AMD Instinct MI300X.

A high-memory AMD accelerator for production inference. Start with one isolated GPU VM or scale to an eight-GPU server without changing the operational model.

Start now View benchmarks

At a glance

Built for memory-intensive serving.

MI300X pairs a large HBM3 footprint with the bandwidth needed to serve open-source models reliably, without turning deployment into a hardware project for your team.

HBM3 memory

192 GB

Memory bandwidth

5.3 TB/s

Compute units

304

Accelerators per server

Inference fit

Capacity that does not get in the way.

The accelerator is only part of the offering. Hot Aisle automates the networking, PXE, operating system, ROCm, virtualization, and billing layers required to make this capacity usable.

Memory for the model you actually want to run

192 GB of local HBM3 gives serving teams room for larger models, longer context, and higher-concurrency deployments without treating memory as an afterthought.

Bandwidth for sustained generation

5.3 TB/s of bandwidth keeps the accelerator supplied as requests move through the model, supporting responsive token generation at production scale.

Isolation that matches the workload

Provision a NUMA-balanced KVM virtual machine for a single GPU, or take a complete eight-GPU server when the workload calls for it.

System profile

The accelerator, in a complete platform.

Infrastructure is more useful when the physical system, virtualization model, and control surface are designed together.

CDNA 3 architecture

CDNA 3 combines eight XCDs, 304 compute units, and 256 MB of AMD Infinity Cache in a dense accelerator designed for model serving.

Eight-GPU system density

Eight MI300X accelerators are available in each Dell PowerEdge XE9680, with a high-bandwidth fabric and full-system automation already in place.

Deploy MI300X

Ready when the workload is.

Create a team, add credit, and provision isolated AMD GPU compute from the terminal UI, API, or CLI. No sales handoff is required.

Quick start