Speed test matrix
gemma-4-12B-it on Intel Arc Pro B65 32GB ×4
vllm · 0.32.0 · BF164 contexts (0–64K) × 5 concurrency levels (1–16) · prefill at 8K, 32K, 64K
Hardware
- Type
- Discrete GPU
- GPU
- Intel Arc Pro B65
- GPU count
- 4
- VRAM
- 128 GB (4 × 32 GB)
- CPU
- AMD EPYC 7343 16-Core Processor
- RAM
- 499 GB
- OS
- Ubuntu 26.04.1 LTS
Engine
- Engine
- vllm
- Version
- 0.32.0
- Quantization
- BF16
- Build
- vllm/vllm-openai-xpu:v0.32.0@sha256:20509fabe50dda62c09c0786d1a6effa299d3dbdddcc49331b93e65144652a8c
Run
- Served model
- g12
- Revision
- 707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7
- Benchmark tool
- llm-inference-bench v0.8.0
- Decode test
- Fixed duration
- Prefill mode
- Measured during decode warm-up
- Time per measurement
- 30 s
- Max output
- 256 tokens
Launch command
docker run --rm -d --name lsv-g12-u-tp4-ctx131k --net=host --shm-size=32g --device /dev/dri:/dev/dri -v /dev/dri:/dev/dri -v /models/weights:/llm/models:ro -v /models/bench/lmx-ctx-20261019/session-a/g12-u-tp4-ctx131k/session:/bench -e ZE_AFFINITY_MASK=0\,1\,2\,3 -e VLLM_XPU_ENABLE_XPU_GRAPH=1 -e DISABLE_ESIMD_GDN_OUTPROJ=0 -e DISABLE_ESIMD_INT4=0 -e DISABLE_ESIMD_INT4_GEMM=0 -e VLLM_OFFLOAD_WEIGHTS_BEFORE_QUANT=1 -e VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 -e VLLM_WORKER_MULTIPROC_METHOD=spawn -e CCL_ZE_IPC_EXCHANGE=sockets --entrypoint vllm vllm/vllm-openai-xpu:v0.32.0@sha256:20509fabe50dda62c09c0786d1a6effa299d3dbdddcc49331b93e65144652a8c serve --model /llm/models/gemma-4-12B-it --host 0.0.0.0 --served-model-name g12 --tensor-parallel-size 4 --pipeline-parallel-size 1 --data-parallel-size 1 --gpu-memory-utilization 0.92 --max-model-len 131072 --max-num-seqs 16 --max-num-batched-tokens 8192 --block-size 64 --kv-cache-dtype auto --mamba-ssm-cache-dtype auto --dtype bfloat16 --enable-prefix-caching --trust-remote-code --limit-mm-per-prompt \{\"image\":0\,\"video\":0\} --cpu-offload-gb 0 --port 8000Prefill
prompt tok/s by context length8K context1,4075,825 ms TTFT
32K context426.076,002 ms TTFT
64K context213.0302,443 ms TTFT
Sustained decode
total tok/s · rows are context length, columns are simultaneous requests| Context | 1 | 2 | 4 | 8 | 16 |
|---|---|---|---|---|---|
| 0 | 50.731 ms TTFT | 99.063 ms TTFT | 191.168 ms TTFT | 360.184 ms TTFT | 648.0135 ms TTFT |
| 8K | 32.4163 ms TTFT | 61.5229 ms TTFT | 117.1286 ms TTFT | 207.3311 ms TTFT | 311.8411 ms TTFT |
| 32K | 20.9Requests queued483 ms TTFT | 36.5660 ms TTFT | 65.8978 ms TTFT | 117.11,234 ms TTFT | 153.61,724 ms TTFT |
| 64K | 16.8Requests queued931 ms TTFT | 27.21,227 ms TTFT | 47.11,627 ms TTFT | 75.81,971 ms TTFT | Over capacity |
Requests queued: the server could not run all requests at once, so some waited in a queue. Over capacity: not run because this many requests at this context would exceed the server's KV cache.
TTFT is measured after the benchmark warms the server's prompt cache, so it mostly excludes prefill.
