NVidia Scientist Calls For Expanded Research Into Parallelism

PALO ALTO, Calif. — Expanded research is needed into techniques for identifying and preserving parallelism in chip applications, according to chip maker Nvidia's chief scientist.

Comparing the futures of general-purpose CPUs and graphics processors (GPUs), David Kirk told the 17th Hot Chips conference on the Stanford Univeristy campus here on Tuesday (Aug. 16) that a crisis looms in programming technology. He said this could not only blight the future CPU market but also bring an end to improvements in graphics performance despite continued improvements in GPUs.

"If we look at the situation of general-purpose CPUs," Kirk said, "we see architects moving rapidly to multithreading and multicore designs. But we don't see a lot more threads to run. What parallelism there may be in algorithms is often concealed or lost altogether in the programming process."

Kirk grounded his pessimism in the experiences of game developers trying to exploit new multicore CPU chips. "We are already seeing some games limited by CPU throughput. We can render images faster than the CPU can send us the data," Kirk said. "But when game developers try to use dual-core CPUs to help, we have seen virtually no benefit to the second CPU core. And if the developer doesn't clearly understand the interactions of the cores with the caches, we have seen the application actually run slower on a dual-core machine."

Kirk contrasted this situation against the entirely different structure inside the GPU. "Graphics has been called embarrassingly parallel," he said. "In effect, each stage in our pipeline, each vertex in the scene and each pixel in the image is independent. And we have put a lot of effort into not concealing this parallelism with programming."

This allows a GPU developer to simply add more vertex processors and shading engines to handle more vertices and more pixels in parallel, as process technology allows. "We are limited by chip area, not by parallelism," Kirk observed.

In comments made before his talk, Kirk pointed out that the GeForce GPU chip used in the Playstation 3 offers significantly more floating-point operations per second than the Cell processor it supports. Kirk's concern is that while GPUs could continue adding execution units to exploit the nearly unlimited parallelism of graphics tasks almost indefinitely, CPUs could already be near the end of scalability. Without more explicitly independent threads or more clearly independent data to work on, he warned, there would be nothing for additional threads or additional cores to do. And those opportunities for parallelism didn't appear to be forthcoming in general-purpose computing.

"Programmers don't understand how to extract what is there," he said. "We need universities to build parallel programming techniques into the undergraduate curriculum, not offer it as a graduate class."

Added Kirk. "Research in parallel languages has stagnated. The work these days is on Java virtual machines and Web programming. We have enough of that now."

One reason CPU designers have turned to parallel architectures has been the crisis in power, he said. "The impact of power on architecture will force architectures to become more parallel," Kirk said. "If the threads aren't there to exploit the hardware, the market will value a multicore CPU less than it values a single-core chip."

That, the Nvidia scientist added, was an outcome that even the CPU vendors couldn't accept.