GRXGPU G100

GRXGPU G100 is an open-source RISC-V GPU built on the Vortex architecture. It is a GPU — not a GPGPU — with hardware tensor cores, warp-group matrix multiply-accumulate (WGMMA), a data transfer accelerator (DXA), and a graphics fixed-function pipeline.

GRXGPU extends Vortex's RISC-V ISA with GPU-native instructions: WGMMA for dense matrix math, DXA for asynchronous tile fetches from global memory into shared memory, and structured sparsity support. The result is a GPU that can run tensor-core GEMM, ray tracing, and graphics workloads on a configurable RISC-V core mesh.

Architecture

• Core mesh: N independent RISC-V cores, each a full in-order pipeline with hardware threads
• Tensor cores (TCU): per-core WGMMA units supporting fp32, fp16, bf16, tf32, fp8, bf8, and 2:4 structured sparsity
• DXA: hardware data transfer accelerator for asynchronous 2D tile fetches (global → shared memory)
• Graphics pipeline: rasterizer, texture units, output mergers
• Ray tracing: BVH traversal, ray-box and ray-triangle intersection
• Memory: per-core LMEM, optional L1/L2/L3 caches, DSMEM for cross-core communication
• Command Processor: hardware-accelerated kernel dispatch and management

Supported Formats

Format	Input	Accumulator	Status
fp32	✅	✅	Verified
fp16	✅	✅	Verified
bf16	✅	✅	Verified
tf32	✅	✅	Verified
fp8	✅	✅	Verified
bf8	✅	✅	Verified
int8	✅	✅	Verified

Performance

Benchmark	Config	Result
SGEMM K=512 (64×64)	fp16 → fp32	IPC 1.752, 275K cycles
SGEMM K=512 (64×64)	bf16 → fp32	IPC 1.752, 275K cycles
TGM FSM K=512	67× instruction reduction vs SW K-loop	7K vs 483K instrs
SGEMM 1024³	G100 config	PASSED, 111M cycles

Directory Structure

• hw/ — RTL hardware sources
• sw/ — Software: kernel, runtime, drivers
• sim/ — Simulators (SimX, RTL)
• tests/ — Regression tests and benchmarks
• docs/ — Documentation and design proposals
• ci/ — Continuous integration scripts
• VX_config.toml — Hardware configuration (single source of truth)

Quick Start

Prerequisites

• Ubuntu 22.04 or compatible
• RISC-V GNU toolchain (installed via ci/toolchain_install.sh)

Build

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=$HOME/tools
./ci/toolchain_install.sh
make -s
make install
export VORTEX_PATH=$(pwd)/install
export PKG_CONFIG_PATH=$VORTEX_PATH/lib/pkgconfig:$PKG_CONFIG_PATH

Run a test

# SGEMM with tensor cores (fp16 → fp32)
cd tests/regression/sgemm_tcu_wg_dxa
make
LD_LIBRARY_PATH=$VORTEX_PATH/lib ./sgemm_tcu_wg_dxa -m 64 -n 64 -k 512

# bf16 SGEMM
cd tests/regression/sgemm_tcu_bf16
make
LD_LIBRARY_PATH=$VORTEX_PATH/lib ./sgemm_tcu_bf16 -m 64 -n 64 -k 512

Run on RTL simulator

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

Configuration

Edit VX_config.toml to change hardware parameters:

VX_CFG_NUM_CORES = 128        # Number of RISC-V cores
VX_CFG_NUM_WARPS = 4          # Warps per core
VX_CFG_NUM_THREADS = 4        # Threads per warp
VX_CFG_ISSUE_WIDTH = 4        # Pipeline issue width
VX_CFG_EXT_TCU_ENABLE = true  # Tensor core unit
VX_CFG_EXT_DXA_ENABLE = true  # Data transfer accelerator
VX_CFG_TCU_WGMMA_ENABLE = true # Warp-group MMA

After editing, re-run ../configure from your build directory to propagate changes.

Compiler Toolchain

GRXGPU uses VOLT (Vortex-Optimized Lightweight Toolchain), an LLVM-based SIMT compiler. See the VOLT repo for build instructions.

Documentation

• Design proposals — architectural proposals and experiment results
• Reply to grxcp team — multi-core bug analysis
• Debugging guide — runtime trace and debug options
• Simulation backends — SimX, RTL, FPGA

Citation

If you use GRXGPU in your research, please cite the original Vortex paper:

@inproceedings{10.1145/3466752.3480128,
  author = {Tine, Blaise and Yalamarthy, Krishna Praveen and Elsabbagh, Fares and Hyesoon, Kim},
  title = {Vortex: Extending the RISC-V ISA for GPGPU and 3D-Graphics},
  year = {2021},
  publisher = {Association for Computing Machinery},
  doi = {10.1145/3466752.3480128},
}

License

Apache License 2.0