The chip is called a General Purpose General Processing Unit (GP GPU). It traces its origins to video cards, which process data sequentially -- in a straight stream, so to speak, hence the name stream processing. This processing approach also works well for certain types of mathematical computations. The GP GPU operates as a co-processor to a CPU -- an Opteron, for instance.
AMD released its first processor in this line, FireStream, a year ago. But the company said it didn't push the initial chip in a widespread way, and focused initially on attracting developers. But with the FireStream 9170, AMD says it is aiming at what it sees as a potentially large market, initially among high performance computing users and, in time, for general business users.
AMD is releasing a software development kit along with the processor, so developers can adapt their applications to work with the accelerator. It has also given the FireStream chip 64-bit double-precision floating point capability. AMD says it is the first chipmaker in the industry to introduce the greater capability.
Patricia Harrell, director of stream computing at AMD, believes that with the right kind of application, the FireStream chip can deliver up to a 10x performance gain without any special application tuning, and a greater gain with tuning.
FireStream technology may be appealing for high-performance computing (HPC) situations such as financial analysis, seismic processing in oil and gas exploration, or life sciences research, as well as in security contexts such as face-recognition or image-matching systems. "This technology has really broad uses across other markets," she said.
Not just throwing CPUs at speed problems
The approach that most HPC users take to improve application performance is to add more microprocessors, said Addison Snell, vice president Tabor Research in San Diego, an HPC market research firm. But he said HPC users are beginning to pay more attention to accelerators such as the GP GPU.
Snell noted, however, that there are challenges with effective GP GPU integration. One of those concerns programmability of applications running with an accelerator -- rather than writing things for a standard processor, developers are trying to get various computational elements to go off the usual chip to a different part of the computer that can do the processing faster, he said. "The other challenge is gaining enough acceleration from doing so that you overcome the latency ????????? the time lag from going off one chip on to another chip," he said.
The software development kit AMD is releasing along with the chip itself should help developers address those programming issues, said Snell.
AMD said its FireStream chip will be released in the first quarter and priced at $1,999.