AMD Instinct MI355X Specs & Datasheet (8-GPU Server)

The AMD Instinct MI355X is AMD’s current flagship data-center accelerator, built on the CDNA 4 architecture. It is the memory-forward answer to NVIDIA’s Blackwell generation: 288 GB of HBM3E per GPU at roughly 8 TB/s of bandwidth.

Like NVIDIA’s HGX platform, the MI355X is sold as a node, not a card. Eight accelerators sit on an OAM/UBB baseboard connected by AMD Infinity Fabric, giving 2,304 GB of pooled HBM3E per node. It is the successor to the MI300X and MI325X, and the first AMD part to reach memory parity with NVIDIA’s Blackwell Ultra tier.

The MI355X 8-GPU node specification. The per-GPU and per-node accelerator, memory, and fabric figures are fixed characteristics of the platform; chassis, host CPU, cooling, site power draw, and final configuration vary by build and are confirmed at quote.

Memory is the MI355X’s defining characteristic. Each accelerator carries 288 GB of HBM3E at about 8 TB/s — so an 8-GPU node pools 2,304 GB of high-bandwidth memory addressable over Infinity Fabric.

That number is the reason to look at the part. An HGX B200 node holds 1,440 GB (180 GB per GPU); the MI355X holds 60% more memory per GPU and matches the 288 GB of NVIDIA’s Blackwell Ultra B300 and GB300. For memory-bound workloads — large models held resident, long-context inference with a big KV cache — capacity per node is frequently the binding constraint, and this is where AMD is most competitive.

On compute, each MI355X delivers roughly ~5 PFLOPS of dense FP8 and ~10.1 PFLOPS of dense FP4/MXFP4, with ~2.5 PFLOPS dense FP16/BF16. All figures here are dense; AMD’s datasheet headline numbers double these with 2:4 sparsity. Dense FP8 and FP4 land close to NVIDIA’s Blackwell-generation per-GPU throughput.

Inside the node, the eight accelerators communicate all-to-all over AMD Infinity Fabric on the OAM baseboard. AMD does not publish a per-GPU aggregate fabric bandwidth for the MI350 series in the directly comparable form NVIDIA publishes NVLink, so no single figure is quoted here — but the number is not the architectural difference that matters. AMD’s 8-GPU node is a well-connected node, and there is no AMD equivalent to a 72-GPU NVLink domain like the GB200 NVL72 or GB300 NVL72 rack.

For scale-out, each GPU can be paired 1:1 with a 400 Gb/s InfiniBand or RoCE NIC over the chassis’s eight low-profile PCIe 5.0 x16 slots, with RDMA. The practical consequence: the MI355X is strongest as a node-scale deployment — where 2,304 GB of pooled memory in one chassis does the work — while NVIDIA retains the advantage when a single model must be sharded coherently across dozens of GPUs in one memory domain.

The MI355X and the air-cooled MI350X are the same CDNA 4 silicon — same 288 GB of HBM3E, same 8 TB/s. What separates them is thermal budget: the MI350X runs at 1,000 W air-cooled, the MI355X at 1,400 W with direct-to-chip liquid. That is the whole reason two chassis exist, and it is usually the first thing that decides which one a site can actually take.

An 8-GPU MI355X baseboard draws up to roughly 11.2 kW at peak from the accelerators alone, before dual EPYC host CPUs (up to 500 W each), system DDR5, NVMe, and scale-out NICs. Supermicro’s 4U liquid platform feeds that with four 6,600 W Titanium supplies in a 2+2 redundant configuration; the 8U air-cooled MI350X alternative uses six 5,250 W supplies in 3+3, carrying a far lighter 8 kW GPU load in a taller chassis.

Two site gates follow, and both are worth settling before anything else:

• Facility water. Direct-to-chip cold plates need a building loop and a CDU. Without one, the MI355X is not deployable at any price — the 8U air MI350X build is the path.
• Rack power and depth. A node in this class is a high-line, multi-circuit device, and the 4U chassis is 35.25" deep at roughly 80 lb before accelerators. Both are worth confirming against the actual rack and PDU before a build is specified.

Exact per-node draw, circuit design, and facility fit depend on the configuration and are confirmed at quote.

How the CDNA 4 flagship lines up against the Blackwell generation, per GPU:

The MI355X ships as a factory-integrated 8-GPU node. The accelerator, memory, and fabric specifications are fixed; chassis, host CPU, cooling, warranty, and lead time depend on the build.

An 8-GPU MI355X server is quoted per configuration on request — there is no meaningful list price, because two nodes with the same eight accelerators can differ substantially in cost. The variables that actually move the number:

• Accelerator count. The 8-GPU baseboard is the standard unit of sale; the accelerators dominate the bill of materials.
• Cooling and chassis. The 4U direct-liquid-cooled build versus the 8U air-cooled MI350X alternative — liquid buys density but requires facility water and a CDU.
• Host CPU and memory. Dual EPYC 9005 (Turin) tier and the system DDR5 capacity feeding the accelerators.
• Scale-out networking. How much 400 Gb/s InfiniBand or RoCE fabric per GPU, and the switching to match.
• Quantity and support scope. Single node versus a multi-rack cluster, and the warranty and on-site service term.
• Lead time. What you need delivered, and when.

For the wider picture of how accelerator servers are actually quoted, see NVIDIA GPU server pricing. Tell us the build you are standing up and we will return pricing, availability, and a facility-fit review. Browse the GPU catalog to compare integrations.