It's a kitten rather than a roar right now, but if the MIAOW project unveiled at last week's Hot Chips conference can get legs, the next year could see the launch of the world's first “open GPU”.
The result of 36 months' development (so far) by a team of 12 developers, MIAOW – the Many-core Integrated Accelerator of Wisconsin – is based on AMD's Southern Islands GPU ISA.
As Nicole Hemsoth writes over at The Register's HPC sister site The Platform, the GPU takes its subset of Southern Islands and adds OpenCL codes to get performance “comparable to existing single-precision GPU results”.
The AMD basis of the project means that MIAOW is going to be confined to the research community for some time. Project leader Dr Karu Sankaralingam told The Platform AMD's only input has come from individuals offering architectural insights, but at some point the boffins are going to have to talk to the chip vendor about its intellectual property.
The chip uses 95 vector, scalar, and memory instructions of the 400-plus available in Southern Islands.
The architecture supports 32 compute units connecting to the layer 2 cache. The cache connects to memory controllers and thence to the device memory, while a host processor handles GPU offload.
As their proof-of-concept, the GPU has been implemented on an FPGA.
At the project page on GitHub, they group notes that MIAOW has an immediate use: “MIAOW implements a compute unit suitable for performing architecture analysis and experimentation with GPGPU workloads. In addition to the Verilog HDL composing the compute unit, MIAOW also includes a suite of unit tests and benchmarks for regression testing,” they write.
“The Vertical Research Group believes that MIAOW can be a useful tool in producing not only more accurate quantitative results when benchmarking GPGPU workloads but also provide context for the architectural complexities of actually implementing newly proposed algorithms and designs that are intended to improve performance or other desired characteristics.”