Speed test matrix
NVIDIA-Nemotron-Nano-9B-v2-quantized.w4a16 on RTX 3090 24GB
vllm · 0.32.0 · W4A165 contexts (0–120K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 120K
Hardware
- Type
- Discrete GPU
- GPU
- RTX 3090
- VRAM
- 24 GB
- CPU
- Intel Core i7-9700K (8C)
- RAM
- 32 GB
- OS
- Ubuntu 24.04.5
Engine
- Engine
- vllm
- Version
- 0.32.0
- Quantization
- W4A16
- Backend
- cuda
- Repository
- github.com/vllm-project/vllm
- Build
- vllm/vllm-openai:v0.32.0-x86_64
Run
- Served model
- bench
- Revision
- main
- 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
- 8,192 tokens
Launch command
# Untuned out-of-the-box baseline. Spec decode: none. KV cache: int4_per_token_head # 1x RTX 3090 24GB (Zagitta/gpu-unlock), 420 W limit, PCIe Gen3 x16 | Intel Core i7-9700K 8C | 32 GB DDR4 | Ubuntu 24.04.5, NVIDIA driver 580.178.04 vllm serve RedHatAI/NVIDIA-Nemotron-Nano-9B-v2-quantized.w4a16 --max-model-len 131072 --gpu-memory-utilization 0.94 --max-num-seqs 128 --enable-prefix-caching --trust-remote-code --kv-cache-dtype int4_per_token_head --attention-backend TRITON_ATTN
Prefill
prompt tok/s by context length8K context5,0971,645 ms TTFT
16K context4,4823,706 ms TTFT
32K context3,7678,779 ms TTFT
64K context2,77323,794 ms TTFT
120K context1,90164,988 ms TTFT
Sustained decode
total tok/s · rows are context length, columns are simultaneous requests| Context | 1 | 2 | 4 | 8 | 16 | 32 | 64 | 128 |
|---|---|---|---|---|---|---|---|---|
| 0 | 128.936 ms TTFT | 243.345 ms TTFT | 441.2311 ms TTFT | 738.0153 ms TTFT | 1,141.2294 ms TTFT | 1,549.7446 ms TTFT | 1,768.0588 ms TTFT | 1,648.6Requests queued1,102 ms TTFT |
| 16K | 119.5532 ms TTFT | 214.3796 ms TTFT | 368.91,324 ms TTFT | 583.62,346 ms TTFT | 822.44,416 ms TTFT | 1,026.78,617 ms TTFT | 1,118.317,290 ms TTFT | Over capacity |
| 32K | 111.21,345 ms TTFT | 191.11,995 ms TTFT | 317.33,287 ms TTFT | 479.95,870 ms TTFT | 641.411,077 ms TTFT | 768.021,563 ms TTFT | Over capacity | Over capacity |
| 64K | 98.81,420 ms TTFT | 158.32,047 ms TTFT | 248.33,317 ms TTFT | 353.65,885 ms TTFT | 446.811,004 ms TTFT | Over capacity | Over capacity | Over capacity |
| 120K | 82.41,292 ms TTFT | 121.11,819 ms TTFT | 180.82,888 ms TTFT | 242.45,050 ms TTFT | Over capacity | Over capacity | Over capacity | 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 | 5,097.0 | — | 1,645 ms | — | — | 1 | — | — |
| Prefill | 16K | 1 | Valid | 4,482.0 | — | 3,706 ms | — | — | 1 | — | — |
| Prefill | 32K | 1 | Valid | 3,767.0 | — | 8,779 ms | — | — | 1 | — | — |
| Prefill | 64K | 1 | Valid | 2,773.0 | — | 23,794 ms | — | — | 1 | — | — |
| Prefill | 120K | 1 | Valid | 1,901.0 | — | 64,988 ms | — | — | 1 | — | — |
| Sustained decode | 0 | 1 | Valid | 128.9 | 128.9 | 36 ms | 7.8 ms | 1.0 | 0 | 3,866 | 30.0 s |
| Sustained decode | 0 | 2 | Valid | 243.3 | 121.8 | 45 ms | 8.2 ms | 2.0 | 0 | 7,298 | 30.0 s |
| Sustained decode | 0 | 4 | Valid | 441.2 | 110.7 | 311 ms | 9.0 ms | 4.0 | 0 | 13,236 | 30.0 s |
| Sustained decode | 0 | 8 | Valid | 738.0 | 92.7 | 153 ms | 10.8 ms | 8.0 | 0 | 22,136 | 30.0 s |
| Sustained decode | 0 | 16 | Valid | 1,141.2 | 71.7 | 294 ms | 14.0 ms | 16.0 | 0 | 34,224 | 30.0 s |
