入門指南排行榜Decode calculator模型OrganizationsReportsMatrix runs硬體基準評測市集租用ProAPI文件
語言
Actual Computer — Every computer, one endpoint

Speed test matrix

gemma-4-E2B-it on H200 NVL 141GB

vllm · 0.26.1rc1.dev638+g52be12cfa · BF165 contexts (8K–123K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 123K, 128K
by @keennaylmx dev importMeasured

Hardware

Type
Discrete GPU
GPU
H200 NVL
VRAM
141 GB
OS
linux

Engine

Engine
vllm
Version
0.26.1rc1.dev638+g52be12cfa
Quantization
BF16
Backend
cuda
Build
cuda-13.2
Commit
52be12cfac0c

Run

Served model
gemma
Revision
main
Decode test
Fixed duration
Prefill mode
Measured during decode warm-up
Time per measurement
30 s
Max output
4,096 tokens

Launch command

"$HOME/env_h200-gemma-vllm/bin/vllm" serve google/gemma-4-E2B-it --served-model-name gemma --trust-remote-code --tensor-parallel-size 1 --reasoning-parser gemma4 --enable-auto-tool-choice --tool-call-parser gemma4 --max-model-len 131072 --gpu-memory-utilization 0.854015 --host 0.0.0.0 --port 8000 --api-key <redacted>

Prefill

prompt tok/s by context length
8K context57K144 ms TTFT
16K context55.5K293 ms TTFT
32K context51.1K633 ms TTFT
64K context40.5K1,594 ms TTFT
123K context27.4K4,520 ms TTFT
128K context26.4K4,879 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K260.284 ms TTFT516.282 ms TTFT1,004.0100 ms TTFT1,903.8149 ms TTFT3,528.8204 ms TTFT5,972.2279 ms TTFT9,011.6332 ms TTFT13.2K1,055 ms TTFT
16K256.1104 ms TTFT506.6143 ms TTFT975.7134 ms TTFT1,811.8251 ms TTFT3,244.4249 ms TTFT5,169.9208 ms TTFT7,296.6940 ms TTFT9,859.51,630 ms TTFT
32K250.6153 ms TTFT491.3140 ms TTFT926.1209 ms TTFT1,644.3368 ms TTFT2,770.6205 ms TTFT4,038.3384 ms TTFT5,363.82,034 ms TTFT6,483.43,482 ms TTFT
64K242.7212 ms TTFT464.2275 ms TTFT841.7288 ms TTFT1,397.2286 ms TTFT2,072.8294 ms TTFT2,880.22,016 ms TTFT3,424.33,746 ms TTFT3,799.97,275 ms TTFT
123K227.8419 ms TTFT418.1496 ms TTFT713.6459 ms TTFT1,065.2603 ms TTFT1,500.92,215 ms TTFT1,861.94,063 ms TTFT2,047.58,044 ms TTFT2,150.715,922 ms TTFT

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.
PhaseContextRequests at onceStatusTotal tok/sPer request tok/sTTFTTime between tokensAverage active requestsCompleted requestsOutput tokensMeasured for
Prefill8K1Valid57K—144 ms——1——
Prefill16K1Valid55.5K—293 ms——1——
Prefill32K1Valid51.1K—633 ms——1——
Prefill64K1Valid40.5K—1,594 ms——1——
Prefill123K1Valid27.4K—4,520 ms——1——
Prefill128K1Valid26.4K—4,879 ms——1——
Sustained decode8K1Valid260.2264.284 ms3.8 ms1.027,80530.0 s
Sustained decode8K2Valid516.2260.982 ms3.8 ms2.0415,48430.0 s
Sustained decode8K4Valid1,004.0255.4100 ms3.9 ms4.0830,11730.0 s
Sustained decode8K8Valid1,903.8240.9149 ms4.2 ms8.01657,11630.0 s
Sustained decode8K16Valid3,528.8223.3204 ms4.5 ms15.816105,86130.0 s
Sustained decode8K32Valid5,972.2192.4279 ms5.2 ms32.032179,17030.0 s
Sustained decode8K64Valid9,011.6148.9332 ms6.7 ms63.764270,34930.0 s
Sustained decode8K128Valid13.2K104.51,055 ms9.6 ms128.00394,78330.0 s
Sustained decode16K1Valid256.1258.1104 ms3.9 ms1.027,68430.0 s
Sustained decode16K2Valid506.6258.1143 ms3.9 ms2.0415,19830.0 s
Sustained decode16K4Valid975.7246.1134 ms4.1 ms4.0829,27030.0 s
Sustained decode16K8Valid1,811.8229.3251 ms4.4 ms8.0854,34930.0 s
Sustained decode16K16Valid3,244.4205.8249 ms4.9 ms16.01697,32130.0 s
Sustained decode16K32Valid5,169.9165.0208 ms6.1 ms31.832155,09930.0 s
Sustained decode16K64Valid7,296.6115.5940 ms8.7 ms64.00218,90030.0 s
Sustained decode16K128Valid9,859.577.61,630 ms12.9 ms128.00295,80830.0 s
Sustained decode32K1Valid250.6253.2153 ms3.9 ms1.027,51830.0 s
Sustained decode32K2Valid491.3248.3140 ms4.0 ms1.9414,73930.0 s
Sustained decode32K4Valid926.1233.4209 ms4.3 ms3.9427,78130.0 s
Sustained decode32K8Valid1,644.3208.8368 ms4.8 ms8.0849,32330.0 s
Sustained decode32K16Valid2,770.6174.3205 ms5.7 ms15.91683,11630.0 s
Sustained decode32K32Valid4,038.3127.7384 ms7.8 ms31.832121,12730.0 s
Sustained decode32K64Valid5,363.883.72,034 ms12.0 ms64.00160,89630.0 s
Sustained decode32K128Valid6,483.450.83,482 ms19.7 ms128.00194,43230.0 s
Sustained decode64K1Valid242.7245.0212 ms4.1 ms1.017,28130.0 s
Sustained decode64K2Valid464.2234.6275 ms4.3 ms2.0213,92430.0 s
Sustained decode64K4Valid841.7212.9288 ms4.7 ms3.9425,24830.0 s
Sustained decode64K8Valid1,397.2175.7286 ms5.7 ms7.9841,91530.0 s
Sustained decode64K16Valid2,072.8131.0294 ms7.6 ms15.91662,18330.0 s
Sustained decode64K32Valid2,880.289.32,016 ms11.2 ms32.0086,40030.0 s
Sustained decode64K64Valid3,424.353.13,746 ms18.8 ms64.00102,72030.0 s
Sustained decode64K128Valid3,799.929.87,275 ms33.6 ms128.00114,04830.0 s
Sustained decode123K1Valid227.8231.8419 ms4.3 ms1.016,83530.0 s
Sustained decode123K2Valid418.1213.1496 ms4.7 ms2.0212,54330.0 s
Sustained decode123K4Valid713.6181.3459 ms5.5 ms4.0421,40730.0 s
Sustained decode123K8Valid1,065.2135.7603 ms7.4 ms7.9831,95630.0 s
Sustained decode123K16Valid1,500.993.22,215 ms10.7 ms16.0045,02430.0 s
Sustained decode123K32Valid1,861.957.44,063 ms17.4 ms32.0055,84030.0 s
Sustained decode123K64Valid2,047.531.78,044 ms31.5 ms64.0061,37630.0 s
Sustained decode123K128Valid2,150.716.915,922 ms59.0 ms128.0064,51030.0 s