Speed test matrix
Qwen3.5-9B on Intel Arc Pro B65 32GB ×4
vllm · 0.32.0 · BF166 contexts (0–200K) × 5 concurrency levels (1–16) · prefill at 8K, 32K, 64K, 128K
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
- q359
- Revision
- c202236235762e1c871ad0ccb60c8ee5ba337b9a
- 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-q359-u-tp4-ctx --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/q359-u-tp4-ctx/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/Qwen3.5-9B --host 0.0.0.0 --served-model-name q359 --tensor-parallel-size 4 --pipeline-parallel-size 1 --data-parallel-size 1 --gpu-memory-utilization 0.92 --max-model-len 262144 --max-num-seqs 16 --max-num-batched-tokens 8192 --block-size 64 --kv-cache-dtype auto --mamba-ssm-cache-dtype float16 --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 context10.4K790 ms TTFT
32K context9,9783,236 ms TTFT
64K context8,9077,232 ms TTFT
128K context7,30717,613 ms TTFT
Sustained decode
total tok/s · rows are context length, columns are simultaneous requests| Context | 1 | 2 | 4 | 8 | 16 |
|---|---|---|---|---|---|
| 0 | 82.841 ms TTFT | 163.158 ms TTFT | 319.368 ms TTFT | 612.888 ms TTFT | 1,132.8130 ms TTFT |
| 8K | 80.669 ms TTFT | 158.4100 ms TTFT | 301.8103 ms TTFT | 559.2106 ms TTFT | 966.5112 ms TTFT |
| 32K | 76.6153 ms TTFT | 146.3Requests queued198 ms TTFT | 270.5256 ms TTFT | 468.4Requests queued279 ms TTFT | 737.2Requests queued288 ms TTFT |
| 64K | 71.4248 ms TTFT | 135.3Requests queued301 ms TTFT | 239.8Requests queued425 ms TTFT | 398.2Requests queued612 ms TTFT | 593.7Requests queued626 ms TTFT |
| 128K | 62.9476 ms TTFT | 112.4Requests queued593 ms TTFT | 191.9Requests queued839 ms TTFT | 295.6Requests queued1,184 ms TTFT | 414.8Requests queued1,491 ms TTFT |
| 200K | 54.7788 ms TTFT | 93.8Requests queued972 ms TTFT | 153.2Requests queued1,360 ms TTFT | 227.6Requests queued1,485 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 | 10.4K | — | 790 ms | — | — | 1 | — | — |
| Prefill | 32K | 1 | Valid | 9,978.0 | — | 3,236 ms | — | — | 1 | — | — |
| Prefill | 64K | 1 | Valid | 8,907.0 | — | 7,232 ms | — | — | 1 | — | — |
| Prefill | 128K | 1 | Valid | 7,307.0 | — | 17,613 ms | — | — | 1 | — | — |
| Sustained decode | 0 | 1 | Valid | 82.8 | 83.6 | 41 ms | 12.0 ms | 1.0 | 11 | 2,482 | 30.0 s |
| Sustained decode | 0 | 2 | Valid | 163.1 | 82.8 | 58 ms | 12.1 ms | 2.0 | 22 | 4,892 | 30.0 s |
| Sustained decode | 0 | 4 | Valid | 319.3 | 81.4 | 68 ms | 12.3 ms | 4.0 | 44 | 9,576 | 30.0 s |
| Sustained decode | 0 | 8 | Valid | 612.8 | 78.5 | 88 ms | 12.7 ms | 7.8 | 80 | 18,382 | 30.0 s |
| Sustained decode | 0 | 16 | Valid | 1,132.8 | 72.8 | 130 ms | 13.7 ms | 15.7 | 144 | 33,983 | 30.0 s |
