AMD's upcoming FireStream processor might be a way for scientists to tap into a lot of performance without breaking the bank.
The company will be demonstrating its FireStream 9170 processor next week at the , and executives spoke this week about the promise of "stream computing." The 9170 is designed to let high-performance computing applications take advantage of the excellent parallel performance of a graphics chip.
The big trend in chip design over the past few years has been parallelism. Instead of trying to crunch all the data through a single path moving as fast as possible, the cool kids are now adding paths so data can flow down multiple outlets. This allows the chip to run at slower speeds, and therefore .
Graphics processing units (GPUs) have been doing this for years. The high-performance discrete graphics chips from companies like Nvidia and AMD's ATI division have been designed with parallel performance in mind for a very long time. Certain types of customers in labs and research facilities would love to be able to tap into that kind of processing power, but GPUs require special programming techniques.
AMD is trying to bridge the gap between PC processors that are easy to program and graphics chips that offer great performance with the FireStream 9170. Think of it as a high-end graphics chip with a lot more memory than usually ships with those products, said Robert Feldstein, vice president of engineering for AMD.
The performance will be there. The 9170 is essentially one of ATI's high-end discrete graphics chips that has been tricked out with more memory and double-precision floating point units, which apparently is better than single precision. It comes with 2GBs of memory, compared with 512MBs of memory on the most powerful ATI graphics chip.
But the programming is still a little tricky. You'll need a software developer's kit, and you'll probably only want to port limited amounts of your code to run on the 9170.
"You don't have a researcher that's trying to port over thousands of lines of legacy code. They have a particular algorithm that (the researcher) knows will run well on a GPU," said Patricia Harrell, director of stream computing for AMD. "You're not worried about changing code for something that gives you an order of magnitude increase (in performance)," she said.
The 9170 isn't going to be out until the first quarter of next year, as AMD's graphics priorities for the holiday season are discrete graphics chips for PCs that all of us can use. It will cost $1,999, which might seem like a lot, but this is something you should be able to add into an existing workstation or server for a performance boost when you need it, rather than buying a fancy server for just a few lines of code.
Eventually, AMD wants to integrate this type of technology directly onto a PC or server processor. It has already announced plans to integrate graphics chips onto PC chips as part of its Fusion project, but it hasn't identified a timeframe for putting its powerful stream computing technology on a PC chip.