Speed test matrix
Qwen3-0.6B on RTX 3090 24GB
vllm · 0.26.0 · BF164 contexts (0–32K) × 5 concurrency levels (1–16) · prefill at 4K, 16K, 32K
Hardware
- Type
- Discrete GPU
- GPU
- RTX 3090
- VRAM
- 24 GB
- CPU
- AMD Ryzen 9 9950X 16-Core Processor
- OS
- linux
Engine
- Engine
- vllm
- Version
- 0.26.0
- Quantization
- BF16
Run
- Served model
- Qwen/Qwen3-0.6B
- Revision
- main
- Benchmark tool
- llm-inference-bench v0.7.7
- Decode test
- Fixed duration
- Prefill mode
- Measured during decode warm-up
- Time per measurement
- 30 s
- Max output
- 8,192 tokens
Launch command
VLLM_USE_FLASHINFER_SAMPLER=0 vllm serve /home/lotto/models/Qwen3-0.6B --served-model-name Qwen/Qwen3-0.6B --max-model-len 40960 --gpu-memory-utilization 0.9
Prefill
prompt tok/s by context length4K context30.8K135 ms TTFT
16K context19.2K845 ms TTFT
32K context11.9K2,703 ms TTFT
Sustained decode
total tok/s · rows are context length, columns are simultaneous requests| Context | 1 | 2 | 4 | 8 | 16 |
|---|---|---|---|---|---|
| 0 | 290.59 ms TTFT | 422.714 ms TTFT | 702.613 ms TTFT | Looped | 1,635.728 ms TTFT |
| 4K | 236.818 ms TTFT | 358.624 ms TTFT | Looped | 803.651 ms TTFT | Over capacity |
| 16K | 170.546 ms TTFT | 228.265 ms TTFT | 316.488 ms TTFT | Over capacity | Over capacity |
| 32K | 121.983 ms TTFT | 150.1112 ms TTFT | 193.6156 ms TTFT | 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 | 4K | 1 | Valid | 30.8K | — | 135 ms | — | — | 1 | — | — |
| Prefill | 16K | 1 | Valid | 19.2K | — | 845 ms | — | — | 1 | — | — |
| Prefill | 32K | 1 | Valid | 11.9K | — | 2,703 ms | — | — | 1 | — | — |
| Sustained decode | 0 | 1 | Valid | 290.5 | 316.1 | 9 ms | 3.2 ms | 1.0 | 1 | 8,714 | 30.0 s |
| Sustained decode | 0 | 2 | Valid | 422.7 | 223.8 | 14 ms | 4.5 ms | 2.0 | 0 | 12,680 | 30.0 s |
| Sustained decode | 0 | 4 | Valid | 702.6 | 189.6 | 13 ms | 5.3 ms | 4.0 | 0 | 21,076 | 30.0 s |
| Sustained decode | 0 | 8 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 0 | 16 | Valid | 1,635.7 | 118.5 | 28 ms | 8.4 ms | 16.0 | 0 | 49,056 | 30.0 s |
| Sustained decode | 4K | 1 | Valid | 236.8 | 241.1 | 18 ms | 4.1 ms | 1.0 | 0 | 7,103 | 30.0 s |
| Sustained decode | 4K | 2 | Valid | 358.6 | 184.2 | 24 ms | 5.4 ms | 2.0 | 0 | 10,758 | 30.0 s |
| Sustained decode | 4K | 4 | Looped | — | — | — | — | — | 0 | 0 | — |
| Sustained decode | 4K | 8 | Valid | 803.6 | 103.8 | 51 ms | 9.6 ms | 8.0 | 0 | 24,104 | 30.0 s |
| Sustained decode | 4K | 16 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 16K | 1 | Valid | 170.5 | 171.7 | 46 ms | 5.8 ms | 1.0 | 0 | 5,113 | 30.0 s |
| Sustained decode | 16K | 2 | Valid | 228.2 | 115.1 | 65 ms | 8.7 ms | 2.0 | 0 | 6,844 | 30.0 s |
| Sustained decode | 16K | 4 | Valid | 316.4 | 79.8 | 88 ms | 12.5 ms | 4.0 | 0 | 9,492 | 30.0 s |
| Sustained decode | 16K | 8 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 16K | 16 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 32K | 1 | Valid | 121.9 | 122.3 | 83 ms | 8.2 ms | 1.0 | 0 | 3,658 | 30.0 s |
| Sustained decode | 32K | 2 | Valid | 150.1 | 75.2 | 112 ms | 13.3 ms | 2.0 | 0 | 4,502 | 30.0 s |
| Sustained decode | 32K | 4 | Valid | 193.6 | 48.5 | 156 ms | 20.6 ms | 4.0 | 0 | 5,808 | 30.0 s |
| Sustained decode | 32K | 8 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 32K | 16 | 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.