COMPUTE ATLAS Supercomputer supply-chain graph
1540 systems 545 sites Sourced data

Parts/Accelerators/AMD Instinct MI250X

Accelerators · CDNA 2 · 2021

AMD Instinct MI250X

Dual-die GPU with 128 GB HBM2e, connected to the host by Infinity Fabric rather than PCIe in the Frontier node design.

Why this part

Two dies on one package, which AMD called a multi-chip module and which software sees as two devices unless told otherwise. Frontier's selection of it was the first time a US exascale flagship went to a vendor other than IBM or Intel at the accelerator layer, and it was a real risk: ROCm in 2021 was substantially less mature than CUDA, and the laboratories took on a porting effort to avoid single-vendor dependence. Frontier reaching number one validated the bet, and the maturity gap remains the honest counterargument.

Systems using this part

Systems
15
Rmax underneath
2.39 EFlop/s
Units recorded
50,592
5 of 15 edges disclose a quantity
System Role Country Units System Rmax Supplied by Rank
HPC6
2024-11
Accelerator Italy - 477.9 PFlop/s AMD #8 →
Svartisen
2024
Accelerator Norway - 7.40 PFlop/s AMD #156 →
Explorer-WUS3
2023
Accelerator USA - 53.96 PFlop/s AMD #48 →
LUMI
2022-06
Accelerator Finland 11,912 379.7 PFlop/s AMD #3 →
Frontier
2022-05
Accelerator USA 37,632 1.35 EFlop/s AMD #8 →
Adastra
2022
Accelerator France - 46.10 PFlop/s AMD #55 →
Setonix
2022
Accelerator Australia - 27.16 PFlop/s AMD #81 →
Dardel
2022
Accelerator Sweden - 9.90 PFlop/s AMD #77 →
Tioga
2022
Accelerator USA - 4.55 PFlop/s AMD #121 →
rzVernal
2022
Accelerator USA 152 - AMD -
Tenaya
2022
Accelerator USA - 3.41 PFlop/s AMD #200 →
Frontier TDS
2021
Accelerator USA - 19.20 PFlop/s AMD #1 →
Crusher
Accelerator USA 768 - AMD -
Lockhart TDS
Accelerator USA - 6.06 PFlop/s AMD #214 →
Ruth
Accelerator USA 128 - AMD -

“Units recorded” sums only the edges where a public source states a quantity. It is a floor, not a total, and should never be read as installed base.

Related news

All news →

Further reading