AMD CPU Roadmap - stagnation coming?

[ARF]

We all know that TSMC N5 and N4 are basically the same process with some small, rather insignificant changes.

This roadmap hints that AMD will move to consumer Zen (6?) Ryzen chips on N3 - the next shrink - as late as the next-next generation, or Zen 6. Because Zen 5C won't be used in Ryzen?

What do you think about this?

AMD CEO, Dr. Lisa Su, To Visit TSMC Next Month To Discuss Future 2nm, 3nm Chip Projects (wccftech.com)

AMD CEO, Dr. Lisa Su, To Visit TSMC Next Month To Discuss Future 2nm, 3nm Chip Projects (wccftech.com)

 

[Mussels]

Sorry, what?


Intel stayed on 14nm for a good 200 years there, how is THIS stagnant?


You do realise that percentagewise, the change from 7nm to 3nm is really really tiny
Like... in a big way. The big tiny. Yeaaaah. Take that, words.

 

[ARF]

Yes, intel used 14nm for several years and this allowed AMD to catch.

AMD will have problems because from 2022 to 2025-2026 it will use the same process, while intel will very aggressively push the more cores agenda.

Intel Next-Gen Desktop CPUs Confirmed: 13th Gen Raptor Lake This Year, 14th Gen Meteor Lake In 2023, 15th Gen Arrow Lake In 2024, 16th Gen Lunar Lake In 2025 (wccftech.com)

 

[TheoneandonlyMrK]

Yes, intel used 14nm for several years and this allowed AMD to catch.

AMD will have problems because from 2022 to 2025-2026 it will use the same process, while intel will very aggressively push the more cores agenda.

View attachment 262734
Intel Next-Gen Desktop CPUs Confirmed: 13th Gen Raptor Lake This Year, 14th Gen Meteor Lake In 2023, 15th Gen Arrow Lake In 2024, 16th Gen Lunar Lake In 2025 (wccftech.com)

I think they have enough levers to pull personally.

 

[Bones]

Yes, intel used 14nm for several years and this allowed AMD to catch.

AMD will have problems because from 2022 to 2025-2026 it will use the same process, while intel will very aggressively push the more cores agenda.

View attachment 262734
Intel Next-Gen Desktop CPUs Confirmed: 13th Gen Raptor Lake This Year, 14th Gen Meteor Lake In 2023, 15th Gen Arrow Lake In 2024, 16th Gen Lunar Lake In 2025 (wccftech.com)

I've seen nothing confirming AMD sticking with the same process at that time (2025-26) in the linked articles.

To be fair, there is mention of 2024+ for Zen 5 but at this time it's also fair to state it that way.

No one, not even AMD knows how development will go and succeed at that time as of today. Because of that, nothing of it is confirmed at this time, we'll know more as those dates get closer.

As these dates approach we'll see how progress towards newer nodes is doing and be able to more accurately predict what's to come at those time period "Targets".
That does mean there could be breakthroughs that allows AMD to proceed to a newer node even faster than what's being predicted today - You never know what's going to be concerning R&D until it happens as fact, not as conjecture.

The same thing applies to Intel, they too could come up with something special along the way that lets them claw their way back on top.

Concerning the statement about Intel, that to me is calling the kettle black by Intel since awhile back they were claiming AMD was "Gluing" cores together - Remember that?

Must have been some damn good glue because it worked, leading to Intel getting clobbered and now Intel, like it or not has to come up with their own glue formula and hope it works even half as well at this point if they really intend to pursue the "Moar cores" agenda they harped about against AMD for so long.

The overall point is it's fine to forecast what you think will be at that time, just don't go saying it today as if it's already written in stone at those times tomorrow because it's not.

 

[Vayra86]

The pessimist in me says that AMD needs to start working on big little in chiplets sooner rather than later if they want to maintain leadership.

 

[SL2]

Zen 2 is 7 nm

Zen 3 is 7 nm

Zen 4 is 5 nm

Zen 5 is 4 nm

Zen 6 is possibly 3 nm

I see no stagnation. Zen 3 should have been a sign of stagnation if anything, look how that turned out. No, we're not talking about availability here.

Also, isn't Zen 5 supposed to be the biggest architectural change since 2017? No, I have no link for that.

This thread makes zero sense.

AMD has been on 7 nm for over three years ( = OMFG STAGNATIONZz) while still dominating the sales with two year old CPU's, which I still don't understand since Alder is fast and price competitive.
Yeah I know all about old AM4 systems being upgraded, but I still am surprised.

