MemPool Flavors: Between Versatility and Specialization in a RISC-V Manycore Cluster
summary
MemPool presents three implemented RISC-V manycore flavors with measured trade-offs: MemPool scales to 1024 PEs via hierarchical tiles (4 cores/tile, 16 tiles/group, 4 groups/cluster) and targets low L1 latency (1–5 cycles across levels); the systolic autonomous data flow adds per-bank queue hardware and load-refill logic to eliminate loads/stores and improves matmul by ~7% over baseline; the vectorial variant replaces tiles with a short-vector core (64 cores, 256 VFUs) and boosts matmul throughput ~60%, and a 64-core Harstream tapeout on GF 12nm already validates the approach.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 1:16Why scale up clusters
- 4:28Baseline MemPool architecture
- 6:35Systolic MemPool variant
- 8:19Vectorial MemPool variant
- 9:42Performance comparison results
- 11:48Trade-offs and ecosystem
related talks
19mNuclei System Technology Releases UX1030H with Full Support for RVA23
01h 15mThe Evolution of Memory Technologies
01h 28mProcessing-in-Memory: Theory and Practice
43mRISC-V CPU Project in Western Digital
21mVentana’s Second Generation RISC-V Processor for Data Center and Other High Performance Applications
17mRISC-V is Ready for Powering the Era of Intelligent General Computing
26mPrincipled algorithm, data structure and system design for Processing-in-Memory (PIM)”
22mMatrix Extensions for RISC-VDelivering on the Promise