All measurements
Latencies are medians. Total tok/s is output across all simultaneous requests.| Phase | Context | Requests at once | Status | Total tok/s | Per request tok/s | TTFT | Time between tokens | Average active requests | Completed requests | Output tokens | Measured for |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Prefill | 8K | 1 | Valid | 1,407.0 | — | 5,825 ms | — | — | 1 | — | — |
| Prefill | 32K | 1 | Valid | 426.0 | — | 76,002 ms | — | — | 1 | — | — |
| Prefill | 64K | 1 | Valid | 213.0 | — | 302,443 ms | — | — | 1 | — | — |
| Sustained decode | 0 | 1 | Valid | 50.7 | 50.8 | 31 ms | 19.7 ms | 1.0 | 7 | 1,519 | 30.0 s |
| Sustained decode | 0 | 2 | Valid | 99.0 | 49.9 | 63 ms | 20.1 ms | 2.0 | 12 | 2,970 | 30.0 s |
| Sustained decode | 0 | 4 | Valid | 191.1 | 48.1 | 68 ms | 20.8 ms | 4.0 | 24 | 5,732 | 30.0 s |
| Sustained decode | 0 | 8 | Valid | 360.1 | 45.7 | 84 ms | 21.9 ms | 8.0 | 48 | 10,796 | 30.0 s |
| Sustained decode | 0 | 16 | Valid | 648.0 | 41.2 | 135 ms | 24.3 ms | 16.0 | 80 | 19,433 | 30.0 s |
| Sustained decode | 8K | 1 | Valid | 32.4 | 33.0 | 163 ms | 30.3 ms | 1.0 | 4 | 971 | 30.0 s |
| Sustained decode | 8K | 2 | Valid | 61.5 | 31.6 | 229 ms | 31.7 ms | 2.0 | 8 | 1,844 | 30.0 s |
| Sustained decode | 8K | 4 | Valid | 117.1 | 30.0 | 286 ms | 33.3 ms | 3.9 | 12 | 3,512 | 30.0 s |
| Sustained decode | 8K | 8 | Valid | 207.3 | 26.7 | 311 ms | 37.5 ms | 7.9 | 24 | 6,212 | 30.0 s |
| Sustained decode | 8K | 16 | Valid | 311.8 | 19.9 | 411 ms | 50.2 ms | 15.9 | 32 | 9,349 | 30.0 s |
| Sustained decode | 32K | 1 | Requests queued | 20.9 | 21.6 | 483 ms | 46.4 ms | 0.9 | 2 | 628 | 30.0 s |
| Sustained decode | 32K | 2 | Valid | 36.5 | 19.0 | 660 ms | 52.6 ms | 2.0 | 4 | 1,094 | 30.0 s |
| Sustained decode | 32K | 4 | Valid | 65.8 | 17.5 | 978 ms | 57.2 ms | 3.9 | 8 | 1,972 | 30.0 s |
| Sustained decode | 32K | 8 | Valid | 117.1 | 15.0 | 1,234 ms | 66.6 ms | 7.8 | 8 | 3,512 | 30.0 s |
| Sustained decode | 32K | 16 | Valid | 153.6 | 10.1 | 1,724 ms | 99.5 ms | 15.6 | 16 | 4,605 | 30.0 s |
| Sustained decode | 64K | 1 | Requests queued | 16.8 | 17.5 | 931 ms | 57.2 ms | 0.9 | 1 | 503 | 30.0 s |
| Sustained decode | 64K | 2 | Valid | 27.2 | 14.1 | 1,227 ms | 71.1 ms | 1.9 | 2 | 816 | 30.0 s |
| Sustained decode | 64K | 4 | Valid | 47.1 | 12.5 | 1,627 ms | 80.1 ms | 3.9 | 4 | 1,412 | 30.0 s |
| Sustained decode | 64K | 8 | Valid | 75.8 | 10.0 | 1,971 ms | 99.7 ms | 7.7 | 8 | 2,270 | 30.0 s |
| Sustained decode | 64K | 16 | Over capacity | — | — | — | — | — | — | — | — |
Requests queued: the server could not run all requests at once, so some waited in a queue. Over capacity: not run because this many requests at this context would exceed the server's KV cache.