News Imagination and Ventana to Build a RISC-V CPU-GPU Platform

Think Silicon (now owned by Synopsis) beat them to it, by almost a year and a half:


Details from the product page:

NEOX™ is a parallel multicore and multithreaded GPU architecture based on the RISC-V RV64IMFC instruction set with adaptive NoC. The number of cores varies from 4 to 64 organized in 1-16 cluster elements, each configured for cache sizes and thread counts. Depending on cluster/core configuration, NEOX™ compute power ranges from 12.8 to 409.6 GFLOPS at 800MHz with support for FP16, and FP32, and SIMD instructions.


Interestingly, both Imagination (UK) and Think Silicon (Greece) are European-based GPU designers. If you've never heard of the latter, that's because they've exclusively targeted the low-power embedded market.

I think RISC-V perhaps isn't ideal for a scale-up GPU, but it's alright for cases where you don't need the best performance and would rather provide more customizability to end customers and leverage existing codebases & toolchains.
 
Ventana Micro already has a RISC-V-based datacenter-grade Veyron V1 CPU IP (introduced in December 2022) that features an eight-wide execution and an up to 3.60 GHz frequency.
Has this ever been seen in the wild, or at least in actual silicon? Some of the features they said they're adding to the V2 make the V1 sound like it's not particularly usable in a true production environment.

As for this:

GKcSSExcjtaz8Vfrt83DwB-970-80.png.webp


I'll believe it when there are independent, 3rd party benchmarks on real hardware. We don't know what kinds of assumptions or modelling went into producing that estimate.

as Jon Peddie rightfully reminds us, a CPU is just a CPU as it needs graphics capabilities to display the results of its work.
Uh, no. Servers (Ventana's primary market) only need a BMC-type solution. You can just drop an ASpeed chip on the board - it doesn't particularly matter what ISA it uses, as there's zero code-sharing between it and the CPU.

It's only for client devices where the matter of integrated graphics becomes important. Even there, it's probably only when you get into the low cost/low power domain that it could start to make sense having flexible retasking of the GPU cores for CPU jobs.
 
Last edited:
  • Like
Reactions: P.Amini