AMD revealed details of its Instinct MI400 artificial intelligence accelerator series at Financial Analyst Day 2025, targeting a 2026 launch and built on the new CDNA 5 architecture. The company claims peak performance of up to 40 FP4 and 20 FP8 PFLOPs, roughly twice the compute of the current MI350. Memory moves from HMB3e to HBM4, increasing per‑GPU capacity from 288 GB to 432 GB and raising total memory bandwidth from 8 TB/s to 19.6 TB/s. Each GPU will also provide 300 GB/s of scale-out bandwidth and broader artificial intelligence data format support alongside expanded AI pipelines.
The MI400 family will include two models. The Instinct MI455X is positioned for large-scale artificial intelligence training and inference, while the MI430X targets high performance computing and sovereign deployments. The MI430X retains the HBM4 memory subsystem but adds hardware support for FP64 and hybrid CPU plus GPU compute capability. AMD is also shifting its packaging approach from CoWoS-S to CoWoS-L, the local silicon interconnect, as part of the MI400 design.
AMD directly compared the MI400 lineup to NVIDIA’s upcoming Vera Rubin series, saying MI400 offers similar compute performance and memory bandwidth while delivering 1.5 times the memory capacity and scale-out bandwidth. Dr. Lisa Su stated the Instinct MI400 accelerators will be available to clients and partners from day one. AMD also confirmed an annual refresh cadence for its Instinct products and said the next-generation MI500 series is already in advanced design stages with a planned 2027 launch.
