NVIDIA is continuing its campaign to nudge the CPU from its dominant position at the center of the computing universe. A trio of announcements this week provides a rough outline of how the company intends to expand its GPU computing footprint.

Cloud Computing Meets the GPU

On Tuesday at the Web 2.0 Summit in San Francisco, NVIDIA announced a new platform that positions the GPU as the engine of a 3D Internet. In a nutshell, the company has constructed a Web services model that employs server-side Tesla GPUs to drive photorealistic imaging to client applications. The idea is to take advantage of the computational muscle of HPC-class GPUs so that high-end imaging applications in areas like medical diagnostics, product design, and manufacturing CAE can be co-located to the cloud. We covered the particulars earlier this week in our feature story.

It’s worth noting that AMD announced something along the same lines back in January of this year when the company revealed plans for a one petaflop GPU-accelerated supercomputer to drive HD content across the Web. The chipmaker called its machine the “AMD Fusion Render Cloud,” but unlike the NVIDIA platform, the supercloud was aimed at online gaming, HD video applications, and film rendering.

At the time, AMD CEO Dirk Meyer said the machine would be powered by 1,000 ATI Radeon HD 4870 processors, and that they plan to have the system up and running by the second half of 2009. In the interim, the company came out with the ATI Radeon HD 5870 GPU, which delivers 2.72 teraflops (single precision) per chip. Using the newer silicon would substantially cut down on the number of GPUs needed for a one petaflop machine. But since AMD has been silent about the GPU supercloud since it was initially announced, it’s conceivable, and even likely, that they shelved the whole project.

NSF Puts GPU Super on Track

On Wednesday, Georgia Tech announced that the NSF is pitching in $12 million over five years to fund a project for two GPU-equipped supercomputers under its Track 2 program. Track 2 is designed to spread federal science money to academia for experimental sub-petascale HPC systems. According to the press release, this is the first Track 2 award to go toward GPU-accelerated supers.

The $12 million will be allocated for the deployment and operation of the HPC machinery, which will be shared across Georgia Tech’s College of Computing, Oak Ridge National Laboratory, and the University of Tennessee, National Institute for Computational Sciences. The systems are targeted for computational science applications, especially biomolecular simulations. Jeffrey Vetter, a computational science who splits his time between Georgia Tech and Oak Ridge National Laboratory, will be the principal investigator for the project, know as Keeneland.

The big winners on the vendor side are HP, who will build the Intel-based HPC systems, and (you guessed it) NVIDIA, who will provide the GPU hardware. The first deployment is slated for “early 2010” and will indeed contain NVIDIA’s next-generation Fermi GPUs. Although the initial systems will be sub-petaflop machines, according to the Keeneland project Web page, in 2012 the supercomputers will be updated to “the next-generation platform and NVIDIA accelerators” and are anticipated to deliver a peak performance of around two petaflops.

Windows 7 Brings GPU Computing API

This week’s debut of Windows 7 brings with it DirectX 11 and the associated DirectCompute API, a Microsoftian invention used to accelerate compute-intensive Windows applications on graphics processors. DirectCompute is essentially Microsoft’s answer to OpenCL for Windows. It is intended to be used in games and other consumer software to speed up multimedia algorithms via the considerable computational prowess of on-board GPUs. This leaves the CPU free to do more mundane tasks, like figuring out what word you’re now misspelling in your document.

Coincidental with the release of Windows 7, NVIDIA decided to remind us that its current crop of DirectX 10 GPUs already support DirectCompute, and its next-gen DirectX 11 Fermi chips will do likewise. Below is a 20 second NVIDIA demo of DirectCompute:

Although obviously NVIDIA didn’t mention it, AMD supports DirectCompute as well, and already has DirectX 11 smarts cooked into its silicon today. Not only that, but the aforementioned ATI Radeon HD 5870 outperforms any current NVIDIA hardware for traditional graphics apps pretty handily. By incorporating all the new GPGPU bells and whistles into Fermi, NVIDIA took a several month hit getting its new architecture to market.

By now it’s clear that the two GPU makers have opted for different strategies. With the CUDA architecture, NVIDIA went aggressively for GPGPU, anticipating that applications and markets for discrete graphics processors will fundamentally shift toward computing over the next several years. AMD took the more conservative approach by sticking more closely to ATI’s graphics roots and deciding time to market plus raw performance will win the day. Time will tell which vendor made the better choice.