Technology

What's built, what's verified, what's next.

Two chips in the open: the GRX930 RISC-V SoC with dual NPUs, and the G100 GPU. A photonic tensor accelerator is in design, and GRXCP ties them together with a CUDA-style programming model. Every row below says how far it has been proven.
GRIDWORKZ PLATFORM

Select a block to view its source, docs and status →

Specifications

Specs matrix

Verified in simTiming closedIn progressDesign
BlockParameterSpecificationStatus
GRX930CPU4 × RV64IMAC, 5-stage in-order, AMO/LR/SCVerified in sim
GRX930CachesPer-core I/D caches (256-bit lines); shared 4-way write-back L2 with invalidation coherenceVerified in sim
GRX930NPU2 × 8×8 systolic GEMM, DMA, command queue (depth 4), sigmoid/tanh activationVerified in sim
GRX930NPU datatypesINT4, INT8, INT16, FP16, BF16 (18/18 sweep cases pass)Verified in sim
GRX930Fabric4-master AXI4 crossbar, 64-bit data; DMA arbiter with equivalence gateVerified in sim
GRX930FPGAArtix-7 xc7a200t (Nexys Video): ~121 MHz post-implementation, 0 failing endpointsTiming closed
GRX930FPGA utilization125,735 LUTs (94%), 67,007 FFs, 55 BRAM, 227 DSPTiming closed
GRX930 test chipProcessGF180MCU 180 nm via wafer.space: single core, 50 MHz, boots from on-chip ROM. GDS due Dec 16, 2026In progress
G100ArchitectureVortex-based RISC-V GPU: tensor cores (WGMMA), DXA, ray tracing, graphics pipelineVerified in sim
G100Tensor formatsfp32, tf32, fp16, bf16, fp8, bf8, int8Verified in sim
G100Discovery config1 cluster × 4 cores, TCU + DXA; SGEMM K=64/512 pass vs host reference, pending independent reviewIn progress
PTAPhotonic tile128 × 64 GEMM tile, planned as a chiplet beside the G100; error-modelled against a frozen MNIST canaryDesign
SoftwareGRXCPCUDA-style runtime (libgrxrt), single-source compiler (grxcc), grxBLAS, grx-smi / grx-prof / grx-sanitizeIn progress
VerificationToolsIcarus Verilog, SymbiYosys formal, Verilator; G100 on SimX and RTL simulationVerified in sim

Sources: grx930, grxgpu and grxcp on GitHub.

Software

GRXCP: the GRX Compute Platform.

A CUDA-class programming platform for the G100 and GRX930. Write __global__ kernels and <<<grid, block>>> launches in one file; grxcc splits and compiles it, and libgrxrt runs it on top of the G100 driver.

  • libgrxrt runtime: grxMalloc, grxMemcpy, streams, events
  • grxcc single-source compiler driver, built on the VOLT LLVM toolchain
  • grxBLAS and friends, including an int8 GEMM route for the PTA
  • grx_cuda_compat.h maps cudaX to grxX, so porting CUDA is one include

In active development and tested on the G100 simulators.

// From grxcp/tests/grxcc/vecadd.grx.cpp
#include <grx/grx.h>

__global__ void axpy(const float* x, float* y,
                     unsigned int n, float a) {
  const unsigned int i = blockIdx.x * blockDim.x + threadIdx.x;
  if (i < n) y[i] = a * x[i] + y[i];
}

// Host side, same file
grxMalloc((void**)&dx, n * sizeof(float));
grxMalloc((void**)&dy, n * sizeof(float));
axpy<<<grid, block>>>(dx, dy, n, alpha);
grxDeviceSynchronize();
grxMemcpy(got.data(), dy, n * sizeof(float), grxMemcpyDefault);