MoonMath AI выпустила open-source HIP-ядро для AMD MI300X, обогнавшее официальный AITER v3

MoonMath AI has open-sourced a HIP kernel for attention computation on AMD MI300X GPU — and in every test configuration, it proved faster than AMD's official AITER v3 library.

Why This Matters

AMD MI300X is AMD's flagship AI accelerator and a direct competitor to NVIDIA H100. Despite comparable specifications in memory and bandwidth, the MI300X has historically fallen behind NVIDIA in the software ecosystem: there was a lack of mature kernels optimized for real inference tasks. AITER (AI Inference and Training Engine Routines) is AMD's official library for such optimizations. This is the baseline against which MoonMath's new kernel compares itself.

How the Kernel Works

The authors applied two key techniques:

• Single-line ASM wrappers — minimal wrappers around AMD assembly instructions that eliminate unnecessary levels of abstraction and give the compiler fewer reasons to generate suboptimal code.
• Eight-wave pipeline — a scheme in which eight wavefronts (the AMD equivalent of warps in NVIDIA) work in parallel and hide memory latency with computation.

Together, these techniques maximize utilization of MI300X matrix blocks without idle cycles. The result is superiority over AITER v3 across all tensor shapes and all rounding modes, which is especially important for quantized inference.

What "Open-Source" Means Here

MoonMath AI released the code under an open license. This means any developer or company can:

• integrate the kernel into their own inference stack on AMD
• study the optimization techniques and apply them to other kernels
• fork and adapt it for specific models or batch sizes

Such releases accelerate AMD's ecosystem faster than internal vendor roadmap releases: the community sees real code, not marketing benchmarks.

What This Signals

A third-party team publicly outperformed AMD on its own hardware — a market signal that MI300X can compete with NVIDIA H100 when equipped with proper low-level optimizations. For infrastructure teams currently choosing between AMD and NVIDIA, there is now another compelling argument in favor of the red camp.