AMD ‘Zen 4’ 5nm Products Will Launch In 2021, 5nm Yield Has Already Crossed 7nm
AMD has been on a red hot streak lately and it looks like it can't get anything wrong. If this report from China Times is to be believed (and this is usually a reliable source) then TSMC's 5nm testing is going very well and the first 3 customers have already been locked in - including AMD. According to the schedule obtained by China Times, AMD's 5nm products will be landing in early 2021 with mass production for 5nm scheduled in 2020.
("5nm" yield "crossing" "7nm" seems to refer to when "7nm" was at the same point in its production, around 2 years ago)
I thought it was possible that Zen 4 would slip into 2022. If it doesn't, that's good news since Intel is supposed to be stepping up its game in 2021 (after preordained failure in 2020).
"Early 2021" (March?) is an aggressive target for a Zen 4 release. However, it might make sense. Features can be moved to Zen 4 instead of debuting with Zen 3, something already indicated by previous rumors. A relatively brief duration between the Zen 3 and Zen 4 releases could work since Zen 3 will be the last generation compatible with the AM4 and SP3 sockets. Some people will upgrade to it on their existing motherboards. Zen 4 is likely to support DDR5, perhaps exclusively. Combine with chip shortages, and there won't be a glut of unsold Zen 3 chips.
1.8x density of TSMC "5nm" is a comparison to "7nm" rather than "7nm+" (source). So the core count could double with a slight increase in the size of or area covered by the chiplets.
My guess is that chiplets will continue to have 8 cores, since that is better for yields/binning. They'll just get smaller and more numerous as needed. I am not sure that core counts are going to increase much for the mainstream Ryzen lineup. 24-32 core "mainstream" Ryzen is possible but seems excessive. I will predict the elimination of desktop 6-core Ryzen CPUs with Zen 4. Just include 2 of the bad chiplets with 4 working cores each, or 1 chiplet with 8 cores, whichever is cheapest. The minimum core count for all new Ryzen chips would match PS5/Xbox core count.
One feature I'm looking out for is L4 cache stacked onto the I/O die. This is an interim step before true monolithic 3D designs, but it could be great for performance. Stacking 4 GB of HBM on the I/O die should cost relatively little. Or if HBM prices are still steep by 2021, 1-2 GB stacks or bigger stacks for the more expensive CPUs.
Zen 4: Even bigger performance leap?
This article is an excellent writeup of what is known or suspected for Zen 3 and Zen 4. I linked directly to the section discussing Zen 4 but you might want to read the whole thing. Some things to note:
* It looks like most Zen 3 CPUs will land just short of that magic (marketing) 5 GHz frequency. Although AMD will probably push to make the 16-core flagship turbo to 5 GHz instead of 4.9 GHz.
* AMD has to improve memory latency to beat Intel's slight lead in games (other areas of improvement include higher frequencies and optimization for AMD CPUs).
Update:
In an interview with Tom's Hardware, AMD's CTO, Mark Papermaster, has revealed that we can expect even more cores on next-generation Ryzen CPUs. AMD recently launched their Ryzen 9 3950X processor, featuring 16 cores, which became an instant hit with entire inventories being cleared away minutes after availability.
There are a lot of interesting details that Mark has mentioned in the interview in particular to the next-generation technologies that would be featured on their processor lineup ranging from Ryzen and EPYC CPUs. The most significant detail and the one I would start this article is with the fact that AMD isn't stopping at just 16 cores. According to AMD, there are now many applications that can scale across multiple cores and threads. The addition of cores is entirely relative to the number of applications that can take advantage of those cores so as long as this balance exists, there would not be a saturation point of cores on next-generation CPUs, whether these be mainstream or the HPC server parts.
I still don't think they will increase core counts for Zen 3. But it's clear that we will see at least a 24-core Ryzen soon.
It makes sense. If an application or game can utilize 2 or 4 cores, I wouldn't think that it could utilize 16, 32, or 64 cores. But with applications and games using 8 or 16 cores, scaling up to use 32+ cores is less of a stretch because there is already significant parallelism. On the gaming front, no less than 8 "real" cores will be used by the next-gen consoles (possibly with background tasks handled by a separate processor). Open world type games with lots of NPCs could take advantage of high core counts.
16 cores is "mainstream" and will become more common after a year or two of price drops. 8 cores is legitimately mainstream and will be coming to AMD's Zen 2 laptop APUs. 6-core CPUs will start to get pushed out of the lineup or become very cheap (such as an Athlon-branded Zen 2 APU).
(Score: 2) by takyon on Friday December 06 2019, @03:00PM
If there is a demand for processors made on these old nodes, so be it. There's a reason that Intel can still make many billions of dollars even while being embarrassed by AMD. There's a great demand for x86 chips, and AMD can't fulfill that with the capacity TSMC has allowed them to use.
In AMD's case, they released the A6-9220C and A4-9120C in 2019. Those are "28nm" dual-core Excavator APUs for laptops. Not great, but not awful in the absolute bottom feeder price segment. Which applies to chips like the dual-core "22nm" Pentium G3420.
Older process nodes could get a new life in the near future as designs like DARPA's 3DSoC could use "130nm", "90nm", "65nm", etc. nodes and still beat chips made on bleeding edge nodes.
I subbed the "Expect More Cores" article, so expect that discussion soon.