Skip to content

Introduction

WAVE is a universal GPU instruction set architecture. Kernels written in Python, Rust, C++, or TypeScript compile to a common .wbin binary. Backends translate that binary for Apple, NVIDIA, AMD, and Intel GPUs.

CUDA, Metal, HIP, and SYCL use different toolchains. WAVE defines shared operations and an execution model that each backend maps to its target. Hardware parameters, such as wave width, remain explicit.

The cross-vendor analysis identifies 11 categories of shared operations:

CategoryOperations
Integer arithmeticAddition, subtraction, multiplication, division, and remainder
Floating-point arithmeticAddition, multiplication, division, and fused multiply-add
Bitwise operationsBoolean operations, shifts, and bit counting
Comparison and selectionComparisons, predicates, and conditional selection
Local memoryAccess to workgroup scratchpad memory
Device memoryAccess to device-wide memory
Atomic operationsIndivisible read-modify-write operations
Wave operationsShuffles, reductions, and ballots
Control flowStructured branches, loops, and calls
SynchronizationBarriers and scoped fences
Type conversionConversions between numeric types

Special registers expose thread and workgroup identifiers, dimensions, and other execution parameters. See the instruction reference for mnemonics and operands.

Source kernel
|
wave-compiler
|
WAVE binary
|
Backend translator
|
Vendor compiler and runtime
|
GPU

The compiler produces WAVE binaries. The assembler wave-asm also produces binaries from assembly, and wave-dis disassembles them. The CPU emulator wave-emu executes binaries for functional testing.

BackendTargetGenerated code
wave-metalApple GPUsMetal Shading Language
wave-ptxNVIDIA GPUsPTX
wave-hipAMD GPUsHIP C++
wave-syclIntel GPUsSYCL C++

GPU dispatch depends on the SDK and installed vendor runtime. The SDK guides describe the available execution paths.

Instructions occupy one or two 32-bit words. The encoding records the operation, operands, and modifiers. The specification defines the instruction fields, register model, and binary layout.

Memory operations and fences use four scopes: wave, workgroup, device, and system. Scope determines which participants an ordering operation covers. The memory model guide explains how to combine memory access, barriers, and fences.

The preprint presents the cross-vendor analysis behind WAVE. Hardware verification includes Apple M4 Pro, NVIDIA T4, and AMD MI300X. The hardware verification page describes the matched Apple and NVIDIA benchmark study and its measurement scope.

Continue to Installation to build the toolchain.