ICS2022

27 June 2022

• https://ics2022.github.io/
• https://meep-project.eu/form/workshop-on-risc-v-and-openpower
• https://meep-project.eu/events/3rd-workshop-risc-v-and-openpower-hpc
• https://libre-soc.org/openpower/sv/SimpleV_rationale/

Luke Leighton bio

Luke Kenneth Casson Leighton specialises in Libre Ethical Technology. He has been using, programming and reverse-engineering computing devices continuously for 44 years, has a BEng (Hons), ACGI, in Theory of Computing from Imperial College, and recently put that education to good use in the form of the Libre-SOC Project: an entirely Libre-Licensed 3D Hybrid CPU-VPU-GPU based on OpenPOWER. He writes poetry and has been developing a HEP Physics theory for the past 36 years in his spare time.

Coherent Distributed Computing: ZOLC, SVP64, OpenCAPI

Deterministic Scheduled Zero-Overhead Loops have a startling property: their deterministic nature allows them to be distributed. Extra-V began to illustrate the potential here, by performing near-Memory Coherent and conditional Graph walking, making full use of OpenCAPI's potential. Snitch also led the way, bringing back Auto-increment Load/Store from the CISC era, but hidden behind Tagged Registers connected to Coherent FIFOs leading indirectly to main Memory. Where both Snitch and Extra-V used limited variants of Deterministic Loops as proof-of-concept to support their overall research, with only rudimentary processing capability, ZOLC is a much more deeply extensive and well-defined Deterministic Loop Control system that can fit directly on top of a standard ISA.

SVP64 takes the Zero-Overhead Loop concept firmly into Supercomputing Vector Processing territory, currently limited to the register file. This talk explores the potential of combining Snitch and Extra-V's pioneering techniques, combining SVP64 and ZOLC, and leveraging OpenCAPI, on top of the OpenPOWER ISA, to create High-performance Coherent Distributed Computing with the potential to run large-scale Parallel Compute tasks at 100% sustained throughput whilst also bringing the potential of Snitch's 85% power-consumption-reduction to bear, using assembly intrinsics at in a normal everyday ubiquitous Software environment: no specialist parallel programming languages or special compilers needed.