Ryzen AI 9 HX 370 (dual-channel) Qwen2.5-32B Q4 decode
3.5 tok/s decode, single-stream
Fields
id
benchmarks/comm-ryzen-ai-9-hx370-qwen25-32b-q4-decodeMetric
decode_tok_s
Value
3.5
Unit
tok/s decode, single-stream
Model
Qwen2.5-32B Q4 dense
Methodology
llama.cpp discussion #19480 measurement; dual-channel LPDDR5X Strix Point laptop/mini-PC
Measured by
third_party
Thermal state
unknown
notes
The desktop-CPU trap: dual-channel ~90-120 GB/s effective lands ~3.5 tok/s on 32B Q4 - 4x slower than a 12-channel Genoa, despite more cores than needed.
Notes
The desktop-CPU trap: dual-channel ~90-120 GB/s effective lands ~3.5 tok/s on 32B Q4 - 4x slower than a 12-channel Genoa, despite more cores than needed.