AMD’s Instinct MI300X and MI300A: Advanced AI Accelerators Blending CDNA 3 and Zen 4

AMD’s upcoming Instinct MI300X and MI300A accelerators, set to launch next month, have generated significant anticipation within the AI sector. Positioned as AMD’s primary focus and strategic priority, these advanced AI solutions are designed to meet industry demands. Scheduled for full unveiling on December 6th during AMD’s “Advancing AI” keynote, the Instinct MI300 family, codenamed MI300, promises to be a chipset powerhouse utilizing TSMC’s advanced packaging technologies.

Everything You Need to Know About AMD Instinct MI300X and MI300A

The AMD Instinct MI300X, a key highlight, targets NVIDIA’s Hopper and Intel’s Gaudi accelerators in the AI segment. Built on the CDNA 3 architecture, it incorporates a mix of 5nm and 6nm IPs, boasting up to 153 billion transistors.

The design features an interposer with 28 dies, including eight HBM3 packages, 16 dummy dies, and four active dies. With 40 compute units per GCD (Graphics Compute Die), totaling 320 Compute & 20,480 core units, the MI300X aims to challenge competitors with its significant memory upgrade of 192 GB HBM3, offering up to 5.2 TB/s bandwidth.

In terms of power consumption, the MI300X is rated at 750W, a 50% increase over its predecessor, the MI250X, and 50W more than the NVIDIA H200. The Instinct MI300A realizes AMD’s promise of an Exascale-class APU, featuring a design similar to the MI300X but incorporating TCO-optimized memory capacities and Zen 4 cores.

With two CDNA 3 GCDs replaced by three Zen 4 CCDs on one active die, the MI300A delivers a total of 24 cores and 48 threads. Key features include 153 billion transistors, up to 24 Zen 4 cores, CDNA 3 GPU architecture, and a memory capacity of 192 GB HBM3. Here are some features of AMD Instinct MI300 Accelerators:

• First Integrated CPU+GPU Package
• Aiming for the Exascale Supercomputer Market
• 153 Billion Transistors
• Up to 24 Zen 4 Cores
• CDNA 3 GPU Architecture
• Up to 192 GB HBM3 Memory
• Up to 8 Chiplets + 8 Memory Stacks (5nm + 6nm process)

AMD aims to offer MI300 AI accelerators in 8-way configurations, featuring SXM designs connecting to mainboards with mezzanine connectors. While competitors like NVIDIA and Intel are also advancing in the AI space, AMD’s MI300 seeks not only to be an alternative but a leader in the AI segment. As AI demand surges, AMD’s formidable solution, emphasizing innovation and performance, is poised to make a significant impact.