUs:
$467 - Xeon 2.8GHz 533FSB
$315 - Xeon 2.66GHz 533FSB
$235 - Xeon 2.4GHz 533FSB
$236 - Xeon 2.0GHz 533FSB
$839 - Opteron 244
$708 - Opteron 242
$280 - Opteron 240
$451 - Athlon XP 3200
$440 - Athlon XP 3200 400
$249 - Athlon XP 3000
$294 - Athlon XP 3000 400
$282 - Athlon MP 2800
$205 - Athlon MP 2600
$158 - Athlon MP 2400
$127 - Athlon MP 2200
$199 - Athlon MP 2100
$122 - Athlon MP 2000
$147 - Athlon MP 1900
$149 - Athlon MP 1800
$115 - Athlon MP 1600
$108 - Athlon MP 1500

This has all happened before...

[hazydave]

When DEC unveiled the first Alpha CPUs, way back when, they called for a 1000x increase over their projected 25 year life cycle. 10x of this was coming from architectural improvements, 10x from clock speed improvements, and 10x from parallelism. The cycle's much shorter these days, also shorter with today's GPUs vs. CPUs, so I don't see the problem. Plus, there are quite a few advantages to being a GPU, far as benchmarking and software goes -- they aren't subject to the same kind of legacy issues that CPUs are.

A couple of years ago, you got around 128 stream processors on a high-end card -- the top of the line stuff today has what, 800 (Radeon 48xx series). This trend show no end it sight... and they might actually be getting more useful, as chips are designed with numbers of stream processors in tight clusters, and at the same time they're getting more general purpose. So much work is done internally on these types of operations, memory bandwidth may need only grow by 10% to match a 100% increase in overall performance.

In a way, it's kind of a cheat, though, compared to CPU performance. GPUs are complex enough, and right now still fairly specific purpose, the only benchmarks you get are actual graphics benchmarks, and some estimate of peak performance. So that's your 570x factor here: it's based on a peak measure of everything going as well as can be expected. In real life, that may not be something you get to see often, if ever.

Also, GPU performance is and will for some time be measured by things the GPUs do well. If the GPU doesn't help, you use the CPU(s) in the system. This seems normal, but what it really means is that GPU performance well largely only be judged by what GPUs do well.. they don't get bad marks for code that's not designed to run on 2400 processors simultaneously. CPUs, on the other hand, will be judged as part of that default expectation... if that great new Athium Coreathon x16 only speeds up Photoshop by a factor of three, that's what people will care about ... not whether it's really peaking at several times that level. CPUs will still have to run non or poorly threaded code, and only get the credit in the public mind for where regular things get faster.

With that said, it's also the case that the CPU companies have been at it longer, and have had access to state-of-the-art chip process for much longer, too. Because of this, CPUs are a more mature technology than GPUs, so one naturally expects their growth to be slower. It's fairly recent than GPU companies have had access to the same kind of process magic that folks like Intel and IBM have had for decades.

Re: Not using free drivers, are you?

[JustNiz]

Yeah I really dont care that my drivers are just binary blobs. I have no desire to mess with my graphics driver anyway.