Slashdot Mirror
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.'"
Then we can use our GPUs as our CPUs!
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.
moox. for a new generation.
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?
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.
You just got troll'd!
Both seem highly unlikely.
"When life gives you lemons, don't make lemonade. Make life take the lemons back!" -- Cave Johnson
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.
Thanks for the heads up, Nvidia! I'll be sure to hold off for 6 years on buying anything with a GPU.
I have to wait six years to play Crysis?
"The difference between genius and stupidity is that genius has it's limits" - Albert Einstein
He constantly runs his mouth without any real thought to what he's saying. It's just attention whoring.
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.
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.
-The art of programming is the pursuit of absolute simplicity.
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.
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.
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.
You can never go home again... but I guess you can shop there.
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:
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.
No sig for you. YOU GET NO SIG!
Will I need a separate power supply or two to run these new video cards? or will they include their own fission reactors?
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.
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.
This game will waste your life. Don't clicky!
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.
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)
"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?"
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.)
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.
Some drink at the fountain of knowledge. Others just gargle.
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.
I hope it looks like this:
http://www.russdraper.com/images/fullsize/bitchin_fast_3d.jpg
"This thing does science so hard, you say, 'I've never seen that much science.'" -Sam
We'll obviously need a Turbo button on it set in the off position until such time that the CPUs catch up.
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
Maybe the high-end ones, but the low-end GPUs are mostly passively cooled and still much more powerful than old GPUs.
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:
Both seem highly unlikely.
It's not a linear relationship.
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
"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.
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.
You can never go home again... but I guess you can shop there.
Either
Neither seem highly unlikely.
A slashdotter who didn't build his own computer is like a Jedi who didn't build his own lightsaber.
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?
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.
"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.