Slashdot Mirror
The Future According To nVidia
NerdMaster writes "Last week nVidia held their Spring 2008 Editor's day, where they presented their forthcoming series of graphics processing units. While the folks at Hardware Secrets couldn't tell the details of the new chips, they posted some ideas of what nVidia is seeing as the future of computing. Basically more GPGPU usage, with the system CPU losing its importance, and the co-existence of ray-tracing and rasterization on future video cards and games. In other words, the 'can of whoop-ass' nVidia has promised to open on Intel."
That's my main influence when I purchase video cards.
"Thanks for all the money you paid to us. We've used it to buy off ISO among other things" -Microsoft
The future according to Sun or IBM.. faster CPUs. The future according to Nvidia... more GPUs .. the future according to Seagate.. exabytes and petabyes, the future according to Minute Maid.. , the future according to Blue Bonnet .. lower cholesterol, the future according to ATT "more bars in more places", the future according to ...
Another paid for article. Yawn.
I'm all for it.
The more competition the better.
Anyone that worries too much about the cost a good GPU adds to the price of a PC, doesn't remember much what it was like when Intel was the only serious player in the CPU market.
This kind of future, to me, spells higher bang for the buck.
"the 'can of whoop-ass' nVidia has promised to open on Intel."
Yep, I'm sure the Intel Devs have all taken a sabbatical.
AT&ROFLMAO
The leading manufacturer of GPUs wants GPUs to become ever more important.
And nVidia has been spouting a lot of this lately. Is the company in trouble and the top executives are now trying to avoid that impression by constantly talking aboit how bright the future is for their company? Quite possible, I would say.
As to the claim that the GPU will replace the CPU: Not likely. This is just the co-processor idea in disguise. Eventually this idea will fade again, except for some very specific tasks. A lot of things cannot be done efficiently on a CPU. I have to say I find the idea of the GPU actually becomming less important, because of more available general-purpose CPUs actually a lot more convincing. And it has the advantage of far simpler hardware. Especially nVidia GPUs have ben riddled with problems for some time now. CPUs (either AMD or Intel) have not been.
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
Extreme Programming - Redundant Array of Inexpensive Developers
I was wondering about this...now that nVidia wants CPU to loose its importance _and_ they started to cooparate with Via on chipsets for Via CPUs (which perhaps aren't the fastest...but I've hard the latest Isaiah core is quite capable), will we see some kind of merge?
One that hath name thou can not otter
nVidia doesn't do the APIs for their cards. They have no properitary API, their native APIs are DirectX and OpenGL. In fact, the advances in those APIs, more specifically DirectX, often determines the features they work on. The graphics card companies have a dialogue with MS on these matters.
This could be an area that OpenGL takes the lead in, as DirectX is still rasterization based for now. However it seems that while DirectX leads the hardware (the new DX software comes out usually about the time the hardware companies have hardware to run it) OpenGL trails it rather badly. 3.0 was supposed to be out by now, but they are dragging their feet badly and have no date when it'll be final.
I imagine that if MS wants raytracing in DirectX, nVidia will support it. For the most part, if MS makes it part of the DirectX spec, hardware companies work to support that in hardware since DirectX is the major force in games. Until then I doubt they'll go out of their way. No reason to add a bunch of hardware to do something if the major APIs don't support it. Very few developers are going to implement something that requires special coding to do, especially if it works on only one brand of card.
I remember back when Matrox added bump mapping to their cards. There was very few (like two) titles that used it because it wasn't a standard thing. It didn't start getting used until later, when all cards supported it as a consequence of having shaders that could do it and it was part of the APIs.
The GPU is a massively parallel computing machine. The cores are 50 MHz. This is fine if you can parallel your needs as you can easily do for video, but for general CPU needs you are executing very slowly if you use only one, 50 MHz core, as you likely will.
Hardware accelarated raytracing could also be interesting for speeding non-realtime rendering such as for making movies!
It arrived Friday. Wow. 430 to 512 BILLION FLOATING POINT OPERATIONS PER SECOND! Need I say any more. Yummy, now I get to play with it!
Ahhhh. You are a Zen Master. Please, teach us more!
I made sure that my current machine had an nVidia graphics chip, so that I could play with stuff like CUDA. But my machine also runs Vista and, some 18 months after its release, there still isn't a stable version of CUDA for Vista. Plus, seeing as my machine's a laptop, I doubt that even the beta drivers available from nVidia would install, seeing as how they're prone to playing silly buggers when it comes to laptop chip support.
So nVidia, instead of spouting off about how great the future's going to be, how about supporting some of the people who are trying to run your hardware in the present, eh?
