This article shows the CPU-only inference with a modern server processor – AMD Epyc 9554. For the LLM model the Google’s Gemma 3 27B Instruct with different quantization are used to show the difference in token generation per second and memory consumption. Gemma 3 27B Instruct is a pretty solid LLM model, offered for free by the Google giant, and in many cases it can be considered to offload some LLM work locally for free. 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 | Gemma 3 27B Instruct | 27B | Q4_K_M | 15.40 GiB | 0 | 14.272 |
| 2 | Gemma 3 27B Instruct | 27B | Q5_0 | 17.94 GiB | 10.66 | 12.75 |
| 3 | Gemma 3 27B Instruct | 27B | Q6_K | 20.64 GiB | 9.6 | 11.526 |
| 4 | Gemma 3 27B Instruct | 27B | Q8_0 | 26.73 GiB | 16.64 | 9.608 |
| 6 | Gemma 3 27B Instruct | 27B | F16 | 50.31 GiB | 41.84 | 5.588 |
The difference between the Q4 and F16 in the tokens per second is 60.84% speed degradation, but the problem is the around 5 tokens per second is on the brick of not usable being too slow. Around 14 tokens per second is good and usable for daily use for a single user.
Here are all the tests output:
1. Gemma 3 27b instruct Q4_K_M
Using unsloth gemma-3-27b-it-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/gemma-3-27b-it-Q4_K_M.gguf | model | size | params | backend | threads | test | t/s | | ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: | | gemma3 27B Q4_K - Medium | 15.40 GiB | 27.01 B | BLAS,RPC | 64 | tg128 | 14.62 ± 0.01 | | gemma3 27B Q4_K - Medium | 15.40 GiB | 27.01 B | BLAS,RPC | 64 | tg256 | 14.54 ± 0.01 | | gemma3 27B Q4_K - Medium | 15.40 GiB | 27.01 B | BLAS,RPC | 64 | tg512 | 14.42 ± 0.01 | | gemma3 27B Q4_K - Medium | 15.40 GiB | 27.01 B | BLAS,RPC | 64 | tg1024 | 14.19 ± 0.00 | | gemma3 27B Q4_K - Medium | 15.40 GiB | 27.01 B | BLAS,RPC | 64 | tg2048 | 13.59 ± 0.01 | build: 66625a59 (6040)
2. Gemma 3 27b instruct Q5_K_M
Using unsloth gemma-3-27b-it-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/gemma-3-27b-it-Q5_K_M.gguf | model | size | params | backend | threads | test | t/s | | ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: | | gemma3 27B Q5_K - Medium | 17.94 GiB | 27.01 B | BLAS,RPC | 64 | tg128 | 13.07 ± 0.00 | | gemma3 27B Q5_K - Medium | 17.94 GiB | 27.01 B | BLAS,RPC | 64 | tg256 | 12.88 ± 0.00 | | gemma3 27B Q5_K - Medium | 17.94 GiB | 27.01 B | BLAS,RPC | 64 | tg512 | 12.96 ± 0.00 | | gemma3 27B Q5_K - Medium | 17.94 GiB | 27.01 B | BLAS,RPC | 64 | tg1024 | 12.77 ± 0.00 | | gemma3 27B Q5_K - Medium | 17.94 GiB | 27.01 B | BLAS,RPC | 64 | tg2048 | 12.07 ± 0.00 | build: 66625a59 (6040)
3. Gemma 3 27b instruct Q6_K
Using unsloth gemma-3-27b-it-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/gemma-3-27b-it-Q6_K.gguf | model | size | params | backend | threads | test | t/s | | ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: | | gemma3 27B Q6_K | 20.64 GiB | 27.01 B | BLAS,RPC | 64 | tg128 | 11.76 ± 0.00 | | gemma3 27B Q6_K | 20.64 GiB | 27.01 B | BLAS,RPC | 64 | tg256 | 11.62 ± 0.00 | | gemma3 27B Q6_K | 20.64 GiB | 27.01 B | BLAS,RPC | 64 | tg512 | 11.66 ± 0.04 | | gemma3 27B Q6_K | 20.64 GiB | 27.01 B | BLAS,RPC | 64 | tg1024 | 11.56 ± 0.01 | | gemma3 27B Q6_K | 20.64 GiB | 27.01 B | BLAS,RPC | 64 | tg2048 | 11.03 ± 0.02 | build: 66625a59 (6040)
4. Gemma 3 27b instruct Q8_0
Using unsloth gemma-3-27b-it-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/gemma-3-27b-it-Q8_0.gguf | model | size | params | backend | threads | test | t/s | | ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: | | gemma3 27B Q8_0 | 26.73 GiB | 27.01 B | BLAS,RPC | 64 | tg128 | 9.73 ± 0.01 | | gemma3 27B Q8_0 | 26.73 GiB | 27.01 B | BLAS,RPC | 64 | tg256 | 9.74 ± 0.00 | | gemma3 27B Q8_0 | 26.73 GiB | 27.01 B | BLAS,RPC | 64 | tg512 | 9.70 ± 0.00 | | gemma3 27B Q8_0 | 26.73 GiB | 27.01 B | BLAS,RPC | 64 | tg1024 | 9.55 ± 0.00 | | gemma3 27B Q8_0 | 26.73 GiB | 27.01 B | BLAS,RPC | 64 | tg2048 | 9.32 ± 0.01 | build: 66625a59 (6040)
5. Gemma 3 27b instruct F16
Using unsloth from gemma-3-27b-it-BF16-00001-of-00002.gguf and gemma-3-27b-it-BF16-00001-of-00002.gguf files.
/root/llama.cpp/build/bin/llama-bench --numa distribute -t 64 -p 0 -n 128,256,512,1024,2048 -m /root/models/tests/gemma-3-27b-it-BF16-00001-of-00002.gguf | model | size | params | backend | threads | test | t/s | | ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: | | gemma3 27B BF16 | 50.31 GiB | 27.01 B | BLAS,RPC | 64 | tg128 | 5.65 ± 0.00 | | gemma3 27B BF16 | 50.31 GiB | 27.01 B | BLAS,RPC | 64 | tg256 | 5.64 ± 0.00 | | gemma3 27B BF16 | 50.31 GiB | 27.01 B | BLAS,RPC | 64 | tg512 | 5.60 ± 0.00 | | gemma3 27B BF16 | 50.31 GiB | 27.01 B | BLAS,RPC | 64 | tg1024 | 5.57 ± 0.00 | | gemma3 27B BF16 | 50.31 GiB | 27.01 B | BLAS,RPC | 64 | tg2048 | 5.48 ± 0.00 | 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