Get 30 times Folding@Home speedup for $450

ATI IS STARTING to talk about Stream Computing in a serious way. Nvidia has its Gelato and was an early proponent of the concept, but ATI is showing that it is far from behind.

Last Friday, ATI showed off several examples of how Stream Computing would fit in to the real world.

Stream Computing uses the GPU as a compute engine rather than strictly for graphics. If you have a problem set that maps to a GPU, it will absolutely scream when compared to a plain old CPU. If you look at the architecture of a GPU, it is a bunch of shaders, tens of them in most cases, each of which can crunch a lot of heavy sums, or at least certain types of heavy sum.

This leaves developers with a problem. How do you split code between the CPU and the GPU? If you do it wrong, you end up with a lot of data shuffling back and forth with little actual work getting done. Do it right, and data gets streamed in, worked on, and streamed out with extraordinary levels of throughput. That is the notion that Dave Orton had when he presented Stream Computing.

There are a bunch of target markets for this type of functionality, hardcore science, climate/weather research, homeland security, advanced stock intelligence/risk assessment, seismic modeling and search. None are what you would call light duty tasks, and most are very time dependent. It does you no good to identify a bad guy with facial recognition after he gets on the plane, or a hot stock option after trading closes for the day.

Stream computing can bring between 10 and 40 times speed boosts to your application if, and this is a very big if, your app will support it. If it does not map well, or is poorly coded, you can lose performance. ATI showed off four companies that use the technology either right now, or will, very soon.

From Left to Right: Chas Boyd of Microsoft, Jeff Yates of Havok, Vijay Pande of Stanford University, Michael Mullaney of Peakstream, and Dave Orton of ATI.

First was Vijay Pande of Stanford University, one of the people behind Folding@Home. Running Folding@Home on a GPU, specifically an X1900 class card. It is seeing between a 20-40 times speedup depending - not least how fast a CPU can feed the card. This would mean a GPU can do what most of a rack of servers can.

Peakstream is a company that makes a middleware layer and tools to facilitate Stream Computing. Michael Mullaney, its VP of marketing, showed off how its tools would allow you to use GPUs for everything from stock trading to oil and gas exploration. The biggest advance it has is a profiler that can peer into the code and point out thrashing and bottlenecks. It provides the API, you use it for whatever code you want.

Chas Boyd, Architect, Graphics Platform Unit at MS was up next to talk about how DX10 and Vista would use many of these stream techniques. Vista goes a lot farther than most of its predecessors by using a GPU to accelerate the desktop, at least in Aero Glass.

Jeff Yates of Havok has the killer app for Stream Computing, physics on a GPU. The demos were nothing they didn't show at Computex, boulders down a hill, tornados whipping debris around, and a new castle demo/game with destructible castles. "Physics" has mindshare right now and this will be by far the most widespread use of the technology.

While no products are yet available, the general feeling is that Stream Computing techniques and technology are here to stay. Early examples promise speedups of tens of times over the fastest CPU out there at a fraction of the cost. If your code fits, Stream may be a very good thing. µ