Wouldn't that be great! It's about time that graphics processing, IO, an other things are sent to their own processors. Anyway, wasn't that done before - Amiga?
Where does IBM/Sony's Cell processor fall in the CPU/GPU battle? IBM most certainly plans to use it in PCs as a CPU, but wasn't most of the initial development focused on making it a better GPUing CPU?
Grandpa: My Homer is not a communist. He may be a liar, a pig, an idiot, a communist, but he is not a porn star.
Funny how things work, isn't it.
I run: Windows, OS X, Linux, FreeBSD. Just because you have a hammer, doesn't mean everything is a nail.
Conan: "... It's time, once again, to look into the future."
... La Bamba's high falsetto ... "In the Year 2000"
Guest: "The future, Conan?"
Conan: "That's right, Let's look to the future, all the way to the year 2000!"
and then
"In the Year 2000"
Maybe instead of worrying about the next generation of GPU's they should be more concerned with fixing their shitty driver support for Vista. I have a laptop that came with Vista but the driver support is so useless and buggy I had to downgrade to XP. The control panel options under XP vs. Vista are night and day.
... Dumbass ...
Because of all the problems I have had with it I doubt I will ever buy another laptop/motherboard with a nVidia product on it.
It's hard to come up with the latest and greatest when nobody will buy your products
Back in the day, if you ran 'math-intensive' software it would look for an 8087 math co-processor and load special code libraries in Lotus 123 to speed up calc performance. Once Intel had the chip real estate to spare though, this special purpose chip got subsumed into the CPU. As Intel keeps driving the transistor count up, they will be perfectly capable of embbeding a full-featured 'streams' processor into their CPUs. It won't happen right away, but it solves the issue of different code libraries (a software company's nightmare) and will be good enough for 80% of the folks out there. High-end graphics will always need a hi-perf, dedicated solution, but the market will be smaller (which I think is nVidias first worry). Add to this some gaming companies recent threats to abandon the PC as a game platform due to piracy concerns, and their market could get much smaller. Hence all the bluster about GPUs taking over for the CPU. Historicaly, it's always been the CPU that takes over special-purpose functions, not the other way around (at least in the Intel space). No one else has the cash and the facilities (same thing really) to drive the transistor count up enough to do this trick.
Nvidia's chief scientist, David Kirk, is really down on raytracing and particularly on dedicated raytracing hardware.
http://scarydevil.com/~peter/io/raytracing-vs-rasterization.html
However... Dr Philipp Slusallek, who demonstrated how even a really slow FPGA implementation of raytracing hardware could kick general purpose processors (whether CPU or GPGPU) butts in 2005, has been working as a "Visiting Professor" at nVidia since October 2007.
They're still playing their cards close to their chest.
Intel obviously sees the threat of the GPU creators, but their attempts at breaking into the GPU market hasn't been very successful.
Their next generation effort is called Larrabee. Which uses multiple x86 cores linked with a ring bus.
It actually reminds me of PS3 SPU setup but Intel is using the GPU functionality as a wedge into the GPU market, instead of pushing it for general computation. But, since standard C code will work on it, you can rewrite the entire stack to be a physics co-processor or fold@home client.
Ultimately, I see the CPU and GPU separation disappear and merge into one chip, much like FPU and sound card functionalities.
AlwaysOnGames Arcade
Why not employ numerous Field Programable Gate Arrays (FPGAs) instead of a CPU? You could program one or more FPGA to be optimized to execute each of the functions the software needs. Need more FLOPs? Program for that. Need scalar computation? program for that. Seven FPGAs running one way and nine running another. At some point, FPGAs may completely replace the CPU as we know it today. The HPC community is already looking at this possibility for some types of computations.
Invenio via vel creo
The Cell would suck in a PC. There's an underpowered, in order PowerPC core and a bunch of SPEs. But SPEs are for signal processing, not general computation. They only have 256K of memory, smaller than the cache inside a desktop CPU. They don't have access to main memory or an MMU. Even if you added an MMU and paged from main mem to the SPEs it still wouldn't help. The bus would be saturated by page misses and the SPEs would spend all their time waiting.
Even in a games console it's probably hard to keep all the SPEs usefully employed.
echo -e 'global _start\n _start:\n mov eax, 2\n int 80h\n jmp _start' > a.asm; nasm a.asm -f elf; ld a.o -o a;
nVidia seems to be a litte late to this game. ;-)
CAn'T CompreHend SARcaSm?
Focus generally seems to be on "bigger" as opposed to "more efficient." Add more cores, increase the frequency, etc etc.
Some other tasks focus on "trimmed down and more efficient" but then tend to fail in the power output arena.
