GPU PROGRAMMING USING RUST AND CUDA

C++ has been the go-to for GPU programming for almost 20 years. Can Rust do the job, and how well?

This book is all about getting hands-on with different toolchains that connect Rust to NVIDIA hardware. Theres RustaCUDA for safe host-side control, the Rust-CUDA project for writing kernels in pure Rust, and NVIDIAs experimental cuda-oxide compiler with its typed launches and async execution graphs.

Were going to build one Cargo workspace that keeps on growing. Itll include device queries, launch planning, Rust-written kernels, memory optimization, parallel reductions and scans, multi-stream pipelines, matrix multiplication benchmarked against cuBLAS, a Monte Carlo option pricer validated against a closed formula, and a complete batched inference application measured against a Python baseline. Well check every result against a CPU reference, and the reports will give accurate numbers, including where libraries outperform hand-written kernels and where experimental toolchains are still a work in progress.

Key Learnings

Launch, synchronize, and verify GPU kernels with ownership-managed device memory.

Write real CUDA kernels using Rust-CUDA and cuda-oxide.

Plan grids, blocks, and warps for 2D workloads.

Accelerate transfer speeds with pinned memory and coalesced access patterns.

Build race-free thread cooperation using shared memory, barriers, and atomics.

Overlap transfers with computation using streams, events, and async Rust pipelines.

Optimize matrix multiplication and benchmark against cuBLAS ceiling.

Wrap CUDA C library safely with handles, error enums, and Drop.

Ship complete batched GPU inference application against Python baselines.

Diagnose performance with Nsight Systems, Nsight Compute, and compute-sanitizer.

Table of Content

New Beneficiary of GPU Computing

Thinking in Threads

Commanding GPU

Writing GPU Kernels

Cleaner Kernels with cuda-oxide

Mastering GPU Memory

Making Threads Cooperate

Keeping GPU Busy

Delivering Real Math

Borrowing NVIDIAs Muscle

Shipping Complete GPU Application

Proving Performance