AMD Instinct MI350 Series: The Complete Guide

AMD’s own framing at launch claimed “up to a 4x generation-on-generation AI compute increase” and a “35x generational leap in inferencing” over MI300X 1. A separate AMD blog states a more specific claim: on Llama 3.1 405B inference in FP4, MI355X “delivers up to 40% better tokens-per-dollar than B200”, naming NVIDIA’s B200 directly rather than an unnamed “competing solution” 3. All of these are AMD’s own comparisons, not independently benchmarked.

AMD’s own MI350P blog states its purpose plainly: “dual-slot drop-in cards for standard air-cooled servers”, built “to run enterprise AI on your existing infrastructure” 2 — a deliberate trade-off of raw performance for compatibility with data centres that were never built for 1,000-watt-plus liquid-cooled chips.

AMD’s ROCm documentation states plainly: “The AMD Instinct MI350 Series GPUs are based on the AMD CDNA4 architecture” 4. AMD’s MI355X product page lists a transistor count of 185 billion — a figure that appears on the live product page but not in either PDF brochure this page also draws on, so it is reported here with that caveat 5.

MI355X figures per AMD’s own product brief, which gives a “peak theoretical performance with sparsity” table 6; AMD’s consumer-facing product page rounds these to MXFP4/MXFP6 10.1 PFLOPS, MXFP8 5 PFLOPS, FP16 matrix 2.5 PFLOPS (5 with sparsity) 5. MI350X’s FP8 figures are not reported here: this page’s own reading of the official MI350X brief returned an apparent extraction error on that row, so rather than publish a possibly wrong number, it is left blank pending a direct re-check of the PDF 8. MI350P figures per AMD’s own product brochure 7.

AMD’s own ROCm documentation gives the exact wording for MI350X/MI355X: the compute chiplets are “built on TSMC’s N3P process to take advantage of improved logic density and performance” and the I/O chiplets are “implemented on TSMC’s N6 process” 4. AMD names Samsung as “the primary HBM3E partner” for MI350X and MI355X memory, and separately says the two companies “will align on primary HBM4 supply” for AMD’s next chip, MI455X 9. AMD’s June 2025 launch material also names Micron as an HBM3E supplier at launch, alongside Samsung.

That is partner and ecosystem naming rather than a confirmed list of cloud providers renting MI350-series chips by the hour. Some specialist cloud providers have since listed MI355X for rent, but this page could not directly verify a current, official price from any of them at the time of writing, so none is stated in the pricing section below.

This is a genuine gap rather than an oversight: unlike MI300X and MI325X, which multiple named cloud providers list with a current price on their own pricing pages, this page found no cloud provider’s own official pricing page confirming a current MI350-series rate. A figure for MI355X circulates on at least one specialist GPU-cloud site, but this page could not directly verify it against that provider’s own current pricing page, so it is not repeated here.

AMD’s own product brief and product page both show FP8 dense at 5.0 PFLOPS rising to 10.1 PFLOPS “with sparsity”, and FP16 matrix at 2.5 PFLOPS rising to 5 PFLOPS with sparsity — the expected roughly-doubling pattern. MXFP4 breaks that pattern: it is listed as a single, flat 10.1 PFLOPS figure with no separate dense/sparse split shown 56. A reader comparing MI355X to a rival chip on MXFP4 alone should check carefully whether the other chip’s figure is dense or sparse before comparing it to AMD’s single MXFP4 number.

AMD’s own June 2024 roadmap release first previewed the architecture that became MI350X/MI355X as part of a multi-year plan that already named MI400-series chips as the following generation 11 — a plan AMD then delivered on, launching MI350X/MI355X in June 2025 1 and MI455X in production in July 2026 12.

AMD’s own description of MI350P’s purpose — “run enterprise AI on your existing infrastructure” 2 — is a direct acknowledgement that MI350X and MI355X, at 1,000W and 1,400W in an OAM module, are simply not an option for a large share of existing enterprise data centres. MI350P’s 600W (configurable down to 450W) PCIe card, at roughly half the memory and compute of its OAM siblings, is AMD’s answer for that market rather than a smaller version of the same product aimed at the same buyer 7.