AMD Instinct MI325X

The MI325X reuses the MI300X's CDNA 3 compute dies but swaps in eight stacks of 12-Hi HBM3E for 256GB total capacity and 6.0 TB/s of bandwidth, both notable jumps over the MI300X's 192GB and 5.3 TB/s. It's positioned as AMD's direct answer to NVIDIA's H200, released to keep pace on memory capacity through 2024 and into 2025 deployments.

We'd recommend the MI325X specifically for teams serving the largest open-weight models (405B-class and up) where fitting more of the model in HBM per GPU reduces the multi-GPU communication overhead that eats into inference throughput. Compute throughput is essentially identical to the MI300X since it's the same CDNA 3 dies, the entire uplift here is memory capacity and bandwidth, so if your workload isn't memory-bound, the extra cost over MI300X buys you less than it might suggest.

Against the NVIDIA H200 (141GB HBM3E, ~4.8 TB/s), the MI325X leads comfortably on both capacity (256GB vs 141GB) and bandwidth (6.0 TB/s vs ~4.8 TB/s), making it AMD's strongest capacity argument yet against NVIDIA's data center lineup. Against its own sibling the MI300X, it's a straightforward upgrade in memory only, same 304 CUs, same CDNA 3 architecture, just more and faster HBM at a higher power draw. Against NVIDIA's Blackwell-generation B200, the picture is closer, NVIDIA's newer architecture claims a compute throughput lead even though the MI325X still holds a per-GPU memory capacity edge. As with all Instinct cards, expect enterprise quote pricing rather than a published retail price.