EPYC Genoa, two DIMMs, 34B Q4_0 decode (llama.cpp #3167)
4.0 tok/s, single-request llama.cpp CPU decode, two of twelve channels populated
Fields
id
benchmarks/llamacpp-epyc-genoa-two-dimm-34b-q4-decodeMetric
decode_tok_s
Value
4
Unit
tok/s, single-request llama.cpp CPU decode, two of twelve channels populated
Model
34B Q4_0 (family not named in the thread, ~34B dense)
Methodology
Community measurement reported in ggml-org/llama.cpp discussion #3167: the same Genoa-class EPYC box as the one-DIMM run after a second 64GB DIMM was added, still on a 34B Q4_0 model under llama.cpp. Decode rose from 1.5 tok/s to 4 tok/s. Exact model family, prompt length, context length and llama.cpp commit are not stated in the captured thread, so only the reported decode rate is recorded. Third-party forum measurement, not a vendor claim and not a controlled bench.
Measured by
third_party
Reproducible
no
Thermal state
unknown
notes
Two of twelve channels populated. 4 tok/s is ~2.7x the one-DIMM figure but still far below a fully populated socket — the 12-channel EPYC 9554 measures ~7 tok/s on the larger 70B Q4_K_M (comm-epyc-9554-llama33-70b-q4-decode), so even two sticks leave most of the bandwidth on the table. Recorded so the channel-population effect is a citable number rather than prose buried in gotcha-epyc-must-populate-all-memory-channels. Accelerator id is the representative Genoa part; the thread does not name the exact SKU.
Notes
Two of twelve channels populated. 4 tok/s is ~2.7x the one-DIMM figure but still far below a fully populated socket — the 12-channel EPYC 9554 measures ~7 tok/s on the larger 70B Q4_K_M (comm-epyc-9554-llama33-70b-q4-decode), so even two sticks leave most of the bandwidth on the table. Recorded so the channel-population effect is a citable number rather than prose buried in gotcha-epyc-must-populate-all-memory-channels. Accelerator id is the representative Genoa part; the thread does not name the exact SKU.