Key points from their site if you don't want to watch video.
AMD Analyst Day 2020: Zen 3, Zen 4, Navi 2X
AMD just held its AMD Financial Analyst Day, where it outlined several developments on the horizon. There’s a lot of information here, so we’re just going to aggregate and parse it, and we may delve deeper into them separately sometime later. A lot of this is information we’ve seen before, some of it is not.
- AMD has shipped more than 260M x86 Zen cores since the launch of Zen in 2017, and according to AMD, the rate is doubling.
- Zen 3 will begin rollout in late 2020, with the full product stack being available by end of 2021. While we (and everyone else) previously understood Zen 3 to use TSMC’s N7+ process, AMD seems to be walking that back, dropping the “+” designation from its most recent roadmaps. AMD clarified that while Zen 3 will use enhancements beyond vanilla 7nm, it didn’t say Zen 3 won’t use EUV, but that it’ll vary by case.
- As a reminder, TSMC currently has two process nodes beyond its first generation N7. There’s the N7P, which is DUV-based. N7P is mostly an optimized continuation of N7 that recycles the same design rules and tools as N7, so it’s IP-compatible. Then there’s the N7+, which is TSMC’s first process that adapts EUV for certain layers. We’re sure to learn more about this in the future.
- AMD is currently developing Zen 4, targeting a 2022 release. Zen 4 cores will be built on 5nm.
- Epyc Milan, the successor to Epyc Rome, is currently on track for a late 2020 launch. Milan will also use the Zen 3 architecture and use an improved iteratrion of 7nm; whether that’s TSMC’s N7P or N7+ remains to be seen. Further out is Genoa, which we’ve briefly mentioned in AMD’s El Capitan Supercomputer news. Epyc Genoa will be Zen 4-based and built on 5nm for a 2022 release.
- AMD’s Infinity Fabric interconnect is now known as Infinity Architecture, beginning with the 3rd generation of the technology. AMD previously announced that El Capitan will use the 3rd-gen Infinity Architecture, and touts an increased CPU to GPU memory coherency. Other than a rough 2022 release date, not much more is known yet.
- AMD is splitting its GPU architectures: RDNA for client, CDNA for compute/data center. As for how CDNA will differentiate from RDNA-based gaming GPUs remains to be seen. AMD shared a roadmap outlining two generations of CDNA: CDNA 1 at 7nm, and CDNA 2 on an unspecified node for 2022. We know CDNA 2 will rely heavily on the 3rd-gen Infinity Architecture. No doubt, El Capitan is set to use CDNA 2.
- Navi 2x seems to be what AMD is targeting for ‘Big Navi’. Navi 2x as will be based on RDNA 2, and according to AMD RDNA 2 will bring a 50% performance-per-watt improvement; this improvement will extend further with RDNA 3, set for 2022. RDNA 2 based GPUs are expected late 2020.
- AMD also unveiled what appears to be a response to Intel’s Foveros technology, in the form of what AMD calls X3D die stacking. AMD is framing X3D die stacking and packaging as a natural evolution from its past MCM and chiplet designs, seeing AMD stack multip die atop a single interposer. Exactly what dies/elements AMD plans to incorporate here aren’t known. However, AMD is promising a 10x increase/improvement in bandwidth density.
