Menu
Back previous/next report: k j

Krai HPE Cray XD670

Submitted by Kraion 2026-02-27 Published on 2026-03-16

Submission Info
Submitter Krai
Submission ID krai-Submission2-SGLang
Submission Status
Submission Date 2026-02-27
Publication Date 2026-03-16
Division Closed
Model
Model Name Llama-3.1-70B
Division Closed
Model Precision BF16
Calibration Link vLLM
Model Notes
Datasets
Dataset Name OpenOrca
Dataset Type Accuracy + Performance
Average Input Tokens 222.33
Dataset Link GitHub

System TPS vs. Tokens/s per User

50100150200System TPS102030Tokens/s per User

System TPS vs. Concurrency

50100150200System TPS0102030# of queries

Time to First Token P95 vs. Concurrency

100200300TTFT P95 (s)0102030# of queries

Tokens/s per User vs. Concurrency

102030Tokens/s per User0102030# of queries
SUT Summary
NVIDIA
System Name HPE Cray XD670
System Availability Available
System Category Datacenter · Cloud
System Size 8x H200-SXM-141GB
Node Types 1
Total Node Count 1
Max Supported Concurrency
Node Type ID 1
Node Count 1
Software
Red Hat
Framework SGLang v0.5.9
Serving Framework SGLang v0.5.9
Inference Backend
Driver
Link to Container
Operating System host: Red Hat Enterprise Linux 9.4; container: Ubuntu 22.04.5 LTS
Filesystem
Other Software lmsysorg/sglang:v0.5.9; rootless Docker v27.4.1; CUDA 13.0; Driver 580.105.08
Software Notes Powered by the KRAI X automation technology
Hardware

Processor

Processor Model Name Intel(R) Xeon(R) Platinum 8562Y+
Processors per Node 2
Cores Per Processor 32
VCPUs Per Processor

Accelerator

Accelerator Model Name NVIDIA H200-SXM5-141GB
Accelerators per Node 8
Memory Type
Memory Capacity 141 GB
Accelerator Interconnect 18x 4th Gen NVLink 900GB/s
Host-Accelerator Interconnect PCIe Gen5 Switch

Host / Storage

Host Memory Capacity 2 TiB
Memory Configuration 32x HPE Cray Supercomputing XD 64GB (1x64GB) Dual Rank x4 DDR5-5600 Registered Standard Memory
Storage Capacity 15 TB SSD; 929 TB HPE ClusterStor E1000
Storage Type HPE Cray XD670 1.92 TB NVMe Read Intensive SFF U.3 7450 SSD; HPE Cray ClusterStor E1000

Networking

Network Interfaces 8x HPE Slingshot SA210S Ethernet 200Gb 1-port PCIe NIC; 8x HPE InfiniBand NDR/Ethernet 400Gb 1-port OSFP PCIe5 x16 MCX75310AAS-NEAT Adapter; 1x HPE InfiniBand HDR/Ethernet 200Gb 2-port QSFP56 PCIe4 x16 MCX653106A-HDAT Adapter
Network Details Infiniband; Ethernet
Network Topology

Other

Other Hardware
Cooling Air-cooled
Hardware Notes
Run Data
Field NameRun 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8Run 9Run 10
Run Date02/25/202602/25/202602/25/202602/25/202602/25/202602/25/202602/25/202602/25/202602/25/202602/25/2026
Concurrency12345711152229
System Tokens/Second32.560.384.5105.9141.1169.2165.1215.9210.7206.8
Tokens/Second per User35.733.131.029.125.923.215.114.89.67.1
TTFT P95 (ms)77394.053346.644717.739311.934046.231178.237216.135919.6149861.7310115.5
Utilization15.1%27.9%39.1%49.1%65.4%78.3%76.4%100.0%97.6%95.8%
Configuration SummaryPP 1, TP 1, DP 8PP 1, TP 1, DP 8PP 1, TP 1, DP 8PP 1, TP 1, DP 8PP 1, TP 1, DP 8PP 1, TP 1, DP 8PP 1, TP 1, DP 8PP 1, TP 1, DP 8PP 1, TP 1, DP 8PP 1, TP 1, DP 8
Config Type
Config Region
Target QPS
Result State
Measured Accuracy Score
Benchmark Version
Result Version
Archive URI
QPS0.13390.24900.34870.43680.58210.69780.68080.89060.86900.8530
Total Output Tokens5,965,0625,949,9865,956,1275,960,4905,958,3915,958,4175,958,4175,958,4175,958,4175,958,417
Average Output Tokens
Run Duration (s)2,608.961,403.031,001.83799.72600.14500.68513.15392.25402.04409.54
Total Queries24,57624,57624,57624,57624,57624,57624,57624,57624,57624,576
Total Queries Issued
Total Queries Failed
Time to First Token (TTFT) P95 (ms)
Minimum3804.23870.33885.03933.83942.83918.14061.54001.14186.33979.4
Average16612.914477.513237.512994.013276.513821.018252.720447.630608.349760.2
P507336.07781.17925.88249.28707.98995.811218.410976.913437.015557.7
P9039348.434464.728613.527112.924656.822656.526358.024353.330126.4115098.6
P9577394.053346.644717.739311.934046.231178.237216.135919.6149861.7310115.5
P99160424.094628.575703.567214.478077.5141026.3203491.0250019.3383753.9568735.2
P999257146.8150344.3128007.498512.9127481.5166551.6233459.8302208.0443234.8875006.1
Maximum432156.7253484.4191495.3177778.9246347.1215834.3239227.1314870.7445388.9986368.3
Time Per Output Token (TPOT) (ms)
Minimum1739.61817.9433.2852.71303.94.53521.04.734080.812.9
Average1913.62062.72189.72318.02574.12819.94209.84170.36056.87683.0
P501895.82043.52170.62297.72552.62798.84174.84126.45975.57576.5
P901971.52136.22281.52421.32702.42958.04376.54364.76267.57942.6
P952014.12193.22348.02496.32786.83049.74512.54505.26446.78176.5
P992301.52526.52718.52945.33244.03609.45176.95359.07921.310225.8
P9993787.83913.64063.64544.16169.56971.710242.010569.617329.421611.3
Maximum5906.510572.87840.710317.311354.816287.133604.165392.8179052.6140525.5
Query Latency (ms)
Minimum6464.47437.56103.96189.96327.97459.111522.58906.313885.914818.1
Average475068.8507246.2537206.7567696.6629275.8688793.61024810.81015719.21470435.51877016.0
P50450588.0480478.9507270.6537205.5594482.9652228.2967515.9960000.71387169.91773763.1
P90798950.6849053.4902221.4949972.51053911.61158329.81719060.91706131.12466617.33149608.3
P95982583.11053935.91109809.91181066.91309625.71436293.42140146.72116034.33067674.33900937.7
P991892682.21946212.82070760.62239232.92456970.42662766.43992459.53942232.05689630.47217454.0
P9992008062.32145613.62268195.72394851.02670375.12923674.64353892.84317420.56249831.77957346.9
Maximum2240850.42281521.12356602.82470723.82728848.53017352.24434690.44431438.46487397.88335438.8