# AMD MI450 The MI400 series is the first accelerator family from this vendor designed from the outset for rack-scale deployment rather than for an eight-accelerator server. The part and the rack were designed together, which is the structural change the generation represents; see [[wiki/AMD Helios|AMD Helios]]. > [!note] The one-sentence version > Designed with its rack rather than for a server, on the newest process, with more memory per accelerator than anything before it. ## What is disclosed Fabrication on a two-nanometer-class process, which places it at the leading edge available; see [[wiki/Foundry|Foundry]] and [[wiki/TSMC|TSMC]]. Twelve stacks of high-bandwidth memory per accelerator, reaching up to 432 gigabytes per device, supplied from three memory manufacturers; see [[wiki/High-Bandwidth Memory|High-Bandwidth Memory]]. A substantial advanced-packaging commitment, reported at ten billion dollars across assembly partners, which is the capacity reservation the part requires to exist in volume; see [[wiki/Advanced Packaging|Advanced Packaging]] and [[wiki/Supply Chain Chokepoint|Supply Chain Chokepoint]]. The family includes variants aimed at different balances of arithmetic, memory, and interconnect, which is the normal shape of an accelerator line serving both training and inference; see [[wiki/Inference Accelerator|Inference Accelerator]]. ## The deployment commitments Several very large customers have signed for this family, and one has indicated a commitment of up to six gigawatts of accelerators from this vendor, beginning with one gigawatt on the rack-scale platform. Those are announcements rather than demonstrated deployments, and they are consequential as announcements because capacity at that scale must be reserved years ahead across foundry, packaging, memory, and power; see [[wiki/Gigawatt Campus|Gigawatt Campus]]. The memory figure is the one to watch. Capacity per accelerator determines how large a model can reside on one device and how many concurrent sequences a serving engine can admit, and both of those translate directly into throughput; see [[wiki/Memory Wall|Memory Wall]] and [[wiki/Continuous Batching|Continuous Batching]]. A memory-forward design is a serving-forward design. ## Relationships - **deployed in:** [[wiki/AMD Helios|AMD Helios]] - **succeeds:** [[wiki/AMD MI350X|AMD MI350X]] - **built from:** [[wiki/Chiplet|Chiplet]] · [[wiki/Advanced Packaging|Advanced Packaging]] · [[wiki/High-Bandwidth Memory|High-Bandwidth Memory]] - **fabricated by:** [[wiki/TSMC|TSMC]] · [[wiki/Foundry|Foundry]] - **connected by:** [[wiki/UALink|UALink]] · [[wiki/Infinity Fabric|Infinity Fabric]] - **programmed through:** [[wiki/ROCm|ROCm]] - **source articles:** [[articles/Vertically Integrating an AI Superpower from AI Factory to Citizen|Vertically Integrating an AI Superpower from AI Factory to Citizen]] · [[articles/The Gigawatt Is the New Measure of War|The Gigawatt Is the New Measure of War]] - **collection route:** [[collections/American Infrastructure|American Infrastructure]] · [[collections/American|American Collection]] ## Evidence posture - **Established:** fabrication on a two-nanometer-class process; twelve stacks of high-bandwidth memory per accelerator reaching up to 432 gigabytes, with three memory suppliers; a reported ten-billion-dollar advanced-packaging commitment across assembly partners; family variants balancing arithmetic, memory, and interconnect differently. - **Announced (not yet demonstrated):** signed commitments from several very large customers, including an indicated commitment of up to six gigawatts of this vendor's accelerators with an initial one gigawatt on the rack-scale platform. - **Strongly indicated:** memory capacity per accelerator as the determinant of serving throughput through residency and batch size; the mechanism is established and per-part figures depend on model and workload. ## Sources - [Helios Is AMD's First AI System To Rival Nvidia Vera Rubin](https://www.cnbc.com/) — CNBC, October 2026; process node, memory stack count and capacity, three memory suppliers, the packaging commitment, customer signings, and the multi-gigawatt commitment.