Erste SchritteLeaderboardDecode calculatorModelleOrganizationsReportsMatrix runsHardwareBenchmarksMarktplatzVermietungenProAPI-Doku
Sprache
Actual Computer — Every computer, one endpoint

Speed test matrix

granite-4.2-3b on H200 NVL 141GB

sglang · 0.5.18 · 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
sglang
Version
0.5.18
Quantization
BF16
Backend
cuda
Build
cuda-13.2

Run

Served model
granite-4.2-3b
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-ibm-sglang/bin/sglang" serve --model-path ibm-granite/granite-4.2-3b --served-model-name granite-4.2-3b --tp 1 --reasoning-parser auto --tool-call-parser auto --context-length 131072 --mem-fraction-static 0.857703 --enable-metrics --host 0.0.0.0 --port 8000 --api-key <redacted> --dtype bfloat16

Prefill

prompt tok/s by context length
8K context47.5K172 ms TTFT
16K context42.2K384 ms TTFT
32K context33K978 ms TTFT
64K context22.1K2,910 ms TTFT
123K context13.7K8,997 ms TTFT
128K context13.5K9,491 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K202.147 ms TTFT358.762 ms TTFT835.682 ms TTFT1,476.5107 ms TTFT2,001.1186 ms TTFT2,427.3332 ms TTFT2,734.5658 ms TTFTOver capacity
16K240.466 ms TTFT104.194 ms TTFT506.4134 ms TTFT1,199.4160 ms TTFT1,103.3339 ms TTFT1,650.4649 ms TTFT1,814.41,254 ms TTFTOver capacity
32K211.490 ms TTFT48.9163 ms TTFT118.1229 ms TTFT211.7386 ms TTFT663.3708 ms TTFT732.31,331 ms TTFTOver capacityOver capacity
64K191.4148 ms TTFT45.0252 ms TTFT89.2416 ms TTFT177.5745 ms TTFT253.61,458 ms TTFTOver capacityOver capacityOver capacity
123K144.8256 ms TTFT32.0479 ms TTFT63.7780 ms TTFT100.81,515 ms TTFTOver capacityOver capacityOver capacityOver capacity

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.
PhaseContextRequests at onceStatusTotal tok/sPer request tok/sTTFTTime between tokensAverage active requestsCompleted requestsOutput tokensMeasured for
Prefill8K1Valid47.5K—172 ms——1——
Prefill16K1Valid42.2K—384 ms——1——
Prefill32K1Valid33K—978 ms——1——
Prefill64K1Valid22.1K—2,910 ms——1——
Prefill123K1Valid13.7K—8,997 ms——1——
Prefill128K1Valid13.5K—9,491 ms——1——
Sustained decode8K1Valid202.1212.447 ms4.9 ms—16,06230.0 s
Sustained decode8K2Valid358.7197.962 ms5.3 ms—210,76230.0 s
Sustained decode8K4Valid835.6207.682 ms4.8 ms—425,06530.0 s
Sustained decode8K8Valid1,476.5183.5107 ms5.5 ms—844,29130.0 s
Sustained decode8K16Valid2,001.1136.2186 ms7.4 ms—1660,02930.0 s
Sustained decode8K32Valid2,427.376.9332 ms13.0 ms—072,81530.0 s
Sustained decode8K64Valid2,734.542.6658 ms23.5 ms—082,04830.0 s
Sustained decode8K128Over capacity————————
Sustained decode16K1Valid240.4241.966 ms4.1 ms—27,21130.0 s
Sustained decode16K2Valid104.152.594 ms19.1 ms—03,12230.0 s
Sustained decode16K4Valid506.4116.2134 ms8.6 ms—015,19230.0 s
Sustained decode16K8Valid1,199.4152.7160 ms6.5 ms—835,98130.0 s
Sustained decode16K16Valid1,103.369.5339 ms14.4 ms—033,08730.0 s
Sustained decode16K32Valid1,650.451.5649 ms19.4 ms—049,50330.0 s
Sustained decode16K64Valid1,814.428.81,254 ms34.7 ms—054,39830.0 s
Sustained decode16K128Over capacity————————
Sustained decode32K1Valid211.4212.590 ms4.7 ms—16,34330.0 s
Sustained decode32K2Valid48.924.5163 ms40.8 ms—01,46630.0 s
Sustained decode32K4Valid118.128.7229 ms34.8 ms—03,54330.0 s
Sustained decode32K8Valid211.726.6386 ms37.6 ms—06,34430.0 s
Sustained decode32K16Valid663.339.3708 ms25.5 ms—019,88830.0 s
Sustained decode32K32Valid732.323.31,331 ms42.9 ms—021,96830.0 s
Sustained decode32K64Over capacity————————
Sustained decode32K128Over capacity————————
Sustained decode64K1Valid191.4191.0148 ms5.2 ms—15,74330.0 s
Sustained decode64K2Valid45.022.5252 ms44.4 ms—01,34830.0 s
Sustained decode64K4Valid89.222.3416 ms44.8 ms—02,67630.0 s
Sustained decode64K8Valid177.521.9745 ms45.7 ms—05,32030.0 s
Sustained decode64K16Valid253.616.11,458 ms62.0 ms—07,60030.0 s
Sustained decode64K32Over capacity————————
Sustained decode64K64Over capacity————————
Sustained decode64K128Over capacity————————
Sustained decode123K1Valid144.8148.6256 ms6.7 ms—14,34230.0 s
Sustained decode123K2Valid32.015.8479 ms63.3 ms—096030.0 s
Sustained decode123K4Valid63.716.0780 ms62.6 ms—01,90830.0 s
Sustained decode123K8Valid100.812.81,515 ms78.1 ms—03,02430.0 s
Sustained decode123K16Over capacity————————
Sustained decode123K32Over capacity————————
Sustained decode123K64Over capacity————————
Sustained decode123K128Over capacity————————

Over capacity: not run because this many requests at this context would exceed the server's KV cache.