AI Chips · Product page

AMD Instinct MI300A: The Complete Guide

CPU and GPU on one chip, sharing one pool of memory — the design behind El Capitan, the world’s fastest supercomputer. Simple to expert, every number links to its source.

Stuck at any point? Ask an AI about this page →
Ask an AI about this chip:CAGPX
AMD Instinct MI300A 10 sections · 2 levels 10 linked sources Checked September 2026
How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.

1.At a glance

SimpleStart here

MI300A is what AMD calls “the world’s first data center APU for HPC and AI” — a single chip that puts CPU cores, GPU cores and shared memory on one package. It launched alongside MI300X on 6 December 2023 and now runs inside El Capitan, the world’s fastest supercomputer 12.

Deep diveThe technical detail

AMD’s own framing: MI300A “features unified memory and cache resources giving customers an easily programmable GPU platform, highly performant compute, fast AI training and impressive energy efficiency” 1. AMD reports roughly 1.9x the performance-per-watt on FP32 HPC and AI workloads compared with its previous chip, MI250X 1.

2.Launch and history

SimpleStart here

On a normal chip, the CPU has its own memory and the GPU has its own separate memory, and data has to be copied between the two before the GPU can use it. MI300A removes that copying step entirely — CPU and GPU cores share one pool of memory on the same chip.

Deep diveThe technical detail

AMD’s own MI300 family overview page calls this “the power of AMD Instinct accelerators and AMD EPYC processors with shared memory”, built for “enhanced efficiency, flexibility, and programmability” 3. That design choice is unusual: most rival AI chips, including AMD’s own MI300X, keep CPU and GPU memory pools separate. AMD’s own reasoning is squarely aimed at scientific supercomputing, where a single simulation can outgrow GPU memory alone and needs to draw on a much larger shared pool without a slow copy step in between.

3.What’s inside it

SimpleStart here

24 “Zen 4” CPU cores and 228 GPU compute units, built from chiplets on one package, sharing memory and cache 4.

Deep diveThe technical detail

Per AMD’s own MI300A data sheet: three CPU chiplets carrying 24 Zen 4 x86 cores in total, six GPU chiplets carrying 228 compute units, 14,592 stream processors and 912 matrix cores, plus 256 MB of Infinity Cache, all connected using AMD’s 4th-generation Infinity Architecture 4. AMD’s product page separately states a transistor count of 146 billion — a figure that appears only on that live product page, not in the press release or the PDF data sheet this page also draws on, so it is reported here with that caveat 5.

4.Full spec table

SimpleStart here

24 CPU cores, 228 GPU compute units, 128 GB of shared HBM3 memory, up to 5.3 TB/s of memory bandwidth 4.

Deep diveThe technical detail
SpecAMD Instinct MI300A
CPU cores24 “Zen 4” x86 cores, three chiplets
GPU compute units228, across six accelerated compute dies
Stream processors / matrix cores14,592 / 912
Memory128 GB HBM3, unified between CPU and GPU
Memory bandwidth5.3 TB/s peak theoretical
Infinity Cache256 MB
Process nodeTSMC 5nm (compute dies) / 6nm (I/O dies)
Transistors146 billion, per AMD’s product page
TDP550W (air/liquid) or up to 760W in liquid-cooled configurations

Figures per AMD’s own MI300A data sheet and product page 45. By contrast, its sibling MI300X carries no CPU cores at all and 192 GB of memory dedicated purely to the GPU side 6 — the two chips trade CPU cores for extra GPU memory from the same underlying platform.

5.Where it’s made

SimpleStart here

Made by TSMC: compute chiplets on a 5-nanometre process, input/output chiplets on a 6-nanometre process — the same manufacturing split AMD uses for MI300X 4.

Deep diveThe technical detail

AMD’s MI300A data sheet states this directly, without the ambiguity that surrounds some earlier AMD chips such as the MI200 series, where AMD names only a process node and not a foundry 47.

6.Which systems use it

SimpleStart here

MI300A is the chip inside El Capitan at Lawrence Livermore National Laboratory — verified the world’s fastest supercomputer in November 2024, and still fastest as of November 2025 82.

