Mark Papermaster, chief expertise officer of AMD, stated at a convention that the corporate’s next-generation Intuition MI300 accelerated processing unit for information facilities and high-performance computing is already up and operating in AMD’s labs. The APU makes use of AMD’s Zen 4 and CDNA 3 architectures and can energy the El Capitan supercomputer that’s anticipated to interrupt the two ExaFLOPS obstacles generally in 2024.
“However with what we introduced and have rolled out with our next-generation Intuition that we have already got again within the labs, the MI300, it’s a true datacenter APU,” stated (opens in new tab) Papermaster on the Wells Fargo 2022 TMT Summit (through SeekingAlpha (opens in new tab)). “It’s a CPU and a GPU acceleration that’s leveraging the Infinity structure to share the identical reminiscence absolutely coherently. It’s all sharing high-bandwidth reminiscence.”
AMD’s Intuition MI300 is a multi-chiplet APU carrying Zen 4 general-purpose x86 cores, and CDNA 3-based compute GPU that share a unified on-package pool of reminiscence comprising of Infinity Cache in addition to HBM. AMD will use TSMC’s N5 (5nm-class) fabrication course of to supply CPU/GPU chiplets for the MI300, however the firm has but to reveal what number of CPU cores and GPU stream processors are set to be packed into the hybrid processor.
Placing CPU cores and GPU accelerators into one bundle and making them use unified reminiscence is the Holy Grail of supercomputing because it not solely combines the most effective of each worlds (basic goal processing and extremely parallel acceleration), however it additionally dramatically simplifies programming and makes it simpler to get all efficiency potential out of execution engines.
The Intuition MI300 APU is about to energy the U.S. Division of Power’s El Capitan supercomputer, which can be put in on the Lawrence Livermore Nationwide Laboratory (LLNL) someday in 2023 and can turn into absolutely operational in mid-2024.
AMD plans to roll out its Intuition MI300 someday in 2023, which is why it’s not shocking that the processor is already up and operating in its labs. In the meantime, it’s unclear whether or not the corporate will provide the MI300 to any shoppers exterior the supercomputer world in mass portions subsequent 12 months.
Since AMD’s Intuition MI300 would be the trade’s first information middle and HPC-grade APU, there can be lots of curiosity in it within the supercomputing neighborhood. In the meantime, AMD’s Intuition MI300 won’t be the one hybrid processor for information facilities packing x86 and GPU sources that can be accessible within the coming years. For instance, Intel is engaged on its codenamed Falcon Shores product with x86 and Xe-HPC cores in the identical bundle and expects to make it accessible in 2024.