This article shows the CPU-only inference with a modern server processor – AMD Epyc 9554. For the LLM model the Alibaba’s Qwen3 32B with different quantization are used to show the difference in token generation per second and memory consumption. Qwen3 32B is a pretty solid LLM model, which is a Causal Language Model type and offers significant advancements in reasoning and causality. It is published by the Alibaba giant, and in many cases it can be considered to offload some LLM work locally for free (especially Maths and IT). The article is focused only showing the benchmark of the LLM tokens generations per second and there are other papers on the quality of the output for the different quantized version.
The testing bench is:
- Single socket AMD EPYC 9554 CPU – 64 core CPU / 128 threads
- 196GB RAM in 12 channel, all 12 CPU channels are populated with 16GB DDR5 5600MHz Samsung.
- ASUS K14PA-U24-T Series motherboard
- Testing with LLAMA.CPP – llama-bench
- theoretical memory bandwidth 460.8 GB/s (according to the official documents form AMD)
- the context window is the default 4K of the llama-bench tool. The memory consumption could vary greatly if context window is increased.
- More information for the setup and benchmarks – LLM inference benchmarks with llamacpp and AMD EPYC 9554 CPU
Here are the results. The first benchmark test is Q4 and is used as a baseline for the diff column below, because Q4 are really popular and they offer a good quality and really small footprint related to the full sized model version.
| N | model | parameters | quantization | memory | diff t/s % | tokens/s |
|---|---|---|---|---|---|---|
| 1 | Qwen3 32B | 32B | Q4_K_M | 18.40 GiB | 0 | 13.228 |
| 2 | Qwen3 32B | 32B | Q5_0 | 21.61 GiB | 11.27 | 11.736 |
| 3 | Qwen3 32B | 32B | Q6_K | 25.03 GiB | 10.08 | 10.552 |
| 4 | Qwen3 32B | 32B | Q8_0 | 32.42 GiB | 17.64 | 8.69 |
The difference between the Q4 and Q8 in the tokens per second is 34.30% speed degradation and even the Q8 is usable with tokens generation between 8-9 per second. Around 14 tokens per second is good and usable for daily use for a single user, which is what the CPU inference would offer.
Here are all the tests output:
1. Qwen3 32B Q4_K_M
Using Qwen Qwen3-32B-Q4_K_M.gguf file.
/root/llama.cpp/build/bin/llama-bench --numa distribute -t 64 -p 0 -n 128,256,512,1024,2048 -m /root/models/tests/Qwen3-32B-Q4_K_M.gguf | model | size | params | backend | threads | test | t/s | | ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: | | qwen3 32B Q4_K - Medium | 18.40 GiB | 32.76 B | BLAS,RPC | 64 | tg128 | 13.57 ± 0.00 | | qwen3 32B Q4_K - Medium | 18.40 GiB | 32.76 B | BLAS,RPC | 64 | tg256 | 13.53 ± 0.01 | | qwen3 32B Q4_K - Medium | 18.40 GiB | 32.76 B | BLAS,RPC | 64 | tg512 | 13.41 ± 0.00 | | qwen3 32B Q4_K - Medium | 18.40 GiB | 32.76 B | BLAS,RPC | 64 | tg1024 | 13.16 ± 0.00 | | qwen3 32B Q4_K - Medium | 18.40 GiB | 32.76 B | BLAS,RPC | 64 | tg2048 | 12.47 ± 0.01 | build: 66625a59 (6040)
2. Qwen3 32B Q5_K_M
Using Qwen Qwen3-32B-Q5_K_M.gguf file.
/root/llama.cpp/build/bin/llama-bench --numa distribute -t 64 -p 0 -n 128,256,512,1024,2048 -m /root/models/tests/Qwen3-32B-Q5_K_M.gguf | model | size | params | backend | threads | test | t/s | | ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: | | qwen3 32B Q5_K - Medium | 21.61 GiB | 32.76 B | BLAS,RPC | 64 | tg128 | 12.00 ± 0.00 | | qwen3 32B Q5_K - Medium | 21.61 GiB | 32.76 B | BLAS,RPC | 64 | tg256 | 12.00 ± 0.00 | | qwen3 32B Q5_K - Medium | 21.61 GiB | 32.76 B | BLAS,RPC | 64 | tg512 | 11.89 ± 0.00 | | qwen3 32B Q5_K - Medium | 21.61 GiB | 32.76 B | BLAS,RPC | 64 | tg1024 | 11.68 ± 0.00 | | qwen3 32B Q5_K - Medium | 21.61 GiB | 32.76 B | BLAS,RPC | 64 | tg2048 | 11.11 ± 0.02 | build: 66625a59 (6040)
3. Qwen3 32B Q6_K
Using Qwen Qwen3-32B-Q6_K.gguf file.
/root/llama.cpp/build/bin/llama-bench --numa distribute -t 64 -p 0 -n 128,256,512,1024,2048 -m /root/models/tests/Qwen3-32B-Q6_K.gguf | model | size | params | backend | threads | test | t/s | | ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: | | qwen3 32B Q6_K | 25.03 GiB | 32.76 B | BLAS,RPC | 64 | tg128 | 10.76 ± 0.00 | | qwen3 32B Q6_K | 25.03 GiB | 32.76 B | BLAS,RPC | 64 | tg256 | 10.77 ± 0.00 | | qwen3 32B Q6_K | 25.03 GiB | 32.76 B | BLAS,RPC | 64 | tg512 | 10.68 ± 0.00 | | qwen3 32B Q6_K | 25.03 GiB | 32.76 B | BLAS,RPC | 64 | tg1024 | 10.52 ± 0.00 | | qwen3 32B Q6_K | 25.03 GiB | 32.76 B | BLAS,RPC | 64 | tg2048 | 10.04 ± 0.01 | build: 66625a59 (6040)
4. Qwen3 32B Q8_0
Using Qwen Qwen3-32B-Q8_0.gguf file.
/root/llama.cpp/build/bin/llama-bench --numa distribute -t 64 -p 0 -n 128,256,512,1024,2048 -m /root/models/tests/Qwen3-32B-Q8_0.gguf | model | size | params | backend | threads | test | t/s | | ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: | | qwen3 32B Q8_0 | 32.42 GiB | 32.76 B | BLAS,RPC | 64 | tg128 | 8.80 ± 0.00 | | qwen3 32B Q8_0 | 32.42 GiB | 32.76 B | BLAS,RPC | 64 | tg256 | 8.83 ± 0.00 | | qwen3 32B Q8_0 | 32.42 GiB | 32.76 B | BLAS,RPC | 64 | tg512 | 8.77 ± 0.00 | | qwen3 32B Q8_0 | 32.42 GiB | 32.76 B | BLAS,RPC | 64 | tg1024 | 8.67 ± 0.00 | | qwen3 32B Q8_0 | 32.42 GiB | 32.76 B | BLAS,RPC | 64 | tg2048 | 8.38 ± 0.01 | build: 66625a59 (6040)
Before all tests the cleaning cache commands were executed:
echo 0 > /proc/sys/kernel/numa_balancing echo 3 > /proc/sys/vm/drop_caches