GRVI Phalanx Update: Plowing The Cloud With Thousands Of RISC V Chickens
summary
A reconfigurable overlay of simple RISC-V cores and a hoplite NoC runs as an FPGA accelerator fabric and compiles C++/GCC kernels into many-core jobs without full bitstream flow. Implementations run on Xilinx UltraScale+ (AWS F1), with a 1,680‑core prototype using 26 MB scratchpad SRAM at ~250 MHz and a plan for ~10,000 cores per large instance (1680 cores, 26 MB SRAM, 250 MHz). Measured fabrics show massive on-chip bandwidth and low power per core, and the system already uses AXI bridges, message/RDMA gateways, SDAccel/OpenCL tooling and bare‑metal C++ for validation (2.5 TB/s memory bandwidth, ~1 Tb/s NoC, 25 mW/core).
AI-written summary · about
chapters (11)
- 0:00Introduction to GRVI Phalanx
- 1:04Accelerator kit overview
- 2:09GRVI processor design
- 3:33Cluster architecture
- 4:15Hoplite NoC details
- 5:17Amazon EC2 F1 deployment
- 6:241680-core FPGA design
- 7:29Software developer kit
- 8:54F1 system layout
- 10:19OpenCL programming model
- 11:43Summary and future
related talks
18mRISC-V Performance Delivered
20mCould RISC-V Become a Force in HPC?
14mRISC-V Grows Up and Goes Big!
15mRISC-V CPUs for the New Era of Heterogeneous Computing
21mKeynote: Building Data Center Scale SoC's using RISC-V at Meta
10mDemo: CORE-V CVA6 Platform - Vendor Neutral, Open Source, 64bit RISC-V Compliant Core for Industry Software Testing & Validation
19mAccelerating Edge AI with the Help of RISC-V
37mKeynote: RISC-V Launchpad Showcase
