Menu
Back previous/next report: k j

Krai Dell PowerEdge XE9680

Submitted by Kraion 2026-06-26

Submission Info
Submitter Krai
Submission ID d2680e81-1dbc-4001-a438-0bab4351a118
Submission Status PUBLISHED
Submission Date 2026-06-26
Publication Date
Division Standardized
Model
Model Name Llama 3 8B
Division Standardized
Model Precision FP16
Calibration Link GitHub
Model Notes W4A8 quantization with vLLM's quantization tools, with fp8 kv cache. No other model transformations applied.
Datasets
Dataset Name CNN-Daily Mail
Dataset Type text
Average Input Tokens 800
Dataset Link GitHub

System TPS vs. Tokens/s per User

5.0k10k15k20k25kSystem TPS50100Tokens/s per User

System TPS vs. Concurrency

5.0k10k15k20k25kSystem TPS05001,0001,5002,000# of queries

Time to First Token P95 vs. Concurrency

0246TTFT P95 (s)05001,0001,5002,000# of queries

Tokens/s per User vs. Concurrency

50100Tokens/s per User05001,0001,5002,000# of queries
SUT Summary
Intel
System Name Dell PowerEdge XE9680
System Availability Available
System Category Datacenter · —
System Size 8x MI300X-192GB
Node Types 1
Total Node Count 1
Max Supported Concurrency 2,048
Node Type ID 0
Node Count 1
Software
Ubuntu
Framework vLLM v0.19.1
Serving Framework vLLM v0.19.1
Inference Backend ROCm
Driver ROCm Driver 6.10.5
Link to Container https://hub.docker.com/layers/rocm/vllm/rocm7.13.0_gfx94X-dcgpu_ubuntu24.04_py3.13_pytorch_2.10.0_vllm_0.19.1
Operating System ubuntu 22.04
Filesystem ext4 ramfs vfat xfs
Other Software ROCm Driver 6.10.5
Software Notes Powered by the KRAI X automation technology
Hardware

Processor

Processor Model Name Intel(R) Xeon(R) Platinum 8460Y+
Processors per Node 2
Cores Per Processor 80
VCPUs Per Processor 160

Accelerator

Accelerator Model Name AMD Instinct MI300X-NPS1-SPX-192GB-750W
Accelerators per Node 8
Memory Type HBM3
Memory Capacity 192GB
Accelerator Interconnect XGMI
Host-Accelerator Interconnect PCIe Gen5 x16

Host / Storage

Host Memory Capacity 2.1T
Memory Configuration 32 x 64GB DDR5-4800
Storage Capacity 3.1 GB NVMe SSD, 12.1 TB SSD
Storage Type NVMe SSD

Networking

Network Interfaces 2x Ethernet
Network Details Ethernet
Network Topology

Other

Other Hardware
Cooling Air-cooled
Hardware Notes
Run Data
Field NameRun 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
Run Date06/26/202606/26/202606/26/202606/26/202606/26/202606/26/202606/26/202606/26/2026
Concurrency3240641282565121,0242,048
System Tokens/Second4,401.95,078.17,106.611,442.016,362.321,290.225,082.627,268.5
Tokens/Second per User137.6127.0111.089.463.941.624.513.3
TTFT P95 (ms)157.2166.0225.8503.9939.91796.93325.76354.4
Utilization16.1%18.6%26.1%42.0%60.0%78.1%92.0%100.0%
Configuration Summaryllama3_1_8b-benchmark · online · c32llama3_1_8b-benchmark · online · c40llama3_1_8b-benchmark · online · c64llama3_1_8b-benchmark · online · c128llama3_1_8b-benchmark · online · c256llama3_1_8b-benchmark · online · c512llama3_1_8b-benchmark · online · c1024llama3_1_8b-benchmark · online · c2048
Config Typeonlineonlineonlineonlineonlineonlineonlineonline
Config Region
Target QPS
Result Statecompletecompletecompletecompletecompletecompletecompletecomplete
Measured Accuracy Score
Benchmark Versiondad7d1adad7d1adad7d1adad7d1adad7d1adad7d1adad7d1adad7d1a
Result Version0.1.00.1.00.1.00.1.00.1.00.1.00.1.00.1.0
Archive URIgs://mlperf-endpoints-blob/runs/7c71d0df-5161-4783-a5df-85e1d133ccdd/archive.tar.gzgs://mlperf-endpoints-blob/runs/0ff21ffd-691f-426e-918d-4d9d1071eb54/archive.tar.gzgs://mlperf-endpoints-blob/runs/abd3c07a-f225-40ed-9564-4f508bb47f67/archive.tar.gzgs://mlperf-endpoints-blob/runs/0524d1f0-e10a-4138-9a39-17b383f3c70b/archive.tar.gzgs://mlperf-endpoints-blob/runs/d84fb2c8-75cd-4a6b-958d-dfa8c9bd6fb1/archive.tar.gzgs://mlperf-endpoints-blob/runs/d9ee6ede-c0a0-459b-9e75-d8fb5895ed81/archive.tar.gzgs://mlperf-endpoints-blob/runs/af087ba5-a0e4-4952-9bc9-f17dffe21f47/archive.tar.gzgs://mlperf-endpoints-blob/runs/4e0753bc-22fb-49f0-934e-6c7b9ba16224/archive.tar.gz
QPS34.390239.672455.520189.3910127.8304166.3300195.9576213.0356
Total Output Tokens17,111,02917,111,03017,111,02317,111,03117,111,02717,111,02817,111,03017,111,029
Average Output Tokens128128128128128128128128
Run Duration (s)3,887.163,369.602,407.781,495.451,045.76803.70682.19627.50
Total Queries133,680133,680133,680133,680133,680133,680133,680133,680
Total Queries Issued133,680133,680133,680133,680133,680133,680133,680133,680
Total Queries Failed00000000
Time to First Token (TTFT) P95 (ms)
Minimum18.820.921.524.228.436.746.580.2
Average82.583.397.0283.5619.21223.32225.23805.6
P5070.068.374.5314.5728.41468.62481.24097.9
P90140.1147.1192.0468.0898.81732.63209.46167.1
P95157.2166.0225.8503.9939.91796.93325.76354.4
P99188.0200.2282.5565.81021.31906.03499.06648.5
P999221.0236.5338.5638.71118.32049.93655.56957.3
Maximum264.4274.5403.2696.31217.22127.93719.77135.4
Time Per Output Token (TPOT) (ms)
Minimum5.555.635.916.357.157.527.4117.2
Average6.687.288.319.0410.914.623.445.2
P506.687.298.378.689.9112.421.343.7
P907.187.829.0610.914.721.835.565.8
P957.327.969.2611.415.323.038.370.4
P997.608.259.6412.116.324.440.974.4
P9997.928.5510.112.817.325.742.876.8
Maximum8.408.9510.515.018.527.644.779.8
Query Latency (ms)
Minimum753.8812.2816.6885.31044.01026.01021.72605.4
Average930.31008.01152.51431.32000.93073.15202.39541.4
P50926.91005.61149.71430.72001.63079.65214.39577.9
P90983.01066.11229.01549.62150.23275.15489.010001.9
P951000.61085.11253.81586.62192.13335.75561.410143.4
P991036.41121.31302.61661.92284.33445.65711.610665.1
P9991077.21162.31356.81747.12410.13565.05990.411592.2
Maximum1149.31217.61467.21961.82461.93630.16290.811829.1