AMD Instinct MI355X
The Instinct MI355X is AMD's flagship 4th-gen CDNA accelerator, an OAM module with 288 GB of HBM3E at 8 TB/s, 256 compute units, and a 1,400 W board power rating, launched 12 June 2025. It is the high-power member of the MI350 series, adds MXFP6 and MXFP4 datatypes, and is built for 8-GPU OAM baseboards joined by Infinity Fabric. GPU Smith specs it as the AMD alternative to B200 and B300 class parts for training and large-model inference, where 288 GB per GPU removes much of the parallelism a smaller-memory part needs.
- 288 GB per GPU is the headline advantage over smaller-memory parts: by arithmetic, a 70B model in FP16 (about 140 GB of weights) fits on one GPU.
- Peak PFLOPS are theoretical at the 2,400 MHz clock and at the datatype named; sustained throughput at 1,400 W depends on cooling and power capping.
- MXFP4 and MXFP6 need framework and kernel support. Verify your serving stack uses them on CDNA 4 before counting on the 10.1 PFLOPS figure.
- The scale-out link is one 128 GB/s port per GPU, so multi-node fabric design (NIC choice, topology) is a separate decision from the node.
- ROCm maturity is the risk, not the silicon: pilot your exact model and framework versions on rented hardware first.
How much HBM does the MI355X have?
288 GB of HBM3E at 8 TB/s peak bandwidth, with full-chip ECC.
What is the MI355X power draw?
AMD lists 1,400 W typical board power for the OAM module.
MI355X vs B200?
On AMD's published peaks the MI355X has more memory (288 GB) and bandwidth (8 TB/s). Delivered performance depends on software; AMD's comparison figures are its own calculations, so benchmark your workload.
Can I buy a single MI355X?
It is an OAM module sold on 8-GPU platforms through OEMs; AMD's PCIe option in this generation is the MI350P.