AMD Technology Roadmap from AMD Financial Analyst Day 2022

1
AMD FAD 2022 Zen 5
AMD FAD 2022 Zen 5

We are covering AMD Financial Analyst Day 2022 with a series of articles today. First, we are going to cover Mark Papermaster, CTO of AMD’s, talk about the technology roadmap.

This is being done live so please excuse typos.

AMD Technology Roadmap from AMD Financial Analyst Day 2022

Even though device scaling has slowed, AMD is investing in the high-performance packaging pace to continue overall device scaling.

AMD FAD 2022 Design Innovation On High Performance Pace
AMD FAD 2022 Design Innovation On High-Performance Pace

This is the increase in engineering investment that AMD has been making. It is amazing to think about this rate of growth.

AMD FAD 2022 Investment In Engineering
AMD FAD 2022 Investment In Engineering

AMD has Zen4 coming out this year, and the denser Zen 4c coming out next year. Zen 5 is coming out in 4nm/3nm in 2024.

AMD FAD 2022 AMD CPU Core Roadmap To Zen 5
AMD FAD 2022 AMD CPU Core Roadmap To Zen 5

AMD gave an update on its core PPA versus Intel. The slide says competitor, but it is Intel as the x86 competitor.

AMD FAD 2022 AMD PPA Versus Intel
AMD FAD 2022 AMD PPA Versus Intel

AMD announced it will add AI acceleration and AVX-512 to Zen 4. Not only will we get more cores with Genoa, but we will expect an 8-10% IPC increase. 96 cores and a big IPC increase is a big deal. AMD should also get some big benefits from the new instructions. Intel is adding more acceleration in Sapphire, so

AMD FAD 2022 Zen 4 Improvements On 5nm
AMD FAD 2022 Zen 4 Improvements On 5nm

AMD also talking abotu Zen 5 for 2024. Like Zen 3 was a big step, Zen 5 will be as well.

AMD FAD 2022 Zen 5
AMD FAD 2022 Zen 5

AMD says that its march of using chiplet technology means that it is better able to build the chips of the future.

AMD FAD 2022 Compute Platform
AMD FAD 2022 Compute Platform

AMD says its 2.5D EFB (this is the MI250X chip) is a higher yield process. The 3D bonding is being used inĀ AMD Milan-X.

AMD FAD 2022 2.5D Elevated Fan Out Bridge
AMD FAD 2022 2.5D Elevated Fan Out Bridge

AMD discussed how both AMD and Xilinx saw similar needs for chiplets and packaging technology. AMD says that the two roadmaps will converge in the future.

AMD FAD 2022 Chiplet And Packaging Leadership With Xilinx Combined
AMD FAD 2022 Chiplet And Packaging Leadership With Xilinx Combined

AMD showed how its Infinity Fabric has evolved with each generation of EPYC.

AMD FAD 2022 AMD Infinity Architecture
AMD FAD 2022 AMD Infinity Architecture

The next generation will include CXL, Xilinx FPGAs, UCIe and more. This is important for the next-gen Instinct that will combine x86 cores and CDNA GPU together and ensure a more efficient shared memory model.

AMD FAD 2022 4th Gen AMD Infinity Architecture
AMD FAD 2022 4th Gen AMD Infinity Architecture

AMD says it has 40+ chiplets already in production and it plans to add more in the future along with adding 3rd party IP and chiplets for customized customer solutions.

AMD FAD 2022 Chiplet Platform Leadership
AMD FAD 2022 Chiplet Platform Leadership

These are being done super fast!

Final Words

More on Zen 5 including Zen 5 with 3D V-Cache, and Zen 5c. More chiplets and Infinity Fabric. A lot of this is forward-looking, but even the nuggets like the AVX-512 and AI coming with Genoa, this is great. There should be more on products later today.

1 COMMENT

  1. > AMD announced it will add AI acceleration and AVX-512 to Zen 4.

    By supporting AVX-512, this ties AMD to substantial engineering costs and package space requirements. Further, AVX-512 uptake has not been as strong as Intel would have liked, so it’s curious to see AMD implement it.

    I therefore wonder if this is a defensive play against oneAPI and if AMD looks to have their own equivalent that spans compute, GPU and Xilinx. This would be a huge commit by AMD but I’m not sure why else they would announce and implement support for these instructions.

    Either way, this is great for the end-user, as AVX-512 is fantastic when the hardware and workload supports it.

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.