The hidden potential in your graphics card: A supercomputer?

[lowlifecat]

http://www.tgdaily.com/2006/09/18/peakstream_supercomputer/

Wolfgang Gruener September 18, 2006 16:49


Redwood City (CA) - The most powerful computing device in your PC may not be that dual-core processor, but your average graphics card. Interest in tapping the hidden processing power in graphics processor has been growing over the past two years, but Peakstream is the first company to actually offer a solution to create a supercomputer based on graphics cards.
Peakstream claims that it has developed a new software platform that can create supercomputers by combining the processing capability of common CPUs with the resources of modern graphics cards. Simply by adding the horsepower of graphics cards to an existing computer, the company claims that the original system can be accelerated by a factor of 20x.

http://www.ad.tgdaily.com/cgi-bin/bd...etition_51.jpg
Compared to graphics processors, main system processors have developed much slower in providing an increase in floating point performance. For example, a Pentium 4 3 GHz processor was estimated to reach about 6 GFLOPs (billion operations per second), while Woodcrest, the server variant of Intel's Core 2 Duo processor, is currently believed to top out at around 24 GFLOPs; according to Intel, a four-processor dual-core Itanium 2 system recently reached 45 GFLOPs.
At least in the floating point discipline, graphics chips are way ahead of the game: ATI recently said that its current high-end X1950 XTX processor brings in 375 GFLOPs, in dual-graphics mode even up to 750 GFLOPs - the equivalent of 31 Xeon 5100 processors. Nvidia's Geforce 7950 GX2 dual-GPU card is rated at 384 GFLOPs and Ageia's physics processor at 96 GFLOPs.
There have been several advances to take advantage of the floating point performance of the GPU when it isn't completely consumed by a video game. For example, Nvidia announced earlier this year to enable physics processing in dual-GPU SLI configurations; ATI calls this process load balancing and talked about opening up the hardware platform to allow software developers to access excess performance capability in graphics cards. Independent projects included BionicFX, which developed a technology to use graphics cards for accelerated audio processing.
Peakstream employs a similar concept and provides software developers an interface to take advantage of graphics cards. The firm's C and C++ based application interface and virtual machine handles work scheduling and memory management and a way for programmers to easily access a combined system of processors - such as x86 or Cell chips - and graphics cards. According to Peakstream, the system can be used with standard developer tools such as Microsoft's Visual Studio or Intel's Compiler software and requires only a "minimal learning curve." The platform is also able to run on unmodified Windows and Linux operating system, the firm said.

Peakstream's platform model

Theoretically, the platform would allow anyone with some programming experience to build a small supercomputer at home. However, the fact that the performance is floating-point focused, Peakstream is aiming for customers in the traditional supercomputing market. The firm believes that especially applications in oil and gas, financial services, defense and academia will be interested in the solution. The company promises that its platform can provide performance increases in industry specific benchmarks between 16x (Monte Carlo Simulation) and 21x (Kirchhoff Migration) over CPU-only systems.
Price may be the main component that will keep private users from taking advantage of the Peakstream platform. The company charges $2000 per computing node for its software, but mentioned that it will provide volume discounts for large cluster systems.

Interesting!!!

[Dr Mordrid]

Remember my "Ah-ha" thread back when the AMD/ATI merger was announced? In it I discussed GPGPU (General Purpose Graphics Processor Unit, AKA GP^2) and why AMD is so enamored with the techniques enabling some of ATI's AVIVO technology;

http://forums.murc.ws/showthread.php?t=57845&highlight=gpgpu

Based on this article in The Register;

http://www.theregister.com/2006/07/24/amd_etc/

The deal, announced today, goes back some time. Last year, at Computex in Taipei, it was apparent that ATI and AMD were falling in love with the idea of using the powerful graphics processor to run computer programs, not just for animating video.

At that show, software developers were invited to the launch of the new dual-core AMD processors, with prototype applications that ran, not on the x86 central processor, but on the graphics chip. Examples included video editors which could handle the output stream live, in real time.

Wikipedia article on GPGPU;

http://en.wikipedia.org/wiki/GPGPU

The first indications are in this VERY interesting post from Beyond3D regarding ATI coming out with GPGPU based cards using what they're calling "FireStream" on their FireGL cards;

ATI to release 'FireStream' GPGPU accelerators?

31-Aug-2006, 09:36.43 Reporter : Rys

Pay close attention to the X server log output from ATI's Linux driver, as we did recently with a FireGL V7350, and you'll see mention of support for what we believe is a brand new class of product featuring a modern graphics processor. Listed in recent drivers is a SKU called FireStream 2U, apparently powered by ATI R580 and given away by the following strings: FireStream 2U (R580 724E), FireStream 2U (R580 724F).

Using R580 and with unique device IDs of 724E and 724F, at least as far as the Linux driver is concerned and confirmed on ATI's vendor ID page, these FireStream products appear to be distinct boards with R580, but with a different purpose than 3D graphics acceleration.

First consider the ATI SIGGRAPH '06 presentation on what they call a Data Parallel Virtual Machine, using a recently released 'close to the metal' API, built specifically for general purpose GPU computing (GPGPU). Then consider that ATI have spent significant resources on GPGPU since the inception of the R5-series of GPUs, lastly bearing in mind that stream processing is a class of problem that exploits massive parellism of a device, and it seems logical that FireStream is a product using a 3D GPU but is not for the explicit purpose of drawing pixels.

We therefore surmise that FireStream products from ATI will come in significantly different configurations to those you're used to seeing for desktop discrete graphics, possibly without any display outputs at all, and almost certainly with different PCI Express slot configurations for use in platforms providing 1x, 2x, 4x and 8x slots -- in addition to PEG16X of course. We invisage stream processing clusters with multiple FireStream boards per system, possibly with multiple GPUs per board, designed to exploit their DPVM efforts and accelerate parallel processing problems that fit a modern GPU, as well as their recent Havok FX relationship for the mainstream. Also, thinking about spy shots of upcoming Radeon desktop boards with NVIDIA-like inter-board connectors, we see no reason why future FireStream/DVPM can't also exploit those 'Crossfire' links for concurrent high-bandwidth connection of multiple FireStream boards, for further performance increases and inter-GPU communication outside of the PCI Express bus.

Whatever the final configuration of the product and how it operates, it seems clear to us that FireStream is a coming reality and the first product of its kind from either of the main desktop GPU IHVs to address a growing and very significant market, and from ATI's perspective exploiting as much as possible the architectural traits they've built into R5-series hardware from day one. More on FireStream as we get it.