SLYD
Read first

The 5-step pipeline.

How energy turns into a deployed, financed, offtake-matched cluster, and what SLYD does at every step.

How SLYD works →
Hardware

New and recovered GPU systems.

NVIDIA and AMD systems through documented manufacturer and qualified channel supply, with financing and deployment coordinated on the same platform.

Explore GPU hardware →
Marketplace

Compute, hardware, and power in one book.

Browse available GPU capacity by accelerator, configuration, region, and price, or bring supply to qualified demand.

Open marketplace →
Pre-qualify

Start with an indicative structure.

Tell us deal size, structure, and offtake. Any range is preliminary and subject to underwriting, diligence, and documentation.

Open Configure →
From the blog

GPU market trends and deployment playbooks.

Infrastructure best practices, hardware comparisons, and industry analysis from the SLYD team.

Read the blog →

Library

AMD accelerators

AMD Instinct MI300X, MI325X and MI355X systems

The Instinct decision has two halves and the software half comes first. This page covers AMD's published specifications by generation, then the ROCm compatibility work that determines whether the hardware is usable for a given stack.

What are AMD Instinct systems?

AMD Instinct accelerators support data-center AI and HPC workloads through OEM and platform-specific systems. Buyers should compare the exact accelerator generation, memory, precision support, interconnect, cooling configuration, ROCm software compatibility, server design, workload, and verified commercial terms.

Manufacturer specifications

Instinct generations as AMD publishes them

These figures come from AMD's product specification pages for each accelerator. The performance rows are peak matrix performance at the stated precision and are dense unless the row says otherwise.

Source: AMD Instinct MI300X, MI325X, MI350X, and MI355X product specification pages, reviewed 18 August 2026. Peak performance figures are AMD's published peak matrix performance at the stated precision.
Specification MI300X MI325X MI350X MI355X
Dedicated memory 192 GB 256 GB 288 GB 288 GB
Memory type HBM3 HBM3E HBM3E HBM3E
Peak memory bandwidth 5.3 TB/s 6 TB/s 8 TB/s 8 TB/s
Peak MXFP4 matrix Not published for this generation Not published for this generation 9.2 PFLOPs 10.1 PFLOPs
Peak FP16 matrix (dense) Not published in this table Not published in this table 2.3 PFLOPs 2.5 PFLOPs
Typical board power 750W peak 1000W peak 1000W 1400W
Form factor OAM module OAM module OAM module OAM module
Bus type PCIe 5.0 x16 PCIe 5.0 x16 PCIe 5.0 x16 PCIe 5.0 x16
Cooling, as published by AMD Passive OAM Passive OAM Passive OAM Passive and Active

AMD publishes MI355X on the 4th Gen AMD CDNA architecture with expanded MXFP6 and MXFP4 datatype support, and publishes 10.1 PFLOPs peak MXFP6 and 5 PFLOPs peak MXFP8 matrix performance alongside the MXFP4 figure above.

Cooling

Cooling is a platform question, not a family rule

It is often said that Instinct accelerators require liquid cooling. AMD's own specifications do not say that. AMD publishes cooling as Passive OAM for MI300X, MI325X, and MI350X, and as Passive and Active for MI355X.

What a real deployment needs comes from the OEM platform and its board power, which ranges from 750W peak on MI300X to 1400W on MI355X. A 1400W accelerator in a dense 8-way platform will usually be delivered as a direct-liquid-cooled system, but that is a property of the server, not a property of the accelerator family. Confirm the cooling design of the exact OEM system being quoted, and check it against what the facility can support.

Software

Validate ROCm before validating the hardware

Instinct runs the ROCm software stack rather than CUDA. For most organizations this is the decision, and it is answered by testing rather than by reading a table.

  1. List the stack that actually has to run

    Frameworks and versions, inference server, quantization and kernel libraries, distributed training libraries, schedulers, monitoring, and anything with a custom kernel. Custom kernels are where portability effort usually concentrates.

  2. Check support against the target ROCm version

    Support is specific to a ROCm release and a platform, not to ROCm in general. Confirm which ROCm version the OEM platform is validated for and whether the stack is supported on that version.

  3. Run the real workload before committing

    Throughput, latency, and accuracy on the actual models, at the actual precision and batch size. Peak specification comparisons do not predict this reliably in either direction.

  4. Cost the migration honestly

    Engineering time to port and validate, ongoing maintenance across two stacks if the estate becomes mixed, and the operational cost of a second toolchain. That work is part of the comparison, not outside it.

Comparison method

How to compare Instinct against NVIDIA defensibly

SLYD does not publish a head-to-head performance verdict between these families. Both vendors publish peak theoretical figures on different datatypes and different sparsity conventions, and those tables cannot be lined up against each other without producing a misleading result.

