I recall AMD in multiple instances refrerring to 7 nm as a long-term node, so it wouldn't surprise me if they skipped 5 nm, or only got into it late for some niche products, and concentrated on a full-node shrink to something like 3 nm instead. 2 years of 7 nm just doesn't sound "long-term" enough for me.
If the semiconductor industry does produce < 5 nm process nodes, they will need to transition away from silicon. After GlobalFoundries canceled development of 7 nm, I feel like a lot is up in the air now.
So it sounds like Zen 3 will mostly be a change to TSMC’s N7+ process which is likely a mild improvement or in prior Intel marking its the tick and Zen 4 would be the tock.
Zen 3 will basically be Zen 2+, i.e. an optimization of the same design with a slightly better node, just like Zen+ was for the original Zen. Zen 4 will be a new design.
It would require a new IO chip with more IFOPs (Infinity Fabric On Package). A 12 IFOP IO chip could connect up 96 cores, in theory. You can see where they would go on the package.
However the memory contention starts becoming an issue, it probably already is with 64 cores, despite the faster DDR4 support. So this may be something that waits for DDR5.
Also heat density on the package. 96 cores, but they would be slower than 64 cores to reduce power consumption and heat generation.
You rather mean Zen 4, which should also support DDR5, have a new socket*, possibly move from 8 to 12 memory channels, etc Zen 3 will merely be an optimized Zen 2 (effectively Zen 2+), not a new design, just as Zen+ was for the original Zen. And if AMD adheres to the same pattern they will *not* release Zen 3 based Epyc CPUs, just as they did not release Zen+ ones.
Zen 3 based CPUs should be the last generation with the same socket (AMD promised up to 2020) and only be released in mainstream and (possibly) Threadripper versions. These CPUs will certainly not have a different number of cores or any other significant new features, that will have to wait for Zen 4.
*It's unclear if you can add PCIe 4.0 support and retain the same socket. Perhaps it's not possible and AMD's promise of keeping the same socket up to 2020 was about their mainstream and perhaps HEDT CPUs, not their server ones, thus they are due to change the socket of Epyc in 2019. But that would mean no PCIe 4.0 support for both Zen 2 *and* Zen 3 mainstream and HEDT CPUs, with such support being added to Zen 4 in 2021. I highly doubt they would do that, so I wonder how they are going to handle this.
With the central IO die there should no longer be a need for the processor cores to have to be designed around a particular memory standard. If all AMD has to do to support DDR5 is release a new IO die we could see DDR5 support very quickly if AMD is happy to create another socket standard. This could add to Intel's headaches as they would have to redesign their main CPUs rather than just a support chip.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
16 Comments
Back to Article
Hul8 - Tuesday, November 6, 2018 - link
I recall AMD in multiple instances refrerring to 7 nm as a long-term node, so it wouldn't surprise me if they skipped 5 nm, or only got into it late for some niche products, and concentrated on a full-node shrink to something like 3 nm instead. 2 years of 7 nm just doesn't sound "long-term" enough for me.Opencg - Wednesday, November 7, 2018 - link
3nm will probably not come until 20-30 years from now if it even comes at allnandnandnand - Wednesday, November 7, 2018 - link
Not true:https://www.eetimes.com/document.asp?doc_id=133097...
Samus - Thursday, November 8, 2018 - link
1nm, sure, but certainly not 3nm. That's <5 years out.PixyMisa - Saturday, November 10, 2018 - link
Real 3nm is unlikely, but marketing 3nm is on track for 2022.Tvegas1980 - Thursday, May 23, 2019 - link
There was an article that was talking about picometer node which is a lot smaller than nanometers so 1 nm is not impossible.Dramaking - Thursday, November 8, 2018 - link
If the semiconductor industry does produce < 5 nm process nodes, they will need to transition away from silicon. After GlobalFoundries canceled development of 7 nm, I feel like a lot is up in the air now.FreckledTrout - Tuesday, November 6, 2018 - link
So it sounds like Zen 3 will mostly be a change to TSMC’s N7+ process which is likely a mild improvement or in prior Intel marking its the tick and Zen 4 would be the tock.nandnandnand - Tuesday, November 6, 2018 - link
If I'm not mistaken, Zen 2 is a 25% improvement (IPC combined with higher clock rates) before accounting for any additional cores.No performance increase is indicated for N7+, but it will probably be better than nothing, and a good breather like 12nm Zen+:
https://www.anandtech.com/show/13445/tsmc-first-7n...
Santoval - Wednesday, November 7, 2018 - link
Zen 3 will basically be Zen 2+, i.e. an optimization of the same design with a slightly better node, just like Zen+ was for the original Zen. Zen 4 will be a new design.Smell This - Wednesday, November 7, 2018 - link
Do Zen2 'chiplets' in EPYC Rome CPUs scale beyond 64 cores?Is this the practical limit of the (old Sea Micro) Infinity ('Freedom') Fabric?
psychobriggsy - Friday, November 9, 2018 - link
It would require a new IO chip with more IFOPs (Infinity Fabric On Package). A 12 IFOP IO chip could connect up 96 cores, in theory. You can see where they would go on the package.However the memory contention starts becoming an issue, it probably already is with 64 cores, despite the faster DDR4 support. So this may be something that waits for DDR5.
Also heat density on the package. 96 cores, but they would be slower than 64 cores to reduce power consumption and heat generation.
So, Zen 3, N7+, DDR5 - 96 cores seems viable.
Santoval - Tuesday, November 13, 2018 - link
You rather mean Zen 4, which should also support DDR5, have a new socket*, possibly move from 8 to 12 memory channels, etc Zen 3 will merely be an optimized Zen 2 (effectively Zen 2+), not a new design, just as Zen+ was for the original Zen. And if AMD adheres to the same pattern they will *not* release Zen 3 based Epyc CPUs, just as they did not release Zen+ ones.Zen 3 based CPUs should be the last generation with the same socket (AMD promised up to 2020) and only be released in mainstream and (possibly) Threadripper versions. These CPUs will certainly not have a different number of cores or any other significant new features, that will have to wait for Zen 4.
*It's unclear if you can add PCIe 4.0 support and retain the same socket. Perhaps it's not possible and AMD's promise of keeping the same socket up to 2020 was about their mainstream and perhaps HEDT CPUs, not their server ones, thus they are due to change the socket of Epyc in 2019. But that would mean no PCIe 4.0 support for both Zen 2 *and* Zen 3 mainstream and HEDT CPUs, with such support being added to Zen 4 in 2021. I highly doubt they would do that, so I wonder how they are going to handle this.
levizx - Saturday, November 10, 2018 - link
IF has nothing to do with Freedom Fabric. FF is more like inter board connection, IF is a direct descendant of HTCheapSushi - Wednesday, November 7, 2018 - link
Do you think DDR5 will come with Zen 3 or Zen 4?RogerAndOut - Thursday, November 8, 2018 - link
With the central IO die there should no longer be a need for the processor cores to have to be designed around a particular memory standard. If all AMD has to do to support DDR5 is release a new IO die we could see DDR5 support very quickly if AMD is happy to create another socket standard. This could add to Intel's headaches as they would have to redesign their main CPUs rather than just a support chip.