AMD’s Instinct MI350P brings 144GB of HBM3E memory to a dual-slot PCIe accelerator, and a new StorageReview benchmark found it particularly strong on chat-sized AI workloads. The results make a useful case for the card, but not a blanket win: long-prompt performance was mixed, and the review unit ran below its 600W maximum.
AMD Watch analysis
What happened
The MI350P is AMD’s first new PCIe Instinct GPU in four years, according to StorageReview’s first look. Its review found that ten of fourteen model-and-precision combinations ran on a single card. The accelerator has 144GB of HBM3E, 4TB/s of memory bandwidth and a configurable board power of up to 600W; the tested card ran at 450W.
The review’s strongest result came on GPT-OSS-120B: 5,057 output tokens per second at 128 concurrent streams on its chat workload. StorageReview measured this as 65% ahead of the NVIDIA RTX PRO 6000 and 50% ahead of the H200 NVL in that test. Its full MI350P review also found the card ahead of the RTX PRO 6000 on several 512-token prompt tests, while results on 8,192-token prompts were mixed.
Why it matters
Memory capacity can decide whether a model fits on one accelerator or has to be split across several. The MI350P’s 144GB gives it more room than the 96GB RTX PRO 6000 reviewed, while its 4TB/s bandwidth sits below the H200 NVL’s 4.8TB/s. That makes the card’s appeal more specific than “more memory, therefore faster”: model, prompt length, power setting and software stack all affect the result.
The comparison also needs its test conditions attached. The MI350P ran at 450W, while the RTX PRO 6000 ran at 600W. StorageReview reports that the MI350P led the RTX PRO 6000 on its chat-sized tests, but that is not a like-for-like power comparison or a promise of the same gains across workloads.
Our read
This is a credible sign that AMD has a serious PCIe option for AI inference, especially where fitting a large model on one card matters. The GPT-OSS result is striking; the mixed long-prompt results are the reminder to match benchmarks to the work you actually plan to run. Before treating this as a buying verdict, look for full-power testing and results on your own model and serving setup. Benchmark charts are useful, but they are not mind readers.
What to watch
- How the MI350P performs at its full 600W power limit in a suitable server.
- Whether other reviewers reproduce the GPT-OSS result and the mixed long-prompt picture.
- How availability, system support and pricing compare with the NVIDIA cards.
Discussion spark: When choosing an AI inference card, would you prioritise fitting a larger model on one GPU, or favour stronger memory bandwidth and more consistent results across prompt lengths?
Sources and evidence
- AMD Instinct MI350P First Look: 144GB of HBM3E in a 600W PCIe Card – StorageReview.com (3 November 2026, 18:44 UTC)
Independent WittyWires tracker for public updates about AMD. Not affiliated with or endorsed by AMD; this is not an official account.