AMD Instinct MI300A
Modular APU combining CPU and GPU tiles for high-performance computing.
The AMD Instinct MI300A is a large accelerated processing unit that integrates CPU and GPU components into a single 3D stacked package. Designed for supercomputers like El Capitan, it uses an SH5 socket and onboard HBM memory instead of traditional DIMMs.
- Combines CPU and GPU tiles in a 3D stacked package
- Uses onboard HBM memory instead of DDR5 DIMMs
- Fits in an SH5 LGA6096 socket similar to EPYC
- Powers the El Capitan 2 Exaflop Supercomputer
- Configured with one-quarter CPU and three-quarters GPU
What it is
The AMD Instinct MI300A is a high-end accelerated processing unit designed for data centres and supercomputing. It combines central processing units and graphics processing units into one module using 3D stacking technology. This approach allows AMD to vary the ratio of compute tiles, with the MI300A featuring one part CPU and three parts GPU. The chip serves as a key component in systems requiring massive parallel processing power, such as the El Capitan supercomputer.
How it came about
AMD revealed details about the MI300 family at the SC23 conference and earlier events in June 2023. The company positioned the chip as a modular building block for next-generation high-performance computing. Observers noted its presence in Gigabyte server hardware during industry showcases, highlighting its role in upcoming deployments by major technology firms and national laboratories.
Inside the chip
The processor uses a modular architecture where GPU, CPU, and high bandwidth memory are stacked together. Unlike standard servers, it does not use DDR5 memory slots. Instead, it relies on onboard HBM for all memory needs. The chip connects via multiple PCIe Gen5 interfaces and fits into an SH5 socket, which shares the LGA6096 form factor with AMD’s EPYC server processors. Support for memory expansion via CXL devices remains unconfirmed by the manufacturer.
How it was made
The MI300A is built using a 3D stacked package design. It utilizes an SH5 carrier that resembles those used for high-end EPYC chips, though in a grey colour rather than orange or green. The physical interface is an LGA6096 socket, allowing it to be installed in motherboards designed for similar high-core-count server processors. Power delivery involves integrated stages typical of modern point-of-load converters.
Where it went
This accelerator targets high-performance computing environments. It is a core component of the El Capitan supercomputer, which aims for two exaflops of performance. The chip is also deployed in large numbers by cloud providers like Microsoft for AI workloads. Server manufacturers such as Gigabyte have integrated the MI300A into their rack-mounted systems, utilizing its socketed design for easier maintenance and upgrades.
Versions and family
The MI300 family includes the MI300X, which is configured entirely with GPU tiles and no CPU cores. The MI300A balances both types of compute resources. Both variants share the same underlying modular building block architecture but serve different workload requirements. The MI300X is often cited for pure AI inference tasks, while the A model handles mixed computational duties.
Why it matters
The introduction of the MI300A marks a shift towards integrated accelerator designs in the supercomputing sector. By combining CPU and GPU resources in a single package, AMD offers a solution for systems that require tight coupling between general-purpose and parallel processing. Its adoption in major supercomputers and cloud infrastructure highlights its significance in the current AI hardware market.
Problems and reception
Industry observers at SC23 noted the physical similarities between the MI300A socket and existing EPYC platforms. Interest focused on its memory architecture and potential support for CXL expansion. The lack of traditional DIMM slots was seen as a distinct design choice aimed at maximizing bandwidth through onboard HBM. Photos from the event confirmed its integration into standard server form factors.
Questions people ask about AMD Instinct MI300A
What is the difference between MI300A and MI300X?
The MI300A contains both CPU and GPU tiles, with a ratio of one-quarter CPU to three-quarters GPU. The MI300X is configured entirely with GPU tiles and has no CPU cores, making it suitable for pure accelerator workloads.
Does the MI300A use standard server memory?
No, the chip does not have DIMM slots for DDR5 memory. It uses onboard high bandwidth memory (HBM) for its memory needs, which provides higher capacity and bandwidth compared to traditional configurations.
What socket does the MI300A use?
The processor uses an SH5 socket with an LGA6096 interface. This design is very similar to the SP5 socket used by AMD EPYC Genoa and Bergamo server CPUs, allowing for familiar installation methods.
Good reading and sources
There is no English Wikipedia article on AMD Instinct MI300A yet, so the facts on this page come from the sources listed above: servethehome.com/giagantic-amd-apu-at-sc23-meet-the-amd-mi300a-and-mi3, fudzilla.com/amd-unveils-instinct-mi300x-and-mi300a-apu-accelerators/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.