AMD Turns Graphics Chip Into High-Performance 'Stream' Processor
AMD adapted an ATI graphics processor for floating point intensive computing. The stream processor will have 48 floating point cores and offer a peak performance of 360 gigaflops.
AMD on Tuesday unveiled a dedicated "stream" processor aimed at the high-performance computing market.
AMD built the processor by adapting an ATI graphics processor for compute-only systems such as workstations and servers, according to Dinesh Sharma, director of enterprise stream computing at AMD. "It was designed to light up pixels on the screen, but we can use those capabilities to solve problems or do financial analysis," he added. "We're using the [graphics processor] in a non-traditional manner."
The stream processor will have 48 floating point cores. Sharma noted that at peak floating point performance, AMD's dual-core processor runs in the 30 gigaflop range. The stream processor should hit 360 gigaflops.
A floating point is a number with a decimal point. Doing calculations with floating points are much more difficult and power consuming. Many scientists, animators, and movie producers base much of their work on floating point calculations.
Stream computing leverages massively parallel processors generally used to calculate and render millions of pixels onto computer monitors hundreds of times each second for 3D graphics applications.
"This is pretty darn clever," said Nathan Brookwood, principal analyst with Insight 64, a market research and consulting firm based in Saratoga, Calif. "This gives users the ability to really dramatically increase the floating point processing capability of a standard X86 based server. Historically, if you wanted to get this kind of performance, you had to buy great big supercomputers. Then AMD realized they had all this floating point ability that they were practically giving away in 3D chips that people were using for gaming. They took the chip, built a more robust platform, and surrounded it with more memory."
Brookwood said ATI, which was bought by AMD this fall, had been working on turning the gaming or graphics processor into a streaming processor for a few years. "ATI did all the work and AMD gets all the glory," he said, adding that this processor should outperform dual-core processors when it comes to floating point intensive tasks.
X86 chips in a dual-core configuration can handle floating core calculations, he explained, but they're not designed to do them at maximum speeds. "They're general purpose processors," Brookwood added. "For the enterprise, [streaming processors] will be important for business supercomputing. For example, if you're a Wall Street firm and your analysts and traders are using very complicated algorithms to figure out if a bond deal is going to be profitable, then they need more floating point capability than they can get with the fastest Opterons or Xeons today."
The streaming processor will be good for calculations, such as seismic migration analysis in the oil and gas industry, according to Sharma. They won't be as good, for instance, for tasks like searching Excel databases. "It's good if you have a lot of data parallelism where you do the same kind of operation on multiple pieces of data. That's where it shines."
According to AMD, hardware platform providers GraphStream Inc., Panta Systems, and Rackable Systems introduced new stream servers to the market on Tuesday. Stream servers use the new AMD Stream Processor and are aimed at the high-performance computing and enterprise markets.
Sharma said the streaming processors are ready for shipment.
"This is a big deal," said Brookwood. "This is going to give some folks doing scientific processing a way to get a lot more bang for their buck."
5 Top Federal Initiatives For 2015As InformationWeek Government readers were busy firming up their fiscal year 2015 budgets, we asked them to rate more than 30 IT initiatives in terms of importance and current leadership focus. No surprise, among more than 30 options, security is No. 1. After that, things get less predictable.