AMD Instinct MI325X Specs & Datasheet (8-GPU Platform)

The AMD Instinct MI325X is an AMD data-center accelerator built on the CDNA 3 architecture. It is the memory-forward member of the Instinct line: 256 GB of HBM3e per GPU at roughly 6 TB/s of bandwidth — the largest single-GPU memory capacity in its class.

Like NVIDIA’s HGX platforms, the MI325X is sold as a platform, not a card. Eight accelerators sit on an OAM baseboard connected by AMD Infinity Fabric, giving 2,048 GB (~2 TB) of pooled HBM3e in one chassis. The MI325X shares the same CDNA 3 compute core as the MI300X; what it adds is more memory (256 GB vs 192 GB) and more bandwidth (6 TB/s vs 5.3 TB/s). The CDNA 4 MI355X is its successor.

The MI325X per-GPU and 8-GPU platform 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.

Memory is the MI325X’s defining characteristic. Each accelerator carries 256 GB of HBM3e at about 6 TB/s — so an 8-GPU platform pools 2,048 GB (~2 TB) of high-bandwidth memory addressable over Infinity Fabric.

That capacity is the reason to look at the part. It is the largest single-GPU memory in its class, and 33% more per GPU than the 192 GB MI300X on the identical CDNA 3 compute core — with bandwidth stepping up from 5.3 TB/s to 6.0 TB/s. For memory-bound workloads — large models held resident, long-context inference with a big KV cache — capacity per GPU is frequently the binding constraint, and this is where the MI325X is most competitive: fewer GPUs to hold a given model, and more headroom per GPU.

On compute, each MI325X 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 platform that is about ~10.5 PFLOPS dense FP16 and ~20.9 PFLOPS dense FP8.

Those are the same CDNA 3 throughput figures as the MI300X — the MI325X’s generational gain is memory and bandwidth, not raw FLOPS. All numbers here are dense; AMD’s datasheet headline figures double these with 2:4 sparsity.

Inside the platform, 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 2,048 GB of pooled HBM3e as one platform, rather than eight isolated cards.

The practical consequence: the MI325X is deployed as a platform-scale unit — where ~2 TB of pooled memory in one chassis does the work of a memory-bound job without sharding a model across separate hosts. Scale-out networking beyond the platform (InfiniBand or RoCE) is specified per build and confirmed at quote.

Each MI325X has a 1,000 W TDP. An 8-GPU baseboard therefore draws on the order of 8 kW from the accelerators alone, before host CPUs, system memory, and conversion losses.

That density is what drives the cooling and facility design of a platform. Chassis, cooling method, and exact per-node site power draw depend on the integrator build and are confirmed at quote.

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

The MI325X ships as a factory-integrated 8-GPU platform. The accelerator, memory, and fabric specifications are fixed; chassis, host CPU, cooling, warranty, and lead time depend on the build.

An 8-GPU MI325X platform is quoted per configuration on request — there is no meaningful list price, because two platforms 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 1,000 W per-GPU TDP drives the cooling and facility design of the build.
• Host CPU and memory. The host CPU tier and system DDR5 capacity feeding the accelerators.
• Scale-out networking. How much InfiniBand or RoCE fabric per GPU beyond the platform, and the switching to match.
• Quantity and support scope. Single platform 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.