EPYC 9454P (12ch DDR5) llama.cpp Qwen3-30B-A3B MoE decode
63.1 tok/s (single-stream)
Fields
id
benchmarks/comm-epyc-9454p-qwen3-moe-30b-decodeMetric
decode_tok_s
Value
63.1
Unit
tok/s (single-stream)
Model
Qwen3-30B-A3B (MoE ~3B active) Q4_K_M-class
Methodology
llama.cpp, single socket 12xDDR5 populated, --numa distribute; MoE so only ~3B active params read per token
Measured by
third_party
Thermal state
unknown