AMD Instinct MI300X Specs & Datasheet (8-GPU Server)

The AMD Instinct MI300X is AMD’s CDNA 3 data-center accelerator, and the part that established the Instinct line as a credible alternative for memory-bound AI work. Each accelerator carries 192 GB of HBM3 at roughly 5.3 TB/s of bandwidth.

Like NVIDIA’s HGX platforms, the MI300X is sold as a node, not a card. Eight accelerators sit on an OAM/UBB baseboard connected by AMD Infinity Fabric, giving 1,536 GB of pooled HBM3 in one chassis. Its successors keep the same shape: the MI325X adds memory and bandwidth on the identical CDNA 3 compute core, and the CDNA 4 MI355X is the current flagship.

What keeps the MI300X relevant is deployability. At 750 W per accelerator it runs air-cooled in a standard rack — no cold plates, no CDU, no building water loop. For a buyer who needs large pooled memory but does not have a liquid-cooled facility, that is frequently the deciding factor.

The MI300X per-GPU and 8-GPU node specification. The 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.

Each MI300X carries 192 GB of HBM3 at about 5.3 TB/s, so an 8-GPU node pools 1,536 GB of high-bandwidth memory addressable over Infinity Fabric.

That capacity is the reason to look at the part. A single MI300X holds a model that would otherwise have to be sharded across several 80 GB GPUs — and sharding costs both latency and engineering time. For memory-bound work — large models held resident, long-context inference with a big KV cache — capacity per GPU is frequently the binding constraint rather than raw FLOPS.

Against its own line, the MI300X gives up memory to its successors: the MI325X carries 256 GB at 6.0 TB/s and the MI355X 288 GB at 8 TB/s. What the MI300X gives back is a lower power envelope and the widest air-cooled availability.

Each MI300X delivers roughly ~1,307.4 TFLOPS (~1.31 PFLOPS) of dense FP16/BF16 and ~2,614.9 TFLOPS (~2.61 PFLOPS) of dense FP8. Across an 8-GPU node that is about ~10.5 PFLOPS dense FP16 and ~20.9 PFLOPS dense FP8.

Those figures are identical to the MI325X — the two parts share the same CDNA 3 compute core, and the MI325X’s generational gain is memory and bandwidth, not FLOPS. CDNA 3 has no FP4 datatype; that arrives with CDNA 4 on the MI355X, which roughly doubles dense FP16 and FP8 as well. All numbers here are dense; AMD’s datasheet headline figures double them with 2:4 sparsity.

Inside the node, the eight accelerators communicate over AMD Infinity Fabric at 896 GB/s. This is the on-baseboard fabric that lets the eight GPUs share their 1,536 GB of pooled HBM3 as one node, rather than eight isolated cards.

For scale-out beyond the node, shipping builds pair each GPU with a 400 Gb/s InfiniBand or RoCE NIC. The practical consequence: the MI300X is deployed as a node-scale unit — where 1,536 GB of pooled memory in one chassis does the work — and there is no AMD equivalent to a rack-scale NVLink domain like the GB200 NVL72. Exact NIC count and fabric tier are specified per build.

Each MI300X has a 750 W TDP, so an 8-GPU baseboard draws on the order of 6 kW from the accelerators alone, before host CPUs, system memory, and conversion losses.

That is the MI300X’s quiet advantage. Six kilowatts of accelerator load is a load a conventional air-cooled rack can carry, which is why both shipping integrations are air-cooled: Supermicro’s 8U AS-8125GS-TNMR2 and Dell’s 6U PowerEdge XE9680. Neither needs facility water or a CDU.

By contrast the CDNA 4 MI355X runs at 1,400 W per accelerator — roughly 11.2 kW per baseboard — and is a direct-to-chip liquid build that will not deploy without a building water loop. If a site cannot take liquid, that difference decides the part. Exact per-node draw and facility fit depend on the configuration and are confirmed at quote.

Where the MI300X sits in the AMD Instinct line. The MI325X shares its CDNA 3 compute core and adds memory and bandwidth; the MI355X is the CDNA 4 successor.

The MI300X 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 MI300X 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.
• Integrator and chassis. The 8U Supermicro build versus the 6U Dell PowerEdge XE9680 — rack units, serviceability, and support model differ.
• Host CPU and memory. The host CPU tier (EPYC, or Xeon on the Dell) and the system DDR5 capacity feeding the accelerators.
• Scale-out networking. How many 400 Gb/s InfiniBand or RoCE NICs per node, 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.