llama.cpp+ZenDNN on EPYC 9755 128c: Mixtral-8x7B-Instruct-v0.1 gguf-q8-0 prefill tok/s (512in/128seq)
954.14 tok/s prefill (llama-batched-bench pp512, 128 parallel seq, aggregate)
Fields
id
benchmarks/amd-epyc9755-zendnn-mixtral8x7b-q8-pp512Metric
prefill_tok_s
Value
954.14
Unit
tok/s prefill (llama-batched-bench pp512, 128 parallel seq, aggregate)
Model
Mixtral-8x7B-Instruct-v0.1
Methodology
VENDOR MEASURED (AMD developer blog, ZD-066/067, Aug 18 2026). AMD EPYC 9755 128-core single socket (one NUMA node pinned, SMT off, 4.06GHz boost, DDR5-6400 12-ch); llama.cpp build b9326, ZenDNN 5.2.2 backend (-DGGML_ZENDNN=ON), ZENDNNL_MATMUL_ALGO=AOCL-DLP, llama-batched-bench 512in/128out/128 parallel seq, flash-attn on, KV cache = model precision, 3-run mean. Mixtral-8x7B-Instruct-v0.1 (gguf-q8-0): ZenDNN 954.14 vs native-CPU 209.84 tok/s prefill. AGGREGATE over 128 seq, not per-request.
Measured by
vendor
Reproducible
yes
notes
Native-CPU baseline 209.84 tok/s -> ZenDNN 954.14 tok/s (4.55x). Vendor-published, single-socket pinned. Decode falls back to CPU path on ZenDNN 5.2.2 (small-batch not yet accelerated).
Notes
Native-CPU baseline 209.84 tok/s -> ZenDNN 954.14 tok/s (4.55x). Vendor-published, single-socket pinned. Decode falls back to CPU path on ZenDNN 5.2.2 (small-batch not yet accelerated).