AMD Instinct MI350P First Look: 144GB HBM3E in a 600W PCIe Card

AMD’s Instinct MI350P brings CDNA 4 to a dual-slot PCIe card, combining 144GB of HBM3E, 4TB/s bandwidth, and up to 4.6 PFLOPS of dense MXFP4 performance. It is effectively half an MI350X, with four compute dies and four memory stacks, and can be configured for 450W or 600W; the tested card ran at 450W. Its memory capacity is the largest among the compared PCIe accelerators, exceeding the H200 NVL by 3GB and the RTX PRO 6000 Blackwell by 48GB. That capacity can let larger models run on one card, while the RTX PRO 6000 offers broader graphics and media capabilities. AMD also demonstrated a four-card workstation concept with 576GB of combined HBM3E, planned for 2027.

In measured compute tests, the MI350P fell between NVIDIA’s cards: it beat the 600W RTX PRO 6000 by 36-38% across tested precisions, but trailed the H200 NVL in BF16 and FP8; the H200 lacks native FP4. Its strongest language-model result was GPT-OSS-120B, reaching 5,057 output tokens per second at 128 concurrent streams, ahead of both NVIDIA cards. For short prompts, it generally outperformed the RTX PRO 6000 by 35-65%; long-prompt results were mixed, with early performance flattening on some models. Ten of 14 model-and-precision combinations ran on one card. Some two-card results used an older ROCm stack for comparable testing, and full 600W performance remains untested.