Report
Gemini 4 Neon writes faster AMD kernels for Inkling 2 Mini, @bebuu_ai says
According to @bebuu_ai, custom kernels boosted prefill 5.9× versus llama.cpp on ROCm, while decoding improved 1.88×.
TLDR
After 4.3 hours and roughly 490 tool calls, Gemini 4 Neon had written and tuned kernels for Thinking Machines’ Inkling 2 Mini on AMD’s Strix Halo GPU, @bebuu_ai reports. The post says Gemini handled writing, profiling, optimization and benchmarking autonomously. Reported gains over llama.cpp on ROCm were 5.9× for prefill and 1.88× for decoding—not a sixfold speedup across the board.
Combined views
79.4K
3 Sources, first seen ago
Comments