AMD Instinct MI350: 288 GB HBM3E, 8 TB/s, and up to 10 PFLOPS FP4 against Blackwell

Training and serving the largest AI models requires data-centre accelerators: huge boards with nothing in common with gaming cards. For years, anyone who needed them bought NVIDIA and that was it, because no credible alternative existed.

On 12 June 2025, at its own Advancing AI event, AMD introduced the Instinct MI350 series, made up of two parts: MI350X and MI355X. Each carries 288 gigabytes of HBM3E memory and moves data at 8 terabytes per second, about one and a half times the previous MI300X generation (which sat at 5.3). In practice: very large models fit entirely in the memory of a single card, without having to split the work across several GPUs.

On compute, AMD's stated figures are 4.6 petaflops FP8 for the MI350X and 5.0 for the MI355X, doubling in FP4 to 9.2 and 10.1 respectively. FP8 and FP4 are low-precision number formats: fewer digits per number, far more calculations per second. The stated comparison is not against the older H200 but against NVIDIA's Blackwell generation, B200 and GB200.

The piece missing on announcement day is the software: ROCm 7, the version of AMD's stack that brings the developer experience closer to CUDA, only arrived between August and September 2025. And it is almost always the software, not the board, that decides whether a customer switches vendor.