 

[ModEl4]

When AMD implements their big-little equivalent whatever multithreading advantage Intel has with the use of Ecores will go poof.
There was rumors that Nova Lake (Q4 2025?) is going to bring huge IPC improvements, but these are just rumors.If this doesn't happen, AMD will probably be competitive with Intel for the next 5 years (unless we have a major China/Taiwan incident)

• Nova Lake (Panther Cove [tentative]/ Darkmont) 2025
  This will mark the biggest architectural change in cpu architecture since the Core architecture is introduced in 2006. Intel is working to build an entirely new architecture from the ground up much like Ryzen with up to 50% cpu performance improvement from lunar lake. This is also the reason why Glenn Hinton returned.

 

[Bill_Bright]

This thread makes zero sense.

I have to agree. Why? Because the "competitive edge" is an erratic, unpredictable, rapidly moving and perhaps most importantly, "blurry" target.

By blurry I mean that "edge" is really multifaceted with several edges and the most important (leading?) edge depends on the individual consumer's perspective/priority and NOT the author of some article, blogger or any forum poster.

Which of the following determines the leading competitive edge?

The fastest clock speed?​

The most cores?​

The most threads?​

The smallest process node/architecture?​

The best efficiency?​

The lowest power consumption?​

The lowest heat generation?​

The highest heat tolerance?​

The highest MIPS (million instructions per second) value?​

The best integrated graphics (if applicable)?​

The best security features?​

The deepest pockets?​

The most sales on Amazon?​

The cheapest price?​

The consumer's personal preference/loyalty?​

All of the above?​

None of the above?​

Depends?​

Ask me tomorrow.

I say AMD and Intel will remain competitive for decades to come, barring any one or more of the following:

A major war,​

A devastating natural disaster,​

A global shortage of raw materials,​

A really stupid upper-management business decision,​

A hostile take-over that obliterates one company or the other,​

A friendly merger,​

A totally unprecedented, massive, upper-level management criminal scandal,​

Something totally unforeseen.​

Truth be told we (everyone) needs both AMD and Intel to succeed and thrive. We need Intel to keep looking over their shoulders to make sure AMD does not pass them (again) and we need AMD to keep nipping at Intel's heels. If they leapfrog and change places again, that's fine. As long as one keeps innovating and the other keeps up, it is a win/win/win.

 

[ARF]

lol, very funny

Which of the following determines the leading competitive edge?

Top performance, performance crown, and the trend of market share - number of sales.

I guess we will know better in several days - Zen 4 vs. Raptor Lake as a starting point.

Do you remember when AMD had 80% sales at mindfactory? No one talks about that anymore.

 

[v12dock]

How many times has Intel had to delay a project and it’s not just a 10nm issue. Sapphire Rapids and ARC are newer projects that they have had to push out continually. I would take Intel’s roadmap with a grain of salt. Sure Pat is “better” CEO than some in the past but I think he a little cocky. He still has this mentality that Intel has superior engineers when AMD has been parrying or beating them for several years now.

1 year ago he said “AMD is in the rear-view mirror" and then a few weeks ago he said AMD has to much momentum? Honestly I think Pat is kind of a jackass, but the more I have to use VMWare I get more and more frustrated on how VMWare operates in which he was the CEO of for 8 years.

 

[evernessince]

Either you are saying TSMC will stagnate and by extension this will harm AMD's processor progression or that AMD will not pursue the latest TSMC nodes in a timely fashion. Either scenario in extremely unlikely. Given recent history it is far more likely that Intel will face more issues regarding cutting edge nodes than TSMC.

 

[mplayerMuPDF]

Either you are saying TSMC will stagnate and by extension this will harm AMD's processor progression or that AMD will not pursue the latest TSMC nodes in a timely fashion. Either scenario in extremely unlikely. Given recent history it is far more likely that Intel will face more issues regarding cutting edge nodes than TSMC.

Why is it "extremely unlikely" that TSMC will stagnate? Past performance is not a guarantee for the future. No reason that TSMC could not start struggling as they run into certain physics related limits. Every golden age comes to an end. If that happens though, AMD will probably be slowed down a bit but I am sure that there will still be some meaningful progress. After all, AMD was still able to advance a decent amount while being stuck on GlobalFoundries (and TSMC if you include cat chips) 28 nm (which was tweaked over time, of course, like Intel's 14 nm).

