AMD announced the MI350P, a new PCIe AI accelerator card with 144GB of HBM3E memory, claiming roughly 40% faster theoretical compute in FP16 and FP8 modes compared to NVIDIA's H200 NVL competitor. The move positions AMD to compete directly in the gigascale AI inference and training market as enterprises evaluate alternative accelerators to reduce GPU supply bottlenecks.
The MI350P sits in AMD's MI300 lineup and targets deployment in standard PCIe infrastructure—a lower integration barrier than NVIDIA's NVLink ecosystem for datacenters not yet all-in on Hopper-based stacks.