AMD Instinct MI300X GPU and MI300A APUs Launched for AI Era

7

AMD Instinct MI300X Architecture

Underpinning the MI300X is a CDNA 3 chip as we mentioned.

AMD Instinct MI300 Family Architecture CNDA 3
AMD Instinct MI300 Family Architecture CNDA 3

Here is the MI300 family modular chiplet package. There are XCDs which are the GPU compute dies. The HBM3 memory is pretty standard, but there are eight stacks. AMD also has IO Dies with AMD Infinity Cache underneath the XCDs.

AMD Instinct MI300X Accelerator Large
AMD Instinct MI300X Accelerator Large

Here is the basics of the XCD. We are not going to re-type what is already on the slide, but this is the basic CDNA 3 accelerator complex.

AMD Instinct MI300 MI300X MI300A Architecture XCD
AMD Instinct MI300 MI300X MI300A Architecture XCD

That new CDNA 3 compute has a number of different numerics. Clearly there was a focus on FP64 compute, but the chip is so big and flexible enough that it can do the FP8, TF32, and more.

AMD Instinct MI300 Family Architecture Compute Enhancements
AMD Instinct MI300 Family Architecture Compute Enhancements

Here is the 128 channel interleaved HBM3 memory interface. AMD also has the 256MB of Infinity Cache with 17TB/s of peak bandwidth.

AMD Instinct MI300X Architecture Memory Subsystem
AMD Instinct MI300X Architecture Memory Subsystem

AMD also has to move data from chiplet to chiplet. AMD is showing bandwidth, and it has things like Infinity Cache to hide this a bit, but there is a latency hit whenever traversing the chiplet architecture.

AMD Instinct MI300X Architecture IO Subsystem
AMD Instinct MI300X Architecture IO Subsystem

There is a lot here, but let us get to the MI300A and see how the MI300 family can also be an APU. If you are looking for the MI300A architecture discussion like the one for MI300X above, skip a page.

7 COMMENTS

  1. It’s a good question which of the MI300A or MI300X is going to be more popular. As a GPU could the MI300X be paired with Intel or even IBM Power CPUs?

    I personally find the APU more interesting. Not because the design is new so much as the fact that real problems are often solved using a mixture of algorithms some of which work well on GPUs and others better suited to CPUs.

  2. I hope to see some uniprocessor MI300A systems hit the market. As of today only quad and octo.
    Maybe a sort of cube form factor, PSU on the bottom, then mobo and gigantic cooler on the top. A SOC compute monster.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

This site uses Akismet to reduce spam. Learn how your comment data is processed.