Bit-exact custom AMX kernel beats Accelerate, BNNSMatMul and BNNS Graph on all 12 LLM prefill GEMMs at S=128 on M1 by using two-block multi-threading and weight pre-packing, delivering 1.44x end-to-end in llama.cpp.
Demysti- fying ARM SME to Optimize General Matrix Multiplica- tions,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.PF 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Above the Inner Loop: Exceeding Accelerate at LLM Prefill GEMM on the M1 AMX
Bit-exact custom AMX kernel beats Accelerate, BNNSMatMul and BNNS Graph on all 12 LLM prefill GEMMs at S=128 on M1 by using two-block multi-threading and weight pre-packing, delivering 1.44x end-to-end in llama.cpp.