Speed test matrix
Qwen3.5-0.8B on RTX PRO 6000 Blackwell Server Edition 96GB
vllm · 0.30.1rc1.dev452+g73a583112 · BF166 contexts (8K–251K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K, 251K, 256K
Hardware
- Type
- Discrete GPU
- GPU
- RTX PRO 6000 Blackwell Server Edition
- VRAM
- 96 GB
- OS
- linux
Engine
- Engine
- vllm
- Version
- 0.30.1rc1.dev452+g73a583112
- Quantization
- BF16
- Backend
- cuda
- Repository
- github.com/vllm-project/vllm
- Build
- cuda-13.2
- Commit
73a5831127a9
Run
- Served model
- qwen
- Revision
- main
- Benchmark tool
- llm-inference-bench v0.7.6
- Decode test
- Fixed duration
- Prefill mode
- Measured during decode warm-up
- Time per measurement
- 30 s
- Max output
- 4,096 tokens
Launch command
env MAX_JOBS=2 FLASHINFER_NVCC_THREADS=1 TORCHINDUCTOR_COMPILE_THREADS=1 "$HOME/env_rtxpro6k-qwen-vllm-73a5831127a9/bin/vllm" serve Qwen/Qwen3.5-0.8B --served-model-name qwen --trust-remote-code --tensor-parallel-size 1 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder --max-model-len 262144 --gpu-memory-utilization 0.83 --host 0.0.0.0 --port 8000 --api-key <redacted> --default-chat-template-kwargs '{"enable_thinking":false}' --attention-backend FLASHINFERPrefill
prompt tok/s by context length8K context107.8K76 ms TTFT
16K context115.6K140 ms TTFT
32K context106.4K304 ms TTFT
64K context85.6K753 ms TTFT
128K context59K2,180 ms TTFT
251K context36.3K6,941 ms TTFT
256K context35.5K7,239 ms TTFT
Sustained decode
total tok/s · rows are context length, columns are simultaneous requests| Context | 1 | 2 | 4 | 8 | 16 | 32 | 64 | 128 |
|---|---|---|---|---|---|---|---|---|
| 8K | 507.832 ms TTFT | 943.650 ms TTFT | 1,819.772 ms TTFT | Looped | 5,407.5137 ms TTFT | 7,824.9249 ms TTFT | 10.7K474 ms TTFT | Looped |
| 16K | 490.646 ms TTFT | 915.865 ms TTFT | 1,761.9104 ms TTFT | 3,103.5147 ms TTFT | Looped | 7,000.5395 ms TTFT | Looped | Looped |
| 32K | 459.568 ms TTFT | 873.194 ms TTFT | 1,648.5134 ms TTFT | 2,880.4242 ms TTFT | Looped | Looped | 7,306.51,465 ms TTFT | 8,365.62,608 ms TTFT |
| 64K | 407.2121 ms TTFT | 775.7186 ms TTFT | 1,475.1220 ms TTFT | 2,520.8297 ms TTFT | Looped | Looped | Looped | Over capacity |
| 128K | 330.0238 ms TTFT | 635.9304 ms TTFT | 1,218.9342 ms TTFT | 1,960.4403 ms TTFT | 2,673.5459 ms TTFT | 3,065.93,170 ms TTFT | Over capacity | Over capacity |
| 251K | 242.1538 ms TTFT | 469.9624 ms TTFT | Looped | 1,388.1632 ms TTFT | Looped | Over capacity | Over capacity | Over capacity |
Looped: the model kept repeating the same text, so the benchmark discarded this measurement. 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 | 107.8K | — | 76 ms | — | — | 1 | — | — |
| Prefill | 16K | 1 | Valid | 115.6K | — | 140 ms | — | — | 1 | — | — |
| Prefill | 32K | 1 | Valid | 106.4K | — | 304 ms | — | — | 1 | — | — |
| Prefill | 64K | 1 | Valid | 85.6K | — | 753 ms | — | — | 1 | — | — |
| Prefill | 128K | 1 | Valid | 59K | — | 2,180 ms | — | — | 1 | — | — |
| Prefill | 251K | 1 | Valid | 36.3K | — | 6,941 ms | — | — | 1 | — | — |
| Prefill | 256K | 1 | Valid | 35.5K | — | 7,239 ms | — | — | 1 | — | — |
| Sustained decode | 8K | 1 | Valid | 507.8 | 510.1 | 32 ms | 2.0 ms | 1.0 | 4 | 15,235 | 30.0 s |
| Sustained decode | 8K | 2 | Valid | 943.6 | 477.0 | 50 ms | 2.1 ms | 1.9 | 6 | 28,308 | 30.0 s |
| Sustained decode | 8K | 4 | Valid | 1,819.7 | 459.7 | 72 ms | 2.2 ms | 4.0 | 12 | 55,096 | 30.3 s |
| Sustained decode | 8K | 8 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 8K | 16 | Valid | 5,407.5 | 344.9 | 137 ms | 2.9 ms | 16.0 | 32 | 162,224 | 30.0 s |
| Sustained decode | 8K | 32 | Valid | 7,824.9 | 252.6 | 249 ms | 4.0 ms | 31.1 | 64 | 234,746 | 30.0 s |
| Sustained decode | 8K | 64 | Valid | 10.7K | 177.3 | 474 ms | 5.6 ms | 62.8 | 64 | 322,371 | 30.0 s |
