शुरुआत करेंलीडरबोर्डDecode calculatorमॉडलOrganizationsReportsMatrix runsहार्डवेयरबेंचमार्कमार्केटप्लेसरेंटलProAPI दस्तावेज़
भाषा
Actual Computer — Every computer, one endpoint

Speed test matrix

GLM-5.3-Flash on RTX PRO 6000 Blackwell Server Edition 96GB ×8

sglang · 0.5.6.post3.dev11345+g028ac64f7 · FP88 contexts (8K–1019K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K, 256K, 512K, 1019K, 1024K
by @keennaylmx dev importMeasured

Hardware

Type
Discrete GPU
GPU
RTX PRO 6000 Blackwell Server Edition
GPU count
8
VRAM
768 GB (8 × 96 GB)
OS
linux

Engine

Engine
sglang
Version
0.5.6.post3.dev11345+g028ac64f7
Quantization
FP8
Backend
cuda
Build
cuda-13.2
Commit
028ac64f797f

Run

Served model
zai-org/GLM-5.3-Flash
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_rtxpro6k-incoai-sglang-pr-028ac64f7/bin/sglang" serve --model-path zai-org/GLM-5.3-Flash --served-model-name zai-org/GLM-5.3-Flash --trust-remote-code --tp 8 --reasoning-parser glm45 --tool-call-parser glm47 --context-length 1048576 --mem-fraction-static 0.78 --enable-metrics --host 0.0.0.0 --port 8000 --api-key <redacted> --kv-cache-dtype fp8_e4m3 --dsa-prefill-backend flashinfer_sparse_mla --dsa-decode-backend flashinfer_sparse_mla --linear-attn-backend triton --attention-backend dsa --fp8-gemm-backend cutlass --moe-runner-backend triton

