Skip go content

AMD Instinct MI300X

Datacenter GPU acceleratorAMD

Dem still dey call am amd instinct mi300x, instinct mi300x

AMDDatacenter GPU accelerator

AMD Instinct MI300X na CDNA 3 datacenter compute accelerator with up to 192 GB of HBM3 memory and 304 compute units inside OAM module.

Wetin you need know

AMD Instinct MI300X na CDNA 3 datacenter compute accelerator with up to 192 GB of HBM3 memory and 304 compute units inside OAM module.

Reference specifications wey dem record no be tested performance or operating-system compatibility verdict. Verify di exact model and platform wit di manufacturer.

Model identity and limitations

AMD Instinct MI300X target GPU compute inside supported datacenter systems. AMD document MI300X systems wey dey use eight-accelerator UBB 2.0 baseboard. Di 192 GB figure na per accelerator; server totals depend on installed modules. Specify weda capacity or throughput figure dey describe one accelerator or di complete server.

Specifications

Manufacturer specification reference

Selected model-specific fields wey dem check against AMD on October 29, 2026. Bandwidth na manufacturer peak specification, no be measured application performance. Power labels follow di manufacturer and dem no dey interchangeable with complete-system consumption.

Manufacturer specification reference
GPU architecture CDNA 3
Module form factor OAM (OCP Accelerator Module)
Compute units 304
Stream processors 19,456
Memory capacity and type Up to 192 GB HBM3 per accelerator
Memory bandwidth Up to 5.3 TB/s, peak theoretical
Maximum board power (TBP) 750 W

Check di complete server platform

Confirm di server model, module identity, firmware and cooling configuration with di system vendor before you buy parts or change software. OAM module need compatible accelerator baseboard and server integration; host-interface specification alone no dey establish physical compatibility.

AMD document MI300X systems wey dey use eight-accelerator UBB 2.0 baseboard. Di 192 GB figure na per accelerator; server totals depend on installed modules. Specify weda capacity or throughput figure dey describe one accelerator or di complete server.

Troubleshoot with di supported software stack

Record di server firmware, operating system, ROCm version, reported device identities and exact error before you make changes. Compare dat combination with AMD and di server vendor support documentation. Separate reported partitions or compute dies from physical accelerator modules.

Run di vendor documented acceptance checks inside maintenance window. Save di previous supported driver/runtime configuration and recovery procedure before you upgrade. If new runtime cause regression, restore di previously supported stack using di system vendor documented method; avoid to delete kernel or driver files by hand.

Sources and verification

Imported reference. Source-specific checks dey beside di supported fields and citations. How we dey do research.

Dis article make sense to you?

Yes or no dey enough. Choose reason if you want, or add short note.

Dis article make sense to you?