Advanced Computing in the Age of AI | Saturday, September 23, 2023

Researchers Take Wraps Off Open Source GPU 

The outlines of an open source hardware platform continue to come into focus with the introduction of what is claimed by university researchers to be the first general purpose graphics processor design.

In a white paper released during a semiconductor design conference last week in Silicon Valley, researchers from the University of Wisconsin-Madison outlined the architecture of the open source GPU dubbed MIAOW, or Many-core Integrated Accelerator of Wisconsin. The researchers described the GPU design as a hardware component of the open source MIOAW platform.

Specifically, they described the device during last week's Hot Chips conference as an open source implementation of AMD's Southern Islands general-purpose GPU (GPGPU) instruction set architecture. The researchers added that the device is capable of running unmodified OpenCL-based applications.

The prototype was essentially designed to demonstrate an open source GPGPU compatible with OpenCL, the standard computing language for cross-platform parallel programming. The researchers said their goal was to emulate a full system, not to compete with commercial designs.

MIAOW (pronounced, you guessed it, "me-ow") was designed by computer science researchers at the University of Wisconsin led by Karu Sankaralingam. With Moore's Law running out of steam, the team said it is attempting to turn chip design on its head by instead focusing on emerging applications "to provide an opportunity to think in a fundamentally different way about how we build processors."

Sankaralingam's team developed the MIAOW core over a three-year period after AMD released the instruction set architecture for its Southern Islands GPU. "The aim when comparing MIAOW to commercial designs was to show that our design is reasonable and that the quantitative results are in similar range," the researchers' noted in their white paper. "We are not quantifying accuracy since we are defining a new microarchitecture and thus there is no reference to compare to. Instead we compare to a nearest neighbor to show trends are similar."

MIOAW implements a subset of the AMD GPU instruction set. The processor design includes a host CPU that assigns a kernel to the GPGPU, the researchers said, and then computes kernel assignments. Among other things, the design includes a configurable number of compute units, a memory controller and a "cached memory hierarchy," the researchers reported.

The current design uses only 95 instructions and 32 compute units. The researchers are reportedly in the process of adding a graphics pipeline to the design.

The team reported that their design fared well in several benchmark tests that compared performance with AMD's current Tahiti GPU. It fell short, however, on other graphics chip benchmarks. Nevertheless, the researchers claim "MIAOW seems to be a balanced design."

While the GPU design represents little more than a working prototype, it does represent another attempt at leveraging open source tools in an ongoing effort to develop a new hardware platform. At the other end of the hardware scale, IBM announced a new Linux mainframe server in August along with several parallel Linux initiatives aimed at developing new software distributions for Linux servers.

About the author: George Leopold

George Leopold has written about science and technology for more than 30 years, focusing on electronics and aerospace technology. He previously served as executive editor of Electronic Engineering Times. Leopold is the author of "Calculated Risk: The Supersonic Life and Times of Gus Grissom" (Purdue University Press, 2016).