MLPerf Inference v6.0

Copyright 2019 - 2025 MLCommons

February 2026

AMD - AMD MI355X Custom Cluster

Submitted by: AMD

Accelerator Details

accelerator_model_nameAMD Instinct MI355X 288GB HBM3e (x87)
accelerators_per_node8
accelerator_memory_capacity288 GB
accelerator_host_interconnectPCIe Gen 5 x16
accelerator_frequency
accelerator_interconnectXGMI
accelerator_interconnect_topology
accelerator_memory_configurationHBM3E
accelerator_on-chip_memories

Processor and Memory Details

host_processor_model_nameAMD EPYC 9575F
host_processors_per_node8
host_processor_core_count64
host_processor_frequency
host_memory_capacity3.0 TB
host_memory_configuration24 x 128GB Micron Technology MTC40F2047S1RC64BB1 QSFF
host_processor_caches
host_processor_interconnect

Other Hardware Details

coolingPassive & Active
disk_controllers
disk_drives
hw_notes
other_hardware
power_management
power_supply_details
power_supply_quantity_and_rating_watts

Network and Interconnect Details

host_network_card_count4 x 10Gbit/s
host_networking2 x Ethernet Controller X710 for 10GBASE-T, 2 x Ethernet Controller 10G X550T
host_networking_topologyEthernet on switching network
network_speed_mbit
nics_enabled_connected
nics_enabled_firmware
nics_enabled_os
number_of_type_nics_installed

Software Details

boot_firmware_version
frameworkPyTorch 2.9.1+git8907517, ROCm 7.0.0
management_firmware_version
nics_enabled_firmware
operating_systemUbuntu 22.04.5 LTS
other_software_stackhipblaslt-1.0.0.70000-38~22.04, vllm-11b6af52, aiter-6af8b687
sw_notes

Results Table

Model Accuracy Target Server Interactive Offline
Metric Performance Metric Performance Metric Performance
llama2-70b-99ROUGE1: 43.9869, ROUGE2: 21.8148, ROUGEL: 28.33, TOKENS_PER_SAMPLE: 265.005Tokens/s 1016380.000Tokens/s 785522.000Tokens/s 1042110.000
llama2-70b-99.9ROUGE1: 44.3868, ROUGE2: 22.0132, ROUGEL: 28.5876, TOKENS_PER_SAMPLE: 265.005Tokens/s 1016380.000Tokens/s 785522.000Tokens/s 1042110.000