Another Look at DSPs for High Performance Computing
With all the talk about energy-efficient flops for high performance computing, digital signal processors could end up being a compelling alternative to the current crop of GPUs and Intel manycore chips. Texas Instruments has been exploring that path in a low-key sort of way for about a year now. Last year, the company launched the TMS320C66, a new line of multicore DSPs built for 4G cellular base stations that they felt could also double as HPC accelerators.
Not much has been said about how those TI chips might fair with real HPC-type workloads. But this week, IEEE Spectrum ran an article describing how a researcher at the University of Madrid used an eight-core version of that processor to run a common matrix-matrix multiplication algorithm. The researcher in question, Francisco D. Igual, found that this particular chip delivered 7.4 gigaflops per watt running his matrix code. According to him, this DSP outperformed other supercomputer platforms.
However, Berkeley Lab’s John Shalf noted that these results were obtained with single precision floating point, which uses just 1/4 the energy of running double precision operations. Even taking that handicap into account thogh, a machine incorporating these DSPs could probably deliver double precision efficiencies on par that of IBM’s Blue Gene/Q, the current champ in performance-per-watt.
The Spectrum article noted that a couple of decades ago, TI produced a family of DSPs that were subsequently used to build a custom supercomputer for quantum chromodynamics applications. That machine, which was known as QCDSP (Quantum Chromodynamics on Digital Signal Processors), was designed by a group of physicists at Columbia University. One of the them was Alan Gara, who went on to work for IBM and became the chief architect of the Blue Gene line of supercomputers.
Building a real supercomputer with these latest DSPs would take some doing, of course. But PCIe-based evaluation kits are available from TI — or at least they were last year, when the DSPs were launched. And although the HPC software stack is rather thin, the company also released a free Linux-based SDK, including C and OpenMP support, that goes with the hardware. If any of this has aroused your interest, the evaluation kit and the SDK look like they’re still available on TI’s website.
Full story at IEEE Spectrum