WAVE - The Universal GPU ISA

WAVE is an open-source research project exploring whether a single GPU instruction set can target every major GPU vendor. The answer is yes. The specification, compiler, four vendor backends, and four language SDKs are all Apache 2.0 licensed.

Any GPU

WAVE defines 11 hardware-invariant primitives distilled from 5,000+ pages of vendor ISA documentation across 16 microarchitectures. One binary format targets Apple Metal, NVIDIA PTX, AMD HIP, and Intel SYCL through dedicated compiler backends. No vendor lock-in, no rewriting kernels per platform.

Any Language

Native SDKs for Python, Rust, C++, and TypeScript let you write, compile, and launch GPU kernels from the language you already use. Install with pip install wave-gpu, cargo add wave-sdk, CMake find_package, or npm install wave-gpu and start computing in minutes.

Proven

Verified on real hardware - Apple M4 Pro, NVIDIA T4, and AMD MI300X - with full spec-level compliance tests for every primitive. The research behind WAVE is published on Zenodo (DOI 10.5281/zenodo.19163452) and in preparation for ASPLOS 2027.

Quick Example

A complete GPU vector addition in Python, from kernel definition to result readback:

import wave_gpu

@wave_gpu.kernel(threads=256)
def vector_add(a, b, out, n):
    tid = wave_gpu.thread_id()
    if tid < n:
        out[tid] = a[tid] + b[tid]

# Allocate device arrays
n = 1024
a = wave_gpu.array([1.0] * n)
b = wave_gpu.array([2.0] * n)
out = wave_gpu.array([0.0] * n)

# Launch - WAVE detects your GPU and selects the right backend
vector_add(a, b, out, n)

print(out.to_list()[:4])  # [3.0, 3.0, 3.0, 3.0]

WAVE compiles the kernel to its portable .wbin binary format, selects the appropriate backend for your hardware (Metal on Apple, PTX on NVIDIA, HIP on AMD, SYCL on Intel), translates to vendor-native code, and dispatches to the GPU. If no supported GPU is detected, the built-in emulator runs the kernel on the CPU so you can develop anywhere.

How It Works

Source kernel → wave-compiler → .wbin → backend → vendor GPU code → GPU

The WAVE compiler lowers your kernel into a compact 32-bit base instruction format (64-bit extended for 3+ operands) using 32 general-purpose registers, 4 predicate registers, and structured control flow. Each backend then translates this portable IR into the vendor’s native ISA.

Next: Introduction - learn what WAVE is and why it exists.