Ryzen AI 9 HX 370 dual-channel Qwen2.5-32B Q4 dense decode - the desktop trap
3.5 tok/s
Fields
id
benchmarks/comm-ryzen-ai-hx370-qwen25-32b-q4-decodeMetric
decode_tok_s
Value
3.5
Unit
tok/s
Model
Qwen2.5-32B Q4 (dense)
Methodology
THIRD-PARTY MEASURED (llama.cpp discussion #19480, cited via runaihome EPYC piece). Strix Point laptop-class chip, dual-channel LPDDR5x only: a dual-channel desktop/mobile part lands ~3.5 t/s on a 32B dense Q4 - recorded as the negative datapoint showing why channel count, not core count, sets CPU decode.
Measured by
third_party
Reproducible
yes
notes
NOTE on accelerator_ids: no HX 370 accelerator record exists in this repo; the field points at the EPYC record it was contrasted with in the source article, which is wrong-but-searchable. If an hx370 record is created, rewire this id. Exists primarily so the 'dual-channel is a trap' claim has a number.
Notes
NOTE on accelerator_ids: no HX 370 accelerator record exists in this repo; the field points at the EPYC record it was contrasted with in the source article, which is wrong-but-searchable. If an hx370 record is created, rewire this id. Exists primarily so the 'dual-channel is a trap' claim has a number.