AMD CTO Predicts 'Attack Of The Killer Micros'

Hester's appearance at ICCAD closely followed AMD's Oct. 25 announcement of its Fusion program, in which AMD will deliver a wide range of merged x86 CPUs with on-board graphics accelerators starting in late 2008. It follows AMD's recent acquisition of graphics specialist ATI Technologies.

The first "attack of the killer micros," Hester said, occurred in the early 1990s, when microprocessor performance began to soar. As a result, he said, microprocessors offer one-tenth the performance at one-hundredth the cost of supercomputers, and provide more productivity.

To understand the upcoming "round two," Hester said, it's important to examine GPU history. He noted that GPUs have moved from simple wire-frame rendering in the 1980s to 32-bit parallel processing engines today. GPUs are moving beyond rendering into dynamic computing, Hester said, offering performance as high as 150 to 250 Gflops/second.

In recent years, Hester said, GPUs have come to offer more performance than CPUs. The gap is around 50x today, he said. "The questions are how to keep CPUs going at the rate they've been going, and also how to put GPU functions on the die," Hester said.

For an example of GPU capability, Hester pointed to Stanford University's Folding@Home project, in which people throughout the world download and run software to speed up protein folding research. 3 years of computation can be reduced to one month on an ATI GPU, he said.

In another decade, Hester said, GPUs will help push supercomputers to the petaflop level. A petaflop is equal to a million billion floating point operations per second.

"A step increase in microprocessor performance per watt per dollar is needed," Hester said. But simply adding more homogeneous CPU cores to a baseline architecture isn't good enough, he said. The solution, he said, is to adopt a heterogeneous architecture with GPU/CPU silicon-level integration. And that's the idea behind the Fusion initiative, which Hester called the most significant evolution in x86 architecture since the "Hammer" in 1999.

Hester noted, however, that there are two significant design challenges in developing such architectures — power management and memory hierarchy. Verification will also be a challenge, he acknowledged.

Since ICCAD is oriented to design automation development, Hester had some comments about how these next-generation architectures will impact CAD tools. He said that multi-core CPUs will drive the need for multi-threaded applications; the integration of GPUs will allow CAD tools to exploit single instruction, multiple data (SIMD) capabilities; and that "professional-grade CAD" at home will be fueled by a massively distributed grid paradigm, as in the Folding@Home project.

"There will be hundreds of millions of systems a year designed to have teraflop-level performance," Hester said. "True supercomputing on a desktop will bring about a new era in high-performance computing."