AMD Instinct MI250X: The Complete Guide

AMD’s own launch material frames the MI200 series — MI250X, MI250 and MI210 — as built to “power exascale systems”, naming Frontier at Oak Ridge National Laboratory as the flagship deployment before it had even run its first benchmark 1. MI250X and MI250 use AMD’s new CDNA 2 architecture; MI210 is a smaller, single-die, PCIe-card version of the same generation 34.

The bet paid off in May 2022: Frontier reached 1.102 exaflops on the industry’s standard benchmark, officially the first machine ever to cross the exascale line, according to both Oak Ridge’s own announcement and the independent TOP500 list 25. Every node in Frontier pairs one AMD EPYC CPU with four MI250X chips — which is why this one chip, more than any other AMD part before it, is the one AMD points to when it talks about entering the top tier of scientific and AI computing.

AMD’s own words for the packaging: MI250X and MI250 use “2.5D Elevated Fanout Bridge (EFB) technology” to connect two Graphics Compute Dies, joined by up to eight Infinity Fabric links per package 1. MI210 drops to a single die with up to three Infinity Fabric links, in a PCIe card rather than the OAM module MI250X and MI250 use 4. AMD has not disclosed a transistor count for any MI200-series chip in its press release, datasheet or product pages; this page does not repeat unsourced third-party figures.

Figures from AMD’s own MI200-series datasheet and MI210 brochure 34. AMD’s MI210 brochure lists FP64 and FP32 matrix performance in one combined row (45.3 TFLOPS); this page reports that figure under FP32 matrix, consistent with how it reads the equivalent combined row for MI250X and MI250. AMD’s own material says “6nm FinFET” but does not name the manufacturing partner on these pages; this page does not add a foundry name AMD did not state.

This is a real gap in AMD’s own public material, not an oversight on this page: the MI200-series datasheet states the process node as 6nm FinFET but names no manufacturing partner 34. A foundry name for MI200 circulates widely in secondary technology coverage, but this page reports only what AMD itself states.

Oak Ridge’s own announcement: Frontier pairs one third-generation AMD EPYC processor with four MI250X GPUs per node, across more than 9,400 nodes — 9,472 CPUs and 37,888 GPUs in total, connected by roughly 90 miles of cabling 2. The independent TOP500.org list for June 2022 records Frontier’s official benchmark result: 1.102 exaflops sustained (Rmax) against a theoretical peak of 1.686 exaflops (Rpeak), on an HPE Cray EX235a system with Slingshot-11 interconnect — TOP500’s own listing calls it “the first true exascale machine” 5. The same system also topped the Green500 efficiency list at 62.68 gigaflops per watt, per Oak Ridge’s own figures 2.

Unlike its successor MI300X, which several cloud providers rent by the hour with published prices, MI250X was sold mainly to supercomputer integrators and enterprise buyers rather than offered as a metered cloud product. Figures for an MI250X hourly rental price circulate on third-party GPU-rental aggregator sites, but none of them is a named cloud provider’s own official pricing page, so this page does not repeat them as fact.

The clearest independent performance record for MI250X is not an AMD benchmark at all — it is Frontier’s own TOP500 result: 1.102 exaflops sustained, the first system ever to do so 5. That is a system-level result (37,888 MI250X GPUs working together), not a single-chip benchmark, and this page reports it as such rather than dividing it down to an implied per-chip figure AMD itself does not publish.

MI100 was AMD’s own first CDNA-architecture chip, announced 16 November 2020, rated by AMD at 11.54 TFLOPS of peak FP64 — AMD’s own release calls it “the first to surpass the 10 teraflops (FP64) performance barrier”, comparing it directly to NVIDIA’s A100 at what AMD stated as 9.7 TFLOPS peak double precision 6. MI250X’s direct successor is the MI300 series, which moved to a newer CDNA 3 architecture and, in one variant, combined CPU and GPU cores on a single chip for the first time 7.

AMD’s own description of MI250X’s packaging — two Graphics Compute Dies bridged together with up to eight Infinity Fabric links 1 — is the same basic idea AMD would later scale much further in MI300X’s eight-chiplet design. Frontier’s Green500-topping efficiency figure (62.68 gigaflops per watt) is also easy to miss next to its exascale headline, but it is the number Oak Ridge itself highlighted alongside the speed record 2.