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

Speed test matrix

GLM-5.3-NVFP4 on H200 SXM 141GB ×8

sglang · 0.5.6.post3.dev10338+g787299797 · NVFP47 contexts (8K–507K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K, 256K, 507K, 512K
by @keennaylmx dev importMeasured

Hardware

Type
Discrete GPU
GPU
H200 SXM
GPU count
8
VRAM
1128 GB (8 × 141 GB)
OS
linux

Engine

Engine
sglang
Version
0.5.6.post3.dev10338+g787299797
Quantization
NVFP4
Backend
cuda
Build
cuda-13.2
Commit
787299797931

Run

Served model
glm
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-glm53flash-dflash2-sglang-pr-37818/bin/sglang" serve --model-path Inferact/GLM-5.3-NVFP4 --served-model-name glm --tp 8 --reasoning-parser glm45 --tool-call-parser glm47 --context-length 524288 --mem-fraction-static 0.80 --enable-metrics --host 0.0.0.0 --port 8000 --api-key <redacted> --moe-runner-backend marlin

Prefill

prompt tok/s by context length
8K context9,653849 ms TTFT
16K context9,6891,675 ms TTFT
32K context9,4713,412 ms TTFT
64K context9,1067,084 ms TTFT
128K context8,40115,341 ms TTFT
256K context7,17535,904 ms TTFT
507K context5,59291,217 ms TTFT
512K context5,56392,586 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K114.264 ms TTFT207.4103 ms TTFT366.3153 ms TTFT622.9183 ms TTFT992.0267 ms TTFT1,513.7560 ms TTFTOver capacityOver capacity
16K114.882 ms TTFT208.1119 ms TTFT366.6189 ms TTFT621.3276 ms TTFT991.0618 ms TTFTOver capacityOver capacityOver capacity
32K114.0121 ms TTFT206.8196 ms TTFT361.8304 ms TTFT617.5601 ms TTFTOver capacityOver capacityOver capacityOver capacity
64K113.5189 ms TTFT205.5339 ms TTFT360.7632 ms TTFT610.21,299 ms TTFTOver capacityOver capacityOver capacityOver capacity
128K112.6322 ms TTFT203.4635 ms TTFT357.41,171 ms TTFTOver capacityOver capacityOver capacityOver capacityOver capacity
256K110.9587 ms TTFT199.31,163 ms TTFTOver capacityOver capacityOver capacityOver capacityOver capacityOver capacity
507K107.41,325 ms TTFTOver capacityOver capacityOver capacityOver 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
Prefill8K1Valid9,653.0—849 ms——1——
Prefill16K1Valid9,689.0—1,675 ms——1——
Prefill32K1Valid9,471.0—3,412 ms——1——
Prefill64K1Valid9,106.0—7,084 ms——1——
Prefill128K1Valid8,401.0—15,341 ms——1——
Prefill256K1Valid7,175.0—35,904 ms——1——
Prefill507K1Valid5,592.0—91,217 ms——1——
Prefill512K1Valid5,563.0—92,586 ms——1——
Sustained decode8K1Valid114.2114.264 ms8.8 ms1.003,42630.0 s
Sustained decode8K2Valid207.4103.8103 ms9.6 ms2.006,22230.0 s
Sustained decode8K4Valid366.391.6153 ms10.9 ms4.0010,98830.0 s
Sustained decode8K8Valid622.978.0183 ms12.8 ms8.0018,68030.0 s
Sustained decode8K16Valid992.062.2267 ms16.1 ms16.0029,76030.0 s
Sustained decode8K32Valid1,513.747.6560 ms21.0 ms32.0045,40830.0 s
Sustained decode8K64Over capacity————————
Sustained decode8K128Over capacity————————
Sustained decode16K1Valid114.8114.882 ms8.7 ms1.003,44530.0 s
Sustained decode16K2Valid208.1104.0119 ms9.6 ms2.006,24430.0 s
Sustained decode16K4Valid366.691.6189 ms10.9 ms4.0010,99630.0 s
Sustained decode16K8Valid621.377.7276 ms12.9 ms8.0018,64030.0 s
Sustained decode16K16Valid991.062.2618 ms16.1 ms16.0029,72830.0 s
Sustained decode16K32Over capacity————————
Sustained decode16K64Over capacity————————
Sustained decode16K128Over capacity————————
Sustained decode32K1Valid114.0114.0121 ms8.8 ms1.003,42030.0 s
Sustained decode32K2Valid206.8103.4196 ms9.7 ms2.006,20430.0 s
Sustained decode32K4Valid361.890.5304 ms11.0 ms4.0010,85230.0 s
Sustained decode32K8Valid617.577.4601 ms12.9 ms8.0018,52030.0 s
Sustained decode32K16Over capacity————————
Sustained decode32K32Over capacity————————
Sustained decode32K64Over capacity————————
Sustained decode32K128Over capacity————————
Sustained decode64K1Valid113.5113.5189 ms8.8 ms1.003,40630.0 s
Sustained decode64K2Valid205.5102.9339 ms9.7 ms2.006,16630.0 s
Sustained decode64K4Valid360.790.7632 ms11.0 ms4.0010,82030.0 s
Sustained decode64K8Valid610.277.31,299 ms12.9 ms8.0018,30430.0 s
Sustained decode64K16Over capacity————————
Sustained decode64K32Over capacity————————
Sustained decode64K64Over capacity————————
Sustained decode64K128Over capacity————————
Sustained decode128K1Valid112.6112.6322 ms8.9 ms1.003,37830.0 s
Sustained decode128K2Valid203.4102.1635 ms9.8 ms2.006,10230.0 s
Sustained decode128K4Valid357.490.21,171 ms11.1 ms4.0010,72030.0 s
Sustained decode128K8Over capacity————————
Sustained decode128K16Over capacity————————
Sustained decode128K32Over capacity————————
Sustained decode128K64Over capacity————————
Sustained decode128K128Over capacity————————
Sustained decode256K1Valid110.9110.8587 ms9.0 ms1.003,32530.0 s
Sustained decode256K2Valid199.3100.41,163 ms10.0 ms2.005,97830.0 s
Sustained decode256K4Over capacity————————
Sustained decode256K8Over capacity————————
Sustained decode256K16Over capacity————————
Sustained decode256K32Over capacity————————
Sustained decode256K64Over capacity————————
Sustained decode256K128Over capacity————————
Sustained decode507K1Valid107.4107.41,325 ms9.3 ms1.003,22230.0 s
Sustained decode507K2Over capacity————————
Sustained decode507K4Over capacity————————
Sustained decode507K8Over capacity————————
Sustained decode507K16Over capacity————————
Sustained decode507K32Over capacity————————
Sustained decode507K64Over capacity————————
Sustained decode507K128Over capacity————————

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