llama.cpp discussion #19480: CPU/server inference thread (EPYC 9454P, Ryzen HX 370 rows)

Fields

id

sources/llama-cpp-discussion-19480-cpu-server-inference-thread-epyc-9454

URL

https://github.com/ggml-org/llama.cpp/discussions/19480

Publisher

llama.cpp community

Kind

forum

Accessed

2026-10-15

notes

Community llama.cpp CPU/server inference thread. Rows cited by runaihome.com and the research-big-memory sweep: EPYC 9454P (12ch DDR5) 63.1 tok/s Qwen3-MoE-30B (~3B act), 11.76 tok/s Qwen3-Next-80B; Ryzen AI 9 HX 370 (dual-channel) 3.5 tok/s Qwen2.5-32B Q4 - the 'desktop trap' showing why channel count, not cores, gates CPU decode.

Notes

Community llama.cpp CPU/server inference thread. Rows cited by runaihome.com and the research-big-memory sweep: EPYC 9454P (12ch DDR5) 63.1 tok/s Qwen3-MoE-30B (~3B act), 11.76 tok/s Qwen3-Next-80B; Ryzen AI 9 HX 370 (dual-channel) 3.5 tok/s Qwen2.5-32B Q4 - the 'desktop trap' showing why channel count, not cores, gates CPU decode.

Cross-references

This record references no other records.