| Sustained decode | 0 | 32 | Valid | 1,549.7 | 48.5 | 446 ms | 20.6 ms | 32.0 | 0 | 46,496 | 30.0 s |
| Sustained decode | 0 | 64 | Valid | 1,768.0 | 27.3 | 588 ms | 36.6 ms | 64.0 | 0 | 53,051 | 30.0 s |
| Sustained decode | 0 | 128 | Requests queued | 1,648.6 | 18.6 | 1,102 ms | 53.8 ms | 90.0 | 0 | 49,410 | 30.0 s |
| Sustained decode | 16K | 1 | Valid | 119.5 | 119.6 | 532 ms | 8.4 ms | 1.0 | 0 | 3,584 | 30.0 s |
| Sustained decode | 16K | 2 | Valid | 214.3 | 106.5 | 796 ms | 9.4 ms | 2.0 | 0 | 6,426 | 30.0 s |
| Sustained decode | 16K | 4 | Valid | 368.9 | 90.5 | 1,324 ms | 11.1 ms | 4.0 | 0 | 11,064 | 30.0 s |
| Sustained decode | 16K | 8 | Valid | 583.6 | 69.6 | 2,346 ms | 14.4 ms | 8.0 | 0 | 17,504 | 30.0 s |
| Sustained decode | 16K | 16 | Valid | 822.4 | 46.5 | 4,416 ms | 21.5 ms | 16.0 | 0 | 24,672 | 30.0 s |
| Sustained decode | 16K | 32 | Valid | 1,026.7 | 26.5 | 8,617 ms | 37.8 ms | 32.0 | 0 | 30,784 | 30.0 s |
| Sustained decode | 16K | 64 | Valid | 1,118.3 | 12.6 | 17,290 ms | 79.5 ms | 64.0 | 0 | 33,536 | 30.0 s |
| Sustained decode | 16K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 32K | 1 | Valid | 111.2 | 111.4 | 1,345 ms | 9.0 ms | 1.0 | 0 | 3,337 | 30.0 s |
| Sustained decode | 32K | 2 | Valid | 191.1 | 94.0 | 1,995 ms | 10.6 ms | 2.0 | 0 | 5,732 | 30.0 s |
| Sustained decode | 32K | 4 | Valid | 317.3 | 75.2 | 3,287 ms | 13.3 ms | 4.0 | 0 | 9,520 | 30.0 s |
| Sustained decode | 32K | 8 | Valid | 479.9 | 53.1 | 5,870 ms | 18.9 ms | 8.0 | 0 | 14,392 | 30.0 s |
| Sustained decode | 32K | 16 | Valid | 641.4 | 31.5 | 11,077 ms | 31.8 ms | 16.0 | 0 | 19,232 | 30.0 s |
| Sustained decode | 32K | 32 | Valid | 768.0 | 15.5 | 21,563 ms | 64.6 ms | 32.0 | 0 | 23,040 | 30.0 s |
| Sustained decode | 32K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 32K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 64K | 1 | Valid | 98.8 | 99.0 | 1,420 ms | 10.1 ms | 1.0 | 0 | 2,963 | 30.0 s |
| Sustained decode | 64K | 2 | Valid | 158.3 | 77.8 | 2,047 ms | 12.9 ms | 2.0 | 0 | 4,746 | 30.0 s |
| Sustained decode | 64K | 4 | Valid | 248.3 | 58.9 | 3,317 ms | 17.0 ms | 4.0 | 0 | 7,448 | 30.0 s |
| Sustained decode | 64K | 8 | Valid | 353.6 | 39.2 | 5,885 ms | 25.5 ms | 8.0 | 0 | 10,608 | 30.0 s |
| Sustained decode | 64K | 16 | Valid | 446.8 | 22.1 | 11,004 ms | 45.3 ms | 16.0 | 0 | 13,392 | 30.0 s |
| Sustained decode | 64K | 32 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 64K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 64K | 128 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 120K | 1 | Valid | 82.4 | 82.5 | 1,292 ms | 12.1 ms | 1.0 | 0 | 2,471 | 30.0 s |
| Sustained decode | 120K | 2 | Valid | 121.1 | 59.7 | 1,819 ms | 16.7 ms | 2.0 | 0 | 3,630 | 30.0 s |
| Sustained decode | 120K | 4 | Valid | 180.8 | 43.3 | 2,888 ms | 23.1 ms | 4.0 | 0 | 5,420 | 30.0 s |
| Sustained decode | 120K | 8 | Valid | 242.4 | 27.5 | 5,050 ms | 36.4 ms | 8.0 | 0 | 7,272 | 30.0 s |
| Sustained decode | 120K | 16 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 120K | 32 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 120K | 64 | Over capacity | — | — | — | — | — | — | — | — |
| Sustained decode | 120K | 128 | 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.