| Sustained decode | 8K | 1 | Valid | 80.6 | 82.1 | 69 ms | 12.2 ms | 1.0 | 10 | 2,420 | 30.0 s |
| Sustained decode | 8K | 2 | Valid | 158.4 | 81.1 | 100 ms | 12.3 ms | 1.9 | 20 | 4,752 | 30.0 s |
| Sustained decode | 8K | 4 | Valid | 301.8 | 77.6 | 103 ms | 12.9 ms | 4.0 | 44 | 9,052 | 30.0 s |
| Sustained decode | 8K | 8 | Valid | 559.2 | 71.6 | 106 ms | 14.0 ms | 7.9 | 72 | 16,777 | 30.0 s |
| Sustained decode | 8K | 16 | Valid | 966.5 | 62.3 | 112 ms | 16.1 ms | 15.7 | 142 | 28,962 | 30.0 s |
| Sustained decode | 32K | 1 | Valid | 76.6 | 79.6 | 153 ms | 12.6 ms | 1.0 | 9 | 2,296 | 30.0 s |
| Sustained decode | 32K | 2 | Requests queued | 146.3 | 76.7 | 198 ms | 13.0 ms | 1.8 | 18 | 4,390 | 30.0 s |
| Sustained decode | 32K | 4 | Valid | 270.5 | 72.0 | 256 ms | 13.9 ms | 3.8 | 36 | 8,102 | 29.9 s |
| Sustained decode | 32K | 8 | Requests queued | 468.4 | 62.6 | 279 ms | 16.0 ms | 7.7 | 64 | 14,048 | 30.0 s |
| Sustained decode | 32K | 16 | Requests queued | 737.2 | 49.6 | 288 ms | 20.2 ms | 15.4 | 112 | 22,112 | 30.0 s |
| Sustained decode | 64K | 1 | Valid | 71.4 | 76.8 | 248 ms | 13.0 ms | 1.0 | 9 | 2,141 | 30.0 s |
| Sustained decode | 64K | 2 | Requests queued | 135.3 | 72.4 | 301 ms | 13.8 ms | 1.9 | 18 | 4,040 | 29.8 s |
| Sustained decode | 64K | 4 | Requests queued | 239.8 | 65.5 | 425 ms | 15.3 ms | 3.6 | 32 | 7,178 | 29.9 s |
| Sustained decode | 64K | 8 | Requests queued | 398.2 | 55.9 | 612 ms | 17.9 ms | 7.2 | 48 | 11,944 | 30.0 s |
| Sustained decode | 64K | 16 | Requests queued | 593.7 | 41.9 | 626 ms | 23.9 ms | 14.9 | 80 | 17,794 | 30.0 s |
| Sustained decode | 128K | 1 | Valid | 62.9 | 71.7 | 476 ms | 13.9 ms | 1.0 | 8 | 1,886 | 30.0 s |
| Sustained decode | 128K | 2 | Requests queued | 112.4 | 65.0 | 593 ms | 15.4 ms | 1.8 | 16 | 3,370 | 30.0 s |
| Sustained decode | 128K | 4 | Requests queued | 191.9 | 55.6 | 839 ms | 18.0 ms | 3.5 | 24 | 5,756 | 30.0 s |
| Sustained decode | 128K | 8 | Requests queued | 295.6 | 44.3 | 1,184 ms | 22.6 ms | 7.0 | 40 | 8,817 | 29.8 s |
| Sustained decode | 128K | 16 | Requests queued | 414.8 | 28.9 | 1,491 ms | 34.6 ms | 14.6 | 48 | 12,443 | 30.0 s |
| Sustained decode | 200K | 1 | Valid | 54.7 | 66.6 | 788 ms | 15.0 ms | 0.8 | 7 | 1,641 | 30.0 s |
| Sustained decode | 200K | 2 | Requests queued | 93.8 | 58.1 | 972 ms | 17.2 ms | 1.7 | 12 | 2,813 | 30.0 s |
| Sustained decode | 200K | 4 | Requests queued | 153.2 | 47.6 | 1,360 ms | 21.0 ms | 3.3 | 20 | 4,594 | 30.0 s |
| Sustained decode | 200K | 8 | Requests queued | 227.6 | 34.1 | 1,485 ms | 29.3 ms | 7.0 | 30 | 6,825 | 30.0 s |
| Sustained decode | 200K | 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.