← Back to Model Beat
Hardware·Jun 22·all news from June 22, 2026

MoonMath AI Open-Sources a HIP Attention Kernel for AMD MI300X That Beats AITER v3 on Every Shape and Rounding Mode

MoonMath AI has released an open-source HIP attention kernel designed to optimize performance on AMD MI300X hardware. By utilizing one-instruction assembly wrappers and an eight-wave pipeline, the implementation surpasses the efficiency of AMD’s own AITER v3 across all tested configurations and rounding modes.

Covered by 1 source

Related stories

HardwareOpenAI and Broadcom unveil LLM-optimized inference chipJun 24 · 11 sourcesHardwareFive Eyes intelligence alliance says frontier AI models could reshape offensive cyber ops in monthsJun 21 · 18 sourcesHardwareAmazon in Talks to Sell Custom AI Chips in Bid to Cut Nvidia’s DominanceJun 18 · 3 sourcesHardwareChina Makes Sweeping Education Reforms to Prepare for AI EraJun 22 · 4 sources