I was wondering how difficult it might be to make a motherboard or graphics card with multi-processors. One small one for general-purpose computing (basic surfing, word-processing, 2d graphics or basic 3d), and a bigger one that could be used to "kick in" when needed, like an overdrive engine, but otherwise non-power and sleeping until needed.
Many CPU's already have power-states, but generally these aren't as efficient as a CPU specifically designed for lower-power/efficiency. Same for GPU's. So how about cores or processors that come online on-demand?
Comment removed based on user account deletion
For me, fullscreen TV output. There's none in 8xxx series and it's broken in older card drivers. See as an example.
Ant(Dude) @ Quality Foraged Links (AQFL.net) & The Ant Farm (antfarm.ma.cx / antfarm.home.dhs.org).
nVidia is scared by Intel's push. Intel it putting a lot of money in getting the brightest graphics-heads out there.
Ray-tracing is for marketing and managers only. Marketing has to make sure that all bases are covered, but marketing people don't really know much about actual technology.
Managers know more, but often not enough to realize that ray-tracing by itself means nothing.
Possibly most graphics programmers are also confused by the feasibility of ray-tracing in real-time. To see so many, so called, experts get dragged into this ray-tracing FUD it's both sad and funny.
It's also sad to see nVidia getting into this silly game just to make sure that they aren't falling behind on the marketing side.
Vector coprocessors, array processors, and all that have been around for ages. Maybe they'll finally catch on. If they do, you can bet that the manufacturer making them will not be a graphics card manufacturer. In fact, by definition, they won't be a graphics card manufacturer, since they will be making co-processors for non-graphics applications.
But I don't think they will catch on. It makes little sense for people to stick extra cards into their machines for computation. Instead, you'll probably see something more like Cell, a combination of CPU plus vectorized coprocessors on a single chip, and if you plug multiple of them into the motherboard, you get the right mix of CPU and coprocessor performance.
Seriously.... Over the past few years Nvidia has shown me that they could really care less if games that aren't brand new will run on their cards. Older games that use palletized textures and 16bit function calls look horrible on the newer cards. This is something they could fix easily in software if they wanted to.
The parent's situation may be somewhat unique, but there's a very much more common situation that hits the same problem, namely EDID blocking or mangling by KVMs. The handling of EDID is patchy at best on a lot of KVMs, even on new ones, so the graphics card can easily reach a wrong conclusion. While it's OK to make assumptions as a default, the settings need to be user-selectable to get around this problem.
There's no market for high-end video cards in linux except for a few CAD tools, and those people don't care at all about open source drivers. So I'm afraid you're a demographic that just doesn't matter to them...
According to nVidia the dream gaming system will consist of quad nVidia GPU cores running on top of a nVidia chipset-equipped motherboard, with nVidia-certified "Enthusiast" system components. Meanwhile the company just will not work on LOWERING the power consumption of their graphics cards. Why do we need one-kilowatt power supplies? Because nVidia says so!
Fuck nVidia.
In a way nVidia's message is the same as that of the Cell ship. There will be more and more use of parallelism, with the CPU (Or a particular CPU on symmetric multi-core systems) acting as a kind of foreman for a troop of processors working in parallel.
Not all of the uses for the gobs of cheap parallel processing power are apparent yet. But people will find cool things to use it for, just as they found cool things to use home computers for. In a way, we are now going through a home supercomputing revolution.
Actually, with most motherboards coming with onboard video (that is usually less powerful than the add-on GPU), this sounds like a really good idea. Of course, in this case you'd need a compatible card (onboard ATI+addon ATI, or onboard Nvidia+addon NVidia). I wonder if it could be standardized so that the lesser-power onboard GPU's could be switched down and allow a passthrough for the addon AGP card (or vise-versa, since the addon card is more likely to have extra ports such as DVI etc than the onboard/motherboard which has limited space).
Possibly Intel, possibly ATI.
But nVidia is the last to publish specs, or any sort of source code. ATI and Intel already do one of the two for pretty much all of their cards.
So, in the long run, nVidia loses. It's possible they'll change in the future, but when you can actually convert a geForce to a Quadro with a soft mod, I very much doubt it'll be anytime soon.
Don't thank God, thank a doctor!
google it if you care...
So true. We only need a minimal computer for the office tasks...however it takes a 3GHZ HT computer to do them at the same time as AV/Firewall :(
:(
The terminal emulator runs slow on anything less
Heck with physics processors and GPU, i need an AV card and i could go back to Pentium 3...
If a company can support Linux (or other OSS OSes) then I have more confidence in them with general all around support in other areas such as OSS standards. This is why I buy nVidia before I even consider ATI or other Windows/MacOS specific driver'ed hardware.