Prefill

prompt tok/s by context length
8K context9,863831 ms TTFT
16K context11.6K1,402 ms TTFT
32K context11.8K2,746 ms TTFT
64K context11.7K5,528 ms TTFT
128K context11.5K11,197 ms TTFT
256K context8,33030,926 ms TTFT
512K context10.3K49,794 ms TTFT
1019K context8,889115,306 ms TTFT
1024K context8,876116,045 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K118.8118 ms TTFT217.1327 ms TTFT383.0467 ms TTFT639.4455 ms TTFT940.1535 ms TTFT1,336.6698 ms TTFT1,915.4704 ms TTFT3,095.71,353 ms TTFT
16K117.5131 ms TTFT219.2220 ms TTFT384.9240 ms TTFT638.2331 ms TTFT932.4540 ms TTFT1,322.2972 ms TTFT1,886.51,844 ms TTFT3,078.78,742 ms TTFT
32K117.2156 ms TTFT219.1254 ms TTFT384.6364 ms TTFT631.5523 ms TTFT933.3964 ms TTFT1,328.31,873 ms TTFT1,886.53,702 ms TTFTOver capacity
64K116.1226 ms TTFT217.3301 ms TTFT380.0479 ms TTFT628.5887 ms TTFT924.11,747 ms TTFT1,315.93,494 ms TTFTOver capacityOver capacity
128K115.7308 ms TTFT215.5472 ms TTFT378.8805 ms TTFT625.21,610 ms TTFT920.33,204 ms TTFTOver capacityOver capacityOver capacity
256K113.0533 ms TTFT210.8918 ms TTFT370.01,792 ms TTFT610.43,586 ms TTFTOver capacityOver capacityOver capacityOver capacity
512K111.01,024 ms TTFT204.72,209 ms TTFT360.04,027 ms TTFTOver capacityOver capacityOver capacityOver capacityOver capacity
1019K103.22,300 ms TTFT191.34,608 ms TTFTOver 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
Prefill8K1Valid9,863.0—831 ms——1——
Prefill16K1Valid11.6K—1,402 ms——1——
Prefill32K1Valid11.8K—2,746 ms——1——
Prefill64K1Valid11.7K—5,528 ms——1——
Prefill128K1Valid11.5K—11,197 ms——1——
Prefill256K1Valid8,330.0—30,926 ms——1——
Prefill512K1Valid10.3K—49,794 ms——1——
Prefill1019K1Valid8,889.0—115,306 ms——1——
Prefill1024K1Valid8,876.0—116,045 ms——1——
Sustained decode8K1Valid118.8118.9118 ms8.4 ms1.003,56530.0 s
Sustained decode8K2Valid217.1108.0327 ms9.3 ms2.006,51430.0 s
Sustained decode8K4Valid383.095.9467 ms10.4 ms4.0011,48830.0 s
Sustained decode8K8Valid639.480.1455 ms12.5 ms8.0019,17630.0 s
Sustained decode8K16Valid940.158.6535 ms17.1 ms16.0028,19230.0 s
Sustained decode8K32Valid1,336.641.8698 ms23.9 ms32.0040,09630.0 s
Sustained decode8K64Valid1,915.429.9704 ms33.4 ms64.0057,47230.0 s
Sustained decode8K128Valid3,095.724.21,353 ms41.3 ms128.0092,80030.0 s
Sustained decode16K1Valid117.5117.6131 ms8.5 ms1.003,52830.0 s
Sustained decode16K2Valid219.2109.6220 ms9.1 ms2.006,57630.0 s
Sustained decode16K4Valid384.996.4240 ms10.4 ms4.0011,56030.0 s
Sustained decode16K8Valid638.280.0331 ms12.5 ms8.0019,15230.0 s
Sustained decode16K16Valid932.458.4540 ms17.1 ms16.0027,96730.0 s
Sustained decode16K32Valid1,322.241.4972 ms24.1 ms32.0039,61630.0 s
Sustained decode16K64Valid1,886.529.61,844 ms33.7 ms64.0056,57630.0 s
Sustained decode16K128Valid3,078.724.48,742 ms41.0 ms128.0092,28830.0 s
Sustained decode32K1Valid117.2117.2156 ms8.5 ms1.003,51530.0 s
Sustained decode32K2Valid219.1109.6254 ms9.1 ms2.006,57430.0 s
Sustained decode32K4Valid384.696.2364 ms10.4 ms4.0011,53630.0 s
Sustained decode32K8Valid631.578.8523 ms12.7 ms8.0018,94430.0 s
Sustained decode32K16Valid933.358.4964 ms17.1 ms16.0027,98530.0 s
Sustained decode32K32Valid1,328.341.61,873 ms24.1 ms32.0039,84030.0 s
Sustained decode32K64Valid1,886.529.63,702 ms33.8 ms64.0056,57630.0 s
Sustained decode32K128Over capacity————————
Sustained decode64K1Valid116.1116.2226 ms8.6 ms1.003,48430.0 s
Sustained decode64K2Valid217.3108.7301 ms9.2 ms2.006,52030.0 s
Sustained decode64K4Valid380.095.0479 ms10.5 ms4.0011,40030.0 s
Sustained decode64K8Valid628.578.6887 ms12.7 ms8.0018,84030.0 s
Sustained decode64K16Valid924.157.91,747 ms17.3 ms16.0027,69630.0 s
Sustained decode64K32Valid1,315.941.33,494 ms24.2 ms32.0039,45630.0 s
Sustained decode64K64Over capacity————————
Sustained decode64K128Over capacity————————
Sustained decode128K1Valid115.7115.7308 ms8.6 ms1.003,46930.0 s
Sustained decode128K2Valid215.5107.8472 ms9.3 ms2.006,46230.0 s
Sustained decode128K4Valid378.895.0805 ms10.5 ms4.0011,36030.0 s
Sustained decode128K8Valid625.278.71,610 ms12.7 ms8.0018,75230.0 s
Sustained decode128K16Valid920.358.53,204 ms17.1 ms16.0027,60030.0 s
Sustained decode128K32Over capacity————————
Sustained decode128K64Over capacity————————
Sustained decode128K128Over capacity————————
Sustained decode256K1Valid113.0113.0533 ms8.9 ms1.003,38630.0 s
Sustained decode256K2Valid210.8105.9918 ms9.4 ms2.006,31630.0 s
Sustained decode256K4Valid370.093.71,792 ms10.7 ms4.0011,08830.0 s
Sustained decode256K8Valid610.478.93,586 ms12.7 ms8.0018,31230.0 s
Sustained decode256K16Over capacity————————
Sustained decode256K32Over capacity————————
Sustained decode256K64Over capacity————————
Sustained decode256K128Over capacity————————
Sustained decode512K1Valid111.0111.11,024 ms9.0 ms1.003,33130.0 s
Sustained decode512K2Valid204.7103.82,209 ms9.6 ms2.006,14430.0 s
Sustained decode512K4Valid360.093.34,027 ms10.7 ms4.0010,80430.0 s
Sustained decode512K8Over capacity————————
Sustained decode512K16Over capacity————————
Sustained decode512K32Over capacity————————
Sustained decode512K64Over capacity————————
Sustained decode512K128Over capacity————————
Sustained decode1019K1Valid103.2103.42,300 ms9.7 ms1.003,09930.0 s
Sustained decode1019K2Valid191.398.84,608 ms10.1 ms2.005,74030.0 s
Sustained decode1019K4Over capacity————————
Sustained decode1019K8Over capacity————————
Sustained decode1019K16Over capacity————————
Sustained decode1019K32Over capacity————————
Sustained decode1019K64Over capacity————————
Sustained decode1019K128Over capacity————————

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