via MarkTechPost
MoonMath AI Open-Sources HIP Attention Kernel for AMD MI300X: Outperforms AITER v3 Across All Shapes and Rounding Modes
In June 2026, the MoonMath AI team open-sourced a bf16 forward attention kernel optimized for AMD’s MI300X GPU. Written entirely in HIP (Heterogeneous Interface for Portability) rather than hand-tuned assembly, the kernel is available under the MIT license. According to the team, their implementation consistently outperforms AITER v3—AMD’s own optimized kernel—on every shape and rounding mode tested. The bare-metal access required for benchmarking was provided by HotAisle, an AMD cloud provider. This release highlights the growing maturity of AMD’s GPU ecosystem for AI workloads, offering developers a high-performance, open-source alternative for attention operations in 2026 and beyond.