OxiCUDA 0.1.2 Released — Pure Rust CUDA Toolkit Replacement

The GPU computing foundation of the COOLJAPAN ecosystem just became fully sovereign.

Today we released OxiCUDA 0.1.2 — a complete, production-grade pure Rust replacement for the entire NVIDIA CUDA Toolkit.

No CUDA SDK. No nvcc. No C/C++ toolchain. No build-time dependencies.
Only runtime dependency is the official NVIDIA driver (libcuda.so / nvcuda.dll).
Just clean, type-safe, memory-safe GPU code that compiles to a single static binary (or WASM) and runs on Turing through Blackwell GPUs — and across multiple vendor backends.

Why OxiCUDA 0.1.2 is a game changer

For decades, high-performance GPU computing meant depending on the massive NVIDIA CUDA Toolkit (with its complex SDK, C++ headers, and vendor lock-in).

These tools are powerful but suffer from:

• C/C++ memory unsafety and segfault risks
• Heavy build-time dependencies and nvcc requirements
• Vendor lock-in (NVIDIA-only)
• Poor portability to Metal, Vulkan, WebGPU, ROCm, or embedded
• Difficulty integrating with safe Rust scientific/AI stacks

OxiCUDA 0.1.2 ends all of that.

It delivers ≥90–95% of native CUDA performance while being 100% memory-safe and auditable.
Notable results (target on supported architectures):

• SGEMM (FP32): ≥95% of cuBLAS
• HGEMM (FP16): ≥95% of cuBLAS (Tensor-Core)
• FlashAttention: ≥90% of FA2
• FFT (power-of-2): ≥90% of cuFFT
• Convolution (FP16): ≥90% of cuDNN

Technical Deep Dive: How We Rebuilt the CUDA Toolkit in Pure Rust

OxiCUDA is organized into 10 volumes + 7 backends (28 crates total), mirroring the CUDA ecosystem while being radically simplified and optimized for Rust:

1. Foundation (4 crates)
   Dynamic driver loading, type-safe DeviceBuffer<T>, RAII memory management, launch! macro.

2. PTX Codegen + Autotuner (2 crates)
   Pure-Rust PTX IR DSL (SM 7.5–10.0), Tensor-Core WMMA/MMA/WGMMA, 3-tier runtime autotuner with disk cache.

3. Linear Algebra (1 crate)
   Full cuBLAS equivalent (GEMM, batched, reductions, elementwise).

4. Deep Learning (1 crate)
   cuDNN replacement: Conv (Winograd/direct/fused), FlashAttention, Norm layers, FP8/INT8/INT4 quant.

5. Scientific Computing (4 crates)
   cuFFT, cuSPARSE, cuSOLVER, cuRAND (Stockham FFT, SpMV/SpGEMM, LU/QR/SVD/Cholesky, Philox RNG).

6–10. Signal Processing, Computation Graph, GPU Training, Inference Engine, Reinforcement Learning
MFCC/STFT, CUDA Graphs, fused optimizers (Adam/AdamW), paged KV-cache, PPO/DQN/SAC, etc.

Backends (7 crates)
Native NVIDIA (via libcuda), Metal, Vulkan, WebGPU, ROCm, LevelZero + generic trait.

Key Rust advantages:

• 100% Pure Rust (zero C/C++/CUDA SDK at build time)
• Type-safe kernels and memory management
• Built-in autotuner selects optimal kernel per GPU
• Full integration with SciRS2, OxiBLAS, OxiONNX, OxiLLaMa

What’s inside 0.1.2 (released April 15)

• Stability improvements and CI hardening
• Expanded backends and autotuner refinements
• Full production readiness of Volumes 1–10
• Production readiness confirmed through 7,263 passing tests across 28 crates
• 253,125 lines of pure Rust — zero warnings, Clippy0 + fail0 enforced

This is the foundation

OxiCUDA is now the official GPU computing backend for the entire COOLJAPAN stack (total ecosystem: 21M+ SLoC Rust, 597 crates, 40+ production-grade libraries):

• SciRS2 / NumRS2 — all GPU-accelerated scientific computing
• OxiBLAS / OxiFFT — linear algebra and signal processing
• OptiRS / OxiLLaMa — training and LLM inference
• OxiMedia / VoiRS — real-time media and speech pipelines
• OxiGDAL / Spintronics / OxiHuman — geospatial, physics, and biomechanical GPU simulations
• ToRSh / OxiRAG — high-throughput tensor and RAG operations

Repository: https://github.com/cool-japan/oxicuda

Star the repo if you want GPU computing without the CUDA Toolkit or C++ toolchain.

The era of “just install the CUDA SDK” is over.

Pure Rust GPU computing is here — fast, safe, portable, and sovereign.

— KitaSan at COOLJAPAN OÜ April 15, 2026