| Sustained decode | 8K | 128 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 16K | 1 | Valid | 490.6 | 493.9 | 46 ms | 2.0 ms | 1.0 | 4 | 14,717 | 30.0 s |
| Sustained decode | 16K | 2 | Valid | 915.8 | 461.5 | 65 ms | 2.2 ms | 2.0 | 6 | 27,474 | 30.0 s |
| Sustained decode | 16K | 4 | Valid | 1,761.9 | 446.5 | 104 ms | 2.2 ms | 4.0 | 12 | 52,858 | 30.0 s |
| Sustained decode | 16K | 8 | Valid | 3,103.5 | 395.0 | 147 ms | 2.5 ms | 8.0 | 24 | 93,106 | 30.0 s |
| Sustained decode | 16K | 16 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 16K | 32 | Valid | 7,000.5 | 225.1 | 395 ms | 4.4 ms | 32.0 | 32 | 210,006 | 30.0 s |
| Sustained decode | 16K | 64 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 16K | 128 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 32K | 1 | Valid | 459.5 | 462.7 | 68 ms | 2.2 ms | 1.0 | 3 | 13,784 | 30.0 s |
| Sustained decode | 32K | 2 | Valid | 873.1 | 441.0 | 94 ms | 2.3 ms | 2.0 | 6 | 26,194 | 30.0 s |
| Sustained decode | 32K | 4 | Valid | 1,648.5 | 421.4 | 134 ms | 2.4 ms | 3.9 | 12 | 49,456 | 30.0 s |
| Sustained decode | 32K | 8 | Valid | 2,880.4 | 371.0 | 242 ms | 2.7 ms | 7.8 | 24 | 86,409 | 30.0 s |
| Sustained decode | 32K | 16 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 32K | 32 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 32K | 64 | Valid | 7,306.5 | 119.2 | 1,465 ms | 8.4 ms | 62.4 | 47 | 219,138 | 30.0 s |
| Sustained decode | 32K | 128 | Valid | 8,365.6 | 66.7 | 2,608 ms | 15.0 ms | 128.0 | 0 | 250,962 | 30.0 s |
| Sustained decode | 64K | 1 | Valid | 407.2 | 412.1 | 121 ms | 2.4 ms | 1.0 | 3 | 12,214 | 30.0 s |
| Sustained decode | 64K | 2 | Valid | 775.7 | 395.0 | 186 ms | 2.5 ms | 1.9 | 6 | 23,271 | 30.0 s |
| Sustained decode | 64K | 4 | Valid | 1,475.1 | 377.6 | 220 ms | 2.6 ms | 4.0 | 12 | 44,251 | 30.0 s |
| Sustained decode | 64K | 8 | Valid | 2,520.8 | 325.9 | 297 ms | 3.1 ms | 7.8 | 16 | 75,626 | 30.0 s |
| Sustained decode | 64K | 16 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 64K | 32 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 64K | 64 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 64K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 128K | 1 | Valid | 330.0 | 335.5 | 238 ms | 3.0 ms | 1.0 | 2 | 9,898 | 30.0 s |
| Sustained decode | 128K | 2 | Valid | 635.9 | 325.0 | 304 ms | 3.1 ms | 2.0 | 4 | 19,076 | 30.0 s |
| Sustained decode | 128K | 4 | Valid | 1,218.9 | 311.6 | 342 ms | 3.2 ms | 3.9 | 8 | 36,569 | 30.0 s |
| Sustained decode | 128K | 8 | Valid | 1,960.4 | 251.1 | 403 ms | 4.0 ms | 7.8 | 16 | 58,803 | 30.0 s |
| Sustained decode | 128K | 16 | Valid | 2,673.5 | 169.7 | 459 ms | 5.9 ms | 15.8 | 16 | 80,206 | 30.0 s |
| Sustained decode | 128K | 32 | Valid | 3,065.9 | 96.6 | 3,170 ms | 10.3 ms | 32.0 | 0 | 91,996 | 30.0 s |
| Sustained decode | 128K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 128K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 251K | 1 | Valid | 242.1 | 247.0 | 538 ms | 4.0 ms | 1.0 | 1 | 7,262 | 30.0 s |
| Sustained decode | 251K | 2 | Valid | 469.9 | 239.7 | 624 ms | 4.2 ms | 2.0 | 2 | 14,097 | 30.0 s |
| Sustained decode | 251K | 4 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 251K | 8 | Valid | 1,388.1 | 177.0 | 632 ms | 5.7 ms | 7.9 | 8 | 41,723 | 30.1 s |
| Sustained decode | 251K | 16 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 251K | 32 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 251K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 251K | 128 | Over capacity | — | — | — | — | — | — | — | — |
Looped: the model kept repeating the same text, so the benchmark discarded this measurement. Over capacity: not run because this many requests at this context would exceed the server's KV cache.