A comparison that survives scrutiny fixes and states all of the following before any number is quoted:

  • Exact accelerator model and platform on both sides
  • Model, precision, and quantization scheme
  • Whether figures assume sparsity or are dense
  • Batch size, sequence length, and concurrency
  • Software versions on both sides
  • Node and cluster configuration
  • Whether the measure is throughput, latency, or cost per unit of work
  • Who ran the test and when

AMD publishes its own comparative claims as calculations by AMD Performance Labs with named endnotes. Those are AMD's claims about peak theoretical performance under AMD's stated conditions, and they are not reproduced here as SLYD findings. For side-by-side specifications, use the GPU database.

Manufacturer systems

Where Instinct accelerators appear

Systems published by the manufacturers in their own current documentation, reviewed 18 August 2026. Named to show where the accelerators appear, not to indicate availability.

Manufacturer-published AMD Instinct implementations. Confirm the current configuration with the manufacturer before relying on any detail.
Manufacturer Published system As described by the manufacturer
Dell PowerEdge XE9785 1OU, 2 x AMD EPYC 9005 processors, 8 x AMD Instinct MI355X
Dell PowerEdge XE9680 6U, published with 8 x AMD Instinct MI300X among its accelerator options
HPE ProLiant Compute XD685 8-way GPU server with AMD Instinct MI355X as one of its accelerator choices, 5U for direct liquid cooling and 6U for air
Lenovo ThinkSystem SR685a V3 Published with AMD Instinct MI300X accelerators
Buyer diligence

Confirm before ordering

  • Exact accelerator generation and count per node
  • OEM platform and its published cooling design
  • Rack power and thermal capacity for that configuration
  • ROCm version the platform is validated for
  • Validated results for the actual software stack
  • CPU, memory, storage, and network configuration
  • Equipment condition and provenance
  • Warranty provider, scope, term, and transferability
  • Availability confirmation with its date and expiration
  • Export, end-use, and geographic restrictions
FAQ

AMD Instinct questions

What is the difference between MI300X, MI325X, and MI355X?

Memory and datatype support, mainly. AMD publishes MI300X with 192 GB of HBM3 at 5.3 TB/s, MI325X with 256 GB of HBM3E at 6 TB/s, and MI355X with 288 GB of HBM3E at 8 TB/s. MI355X is built on the 4th Gen AMD CDNA architecture with expanded MXFP6 and MXFP4 datatype support, which the earlier generations do not have.

Do AMD Instinct accelerators require liquid cooling?

No, not as a family requirement. AMD publishes cooling as Passive and Active for MI355X, and as Passive OAM for MI350X, MI325X, and MI300X. What a specific deployment needs is determined by the OEM platform and its board power, which AMD publishes as 750W peak for MI300X, 1000W peak for MI325X, 1000W for MI350X, and 1400W for MI355X.

What is the peak FP4 performance of MI355X?

AMD publishes 10.1 PFLOPs peak Microscaling four-bit precision matrix performance, written as MXFP4, for MI355X, and 9.2 PFLOPs for MI350X. AMD also publishes 10.1 PFLOPs peak MXFP6 and 5 PFLOPs peak MXFP8 for MI355X. Higher figures circulating for this part do not match AMD's current product specification page.

Will my AI software run on AMD Instinct?

That is the question to resolve before the hardware question. AMD Instinct runs the ROCm software stack rather than CUDA. Major frameworks and popular inference servers support ROCm, but coverage varies by framework version, model, kernel, and library, so the specific stack in use should be validated against the ROCm version supported on the target platform.

How should AMD Instinct be compared against NVIDIA accelerators?

On the actual workload, not on peak theoretical numbers. The two vendors publish different datatypes, different sparsity conventions, and different platform boundaries, so specification tables are not directly comparable. A defensible comparison fixes the model, precision, batch size, software version, and system configuration, and states all of them.

Which manufacturers build AMD Instinct systems?

Server manufacturers publish their own Instinct implementations. Dell publishes PowerEdge XE9785 with 8 AMD Instinct MI355X and PowerEdge XE9680 with 8 AMD Instinct MI300X, HPE publishes ProLiant Compute XD685 with MI355X as one of its accelerator options, and Lenovo publishes ThinkSystem SR685a V3 with MI300X accelerators.

What should I confirm before ordering an Instinct system?

The exact accelerator generation and count, the OEM platform and its cooling design, rack power and thermal capacity for that configuration, the ROCm version and validated software stack, network and storage configuration, and the condition, warranty, availability, and commercial terms for the specific transaction.

Sources for the specifications on this page

Specifications belong to the manufacturers named and are reproduced as published. They are not SLYD guarantees of performance, configuration, or availability.

Scope an AMD Instinct requirement

Share the workload, the software stack that has to run, node count, and facility constraints. The software validation usually drives the schedule more than the hardware does.

Page updated: August 18, 2026

Reconnecting to the server...

Please wait while we restore your connection

An unhandled error has occurred. Reload 🗙