AMD Instinct MI355X: the CDNA 4 accelerator, explained

What the MI355X is

The MI355X is AMD's answer to Blackwell, and its pitch is memory. The CDNA 4 flagship, launched mid-2025 at the top of the Instinct MI350 series, is a 3nm chiplet design with 256 compute units and 288 GB of HBM3e per GPU across ten stacks at roughly 8 TB/s. That matches the B300 on memory and beats the 192 GB on the B200 and the H100 generation outright: more model, more context, and larger KV caches resident on a single accelerator.

It ships in the OAM (OCP Accelerator Module) form factor and drops into industry-standard UBB 2.0 baseboards, so an MI355X node is an 8-GPU server much like an HGX node in shape and deployment. The GPUs connect over 4th-generation Infinity Fabric in a fully meshed topology, and the software stack is ROCm, AMD's open CUDA alternative, which by 2026 has matured into a serious production runtime for inference. The MI355X does not try to beat NVIDIA on the widest software moat; it competes on memory per dollar, an open stack, and a real second source of supply.

The MI355X is aimed squarely at the inference era, and its case rests on memory. At serving time the KV cache, the per-conversation working memory of a large language model, competes with model weights for HBM. With 288 GB per GPU the MI355X holds larger models, longer contexts, and more concurrent sessions before you shard across accelerators or evict cache. For memory-bound workloads, reasoning models with long chains of thought, agents carrying large contexts, and high-concurrency APIs, capacity per GPU often matters more than peak FLOPS, and this is exactly where the MI355X is strongest. It also supports FP4 and FP6, so quantized inference runs at the narrow precisions serving stacks are converging on.

The software question used to be the deciding one, and in 2026 it is much less so. ROCm now provides day-one or near day-one support for the mainstream open inference engines, vLLM and SGLang among them, so most modern serving stacks run on MI355X without a rewrite. The commercial argument follows: a credible second source disciplines pricing and shortens lead times, AMD hardware typically prices below the equivalent NVIDIA part, and for a workload dominated by memory footprint the total cost of ownership can land in AMD's favor. The honest caveat: validate your specific model and kernels on ROCm first, because coverage is broad but not yet universal.

An 8-GPU MI355X platform pools eight accelerators into 2.3 TB of Infinity Fabric-connected HBM3e (8 x 288 GB), the same memory class as an NVIDIA B300 node. Today's frontier open models are mixture-of-experts, so every expert must sit in HBM even though only a slice fires per token: total parameters, not the active count, set the memory floor. Weights-only arithmetic (1 byte per parameter at FP8, half that at FP4) shows what fits where:

Weights-only figures at published parameter counts. Production serving adds KV cache, activations and framework overhead, which is why the memory pool is the binding constraint. The point holds even for 1T-plus models: the 2.3 TB pool keeps a trillion-parameter MoE resident on a single node, no cross-node sharding required.

MI355X specs

Official AMD CDNA 4 figures; exact clocks vary by OEM and cooling. Interconnect is 4th-gen Infinity Fabric (fully meshed) with a PCIe Gen 5 host link. Real inference throughput and cost per token depend on the model and serving stack, not raw FLOPS alone.

MI355X vs NVIDIA B200

The clearest contrast is memory. The MI355X carries 50% more HBM per GPU than the B200 at comparable bandwidth and FP4 throughput, which favors memory-bound serving. The B200 answers with CUDA and the broadest tooling and support. For a workload gated by memory footprint the MI355X is compelling; for one gated by ecosystem depth the B200 is the safer default.

MI355X vs NVIDIA B300

Here the two match on memory at 288 GB per GPU, so the comparison comes down to compute, ecosystem and cost. The B300 leads on peak FP4 throughput and rack-scale NVLink domains; the MI355X answers with the same memory class as a genuine second source, an open stack, and typically a lower acquisition cost per GB. For memory-bound serving where peak FLOPS are not the bottleneck, the MI355X can deliver B300-class capacity at a better price.

What an MI355X costs

AMD wins on price, and that is the point. Instinct capacity generally lands below the equivalent NVIDIA part, so for a memory-bound workload you get B300-class capacity per GPU at a lower rate, and a viable second source keeps the whole market honest. The pattern underneath is familiar: while supply is tight, on-demand carries a premium, spot appears only in windows, and reserved terms land well below on-demand, which is where serious volume transacts.

The MI355X does not yet have its own dedicated live index, so for AMD levels see the live MI300X price index, the closest tracked AMD Instinct part, as a proxy for how AMD capacity prices against NVIDIA. The full GPU price index covers the rest of the fleet if you want to compare AMD and NVIDIA economics side by side.

MI355X availability in Europe

Era Compute sources MI355X capacity in European datacenters with the power and liquid cooling that CDNA 4 density demands, from a single 8-GPU node up to larger clusters, with per-GPU pricing that falls as the cluster grows. This is reserved capacity: contracted on multi-year terms and qualified per customer, not on-demand rental.

For European buyers the placement is the point. EU-sited capacity keeps training data and inference traffic inside GDPR and data-residency boundaries, and European facilities frequently undercut equivalent US capacity.

Questions, answered