Deep diveThe technical detail

Lawrence Livermore’s own November 2024 announcement: El Capitan reached 1.742 exaflops on the standard Linpack benchmark against a theoretical peak of 2.79 exaflops, and the lab states plainly, “the AMD Instinct MI300A APUs are truly the heart of El Capitan’s extraordinary capabilities” 8. The independent TOP500 list for November 2024 confirms the same 1.742-exaflop result, records the system’s power draw at 29,581 kW and its efficiency at 60.3 gigaflops per watt, and names an HPE Cray EX255a system built around AMD’s 4th-generation EPYC processors and MI300A accelerators 9. TOP500’s own listed theoretical peak for the same system is 2.746 exaflops, a little below the 2.79 exaflops in Lawrence Livermore’s figure above — a small gap between two official sources rather than an error on this page’s part. A year later, Lawrence Livermore’s own follow-up confirms El Capitan retained the title in November 2025, now measured at 1.809 exaflops 2. Neither Lawrence Livermore, HPE nor TOP500 states an exact count of MI300A chips in El Capitan in the official material this page could find — only that the system has more than 11,000 nodes and roughly 11 million CPU/GPU cores combined, so this page does not state a specific chip count 109.

7.Official pricing

SimpleStart here

Not published. AMD sells MI300A to supercomputer builders and enterprise HPC customers rather than renting it by the hour, and this page found no official price for it anywhere.

Deep diveThe technical detail

This is a genuine difference from MI300X, which several cloud providers rent by the GPU-hour with published prices 6. MI300A is positioned for large, custom HPC deployments — like El Capitan — sold through system integrators such as HPE rather than offered as a metered cloud product, and no AMD, OEM or cloud-provider source in this research states a price for it.

8.Real-world performance

SimpleStart here

AMD’s own headline figure: roughly 1.9x the performance-per-watt on FP32 HPC and AI workloads, compared with its previous chip, MI250X 1.

Deep diveThe technical detail

El Capitan’s own benchmark results are the clearest independent evidence of MI300A’s real-world performance: 1.742 exaflops sustained at launch, rising to 1.809 exaflops a year later, both figures independently confirmed by TOP500 and Lawrence Livermore rather than self-reported by AMD 829. As with any system-level result, this reflects the whole machine — more than 11,000 nodes working together — not a single-chip benchmark.

9.What came before, what came next

SimpleStart here

MI300A launched the same day as MI300X, as part of the same MI300 family — it does not have a separate predecessor or successor of its own the way MI300X does 1.

Deep diveThe technical detail

Both chips trace back to the same underlying MI300 chiplet platform, announced together on 6 December 2023 1. AMD has not announced a direct successor APU to MI300A by name; its MI250X predecessor (announced 2021) had no CPU cores at all, making MI300A the first chip in the Instinct line to combine both 7.

10.Hidden in plain sight

SimpleStart here

Sharing memory between CPU and GPU sounds like a small technical detail, but it is the specific design choice AMD credits for MI300A’s efficiency — and the reason the world’s fastest supercomputer runs on it rather than on separate CPU and GPU chips.

Deep diveThe technical detail

Most AI accelerators, including AMD’s own MI300X, keep GPU memory separate from system memory, trading flexibility for raw GPU memory capacity 6. MI300A trades some of that capacity (128 GB versus MI300X’s 192 GB) for a unified pool a CPU and GPU can both reach directly — AMD’s own explanation is “enhanced efficiency, flexibility, and programmability” for large-scale scientific software that mixes ordinary CPU code with GPU-accelerated code in the same program 3.

11.Sources

10 sources, checked September 2026. Where NVIDIA or another maker has not published a figure, this page says so rather than estimate it.

  1. AMD: AMD Delivers Leadership Portfolio of Data Center AI Solutions with AMD Instinct MI300 Series (6 Dec 2023)Official
  2. Lawrence Livermore National Laboratory: El Capitan retains title as world’s fastest supercomputer in latest Top500 list (17 Nov 2025)Official
  3. AMD: Instinct MI300 series family overview pageOfficial
  4. AMD: AMD Instinct MI300A data sheetOfficial
  5. AMD: Instinct MI300A product pageOfficial
  6. AMD: Instinct MI300X product pageOfficial
  7. AMD Investor Relations: New AMD Instinct MI200 Series Accelerators (8 Nov 2021)Official
  8. Lawrence Livermore National Laboratory: Lawrence Livermore National Laboratory’s El Capitan verified as world’s fastest supercomputer (18 Nov 2024)Official
  9. TOP500.org: TOP500 list, November 2024: El CapitanOfficial
  10. Hewlett Packard Enterprise: HPE delivers world’s fastest direct liquid-cooled exascale supercomputer, El Capitan, for Lawrence Livermore National LaboratoryOfficial

Ask an AI about this page

Opens your assistant with this page as the source, and a question rather than a summary. It will ask what you are building before it answers.

ChatGPTClaudeGeminiPerplexityGrok

Nothing is sent from here. The link carries only this page’s title and address.