Skip to content

G100 GPU

The G100 is an open-source RISC-V GPU built on the Vortex architecture. Source: kubeworkz/grxgpu.

Architecture

  • Core mesh: independent RISC-V cores, each an in-order pipeline with hardware threads
  • Tensor cores (TCU): per-core warp-group matrix multiply-accumulate (WGMMA), with 2:4 structured sparsity
  • DXA: data transfer accelerator for asynchronous 2D tile fetches from global to shared memory
  • Graphics: rasterizer, texture units, output mergers
  • Ray tracing: BVH traversal, ray-box and ray-triangle intersection
  • Memory: per-core local memory, optional L1/L2/L3, DSMEM across cores
  • Command processor: hardware kernel dispatch

Tensor formats

fp32, tf32, fp16, bf16, fp8, bf8 and int8 are verified as both inputs and accumulators.

Status

The current discovery configuration is one cluster of four cores with TCU and DXA enabled. Tensor-core SGEMM (sgemm_tcu_wg_dxa, K = 64 and K = 512) passes against a host reference on the SimX and RTL simulators, pending independent review. Full-chip FPGA closure, the 128-core flagship configuration and any ASIC process are not established yet. See docs/grxgpu_g100_product_brief.md in the repo for the evidence ledger.

Build and run

Terminal window
git clone --depth=1 --recursive https://github.com/kubeworkz/grxgpu.git
cd grxgpu
sudo ./ci/install_dependencies.sh
mkdir build && cd build
../configure --xlen=32 --tooldir=$HOME/tools
./ci/toolchain_install.sh
make -s

Then run a tensor-core SGEMM, or a test on the RTL simulator:

Terminal window
./ci/blackbox.sh --cores=1 --app=vecadd --driver=rtlsim

The grxgpu README has the full steps, including installing and running the regression tests.

Programming it

Use GRXCP for CUDA-style kernels, or Vortex’s runtime and the VOLT compiler directly.