 

[Wirko]

The pessimist in me says that AMD needs to start working on big little in chiplets sooner rather than later if they want to maintain leadership.

When AMD implements their big-little equivalent whatever multithreading advantage Intel has with the use of Ecores will go poof.

AMD's big-little is not a given. It's more likely in notebook-oriented APUs, which come to desktop later.

But what about Zen CCDs? Of course they are exploring architectural changes to achieve (a) performance per thread very high when few threads are running and (b) total performance per mm2 very high with highly multithreaded loads.
But also (c) both of those goals in a single chiplet if possible, in order to bring some benefits of big-little to midrange, single-CCD Ryzens.
Not to forget that (d) same chiplet goes into server CPUs where MT performance is what really counts, in most cases.

But AMD may take more exotic path, or paths, to get there. One possibility is more than 2-way SMT. Not necessarily 4-way but 3-way. They're not afraid of odd numbers like "three", and they've gathered enough experience with SMT by now, so they can assess both costs and benefits.

Another possibility is the return of some elements of their clustered multithreading (CMT) - without abandoning SMT, of course. Pairs of cores can share some structures, L2 for example, in a manner similar to Intel's E-clusters. They've gathered enough experience with CMT by now, maybe the engineers can extract something good from it.

Or they may see some benefit in implementing IBM's combined L2-L3 cache in some form, license it and make it work. IBM Z architecture is very distant from x86 - but its CPU package has two chiplets of 8 cores each, which isn't so distant.

Is there any solid info regarding increased number of cores per CCD in Zen 5? It would be weird, for both Ryzen and Epyc, if they stay at 8.

Also, I want to believe that the 3D cache die can extend L2 too, not just L3. Just a crazy thought. Too crazy.

 

[mechtech]

We all know that TSMC N5 and N4 are basically the same process with some small, rather insignificant changes.

This roadmap hints that AMD will move to consumer Zen (6?) Ryzen chips on N3 - the next shrink - as late as the next-next generation, or Zen 6. Because Zen 5C won't be used in Ryzen?

What do you think about this?
View attachment 262731
AMD CEO, Dr. Lisa Su, To Visit TSMC Next Month To Discuss Future 2nm, 3nm Chip Projects (wccftech.com)


AMD CEO, Dr. Lisa Su, To Visit TSMC Next Month To Discuss Future 2nm, 3nm Chip Projects (wccftech.com)

ARM and RISC will make them both (Intel & AMD) irrelevant by 2026

 

[Shrek]

The industry is hitting a wall, given that a silicon atom is about 0.2 nm across.

There's a limit to how small we can make transistors, but the solution is photonic chips — Quartz (qz.com)

 

[Steevo]

Considering their performance is neck and neck, it’s more of a process optimized and which minor difference on what application, or what hardware acceleration parts have been added.

I don’t think we will see double digit IPC increases ever again due to architecture changes. Increases from better memory, lower latency, prefetch and speculative execution on small cores with cached results maybe

 

[Shrek]

On chip RAM (like the Apple M series) will help with access speed, but I hope it will also allow for additional external RAM (unlike the M series)

 

[Wirko]

The industry is hitting a wall, given that a silicon atom is about 0.2 nm across.

There's a limit to how small we can make transistors, but the solution is photonic chips — Quartz (qz.com)

That's until you calculate how big a photon is.

 

[ModEl4]

AMD's big-little is not a given. It's more likely in notebook-oriented APUs, which come to desktop later.

But what about Zen CCDs? Of course they are exploring architectural changes to achieve (a) performance per thread very high when few threads are running and (b) total performance per mm2 very high with highly multithreaded loads.
But also (c) both of those goals in a single chiplet if possible, in order to bring some benefits of big-little to midrange, single-CCD Ryzens.
Not to forget that (d) same chiplet goes into server CPUs where MT performance is what really counts, in most cases.

But AMD may take more exotic path, or paths, to get there. One possibility is more than 2-way SMT. Not necessarily 4-way but 3-way. They're not afraid of odd numbers like "three", and they've gathered enough experience with SMT by now, so they can assess both costs and benefits.

Another possibility is the return of some elements of their clustered multithreading (CMT) - without abandoning SMT, of course. Pairs of cores can share some structures, L2 for example, in a manner similar to Intel's E-clusters. They've gathered enough experience with CMT by now, maybe the engineers can extract something good from it.

Or they may see some benefit in implementing IBM's combined L2-L3 cache in some form, license it and make it work. IBM Z architecture is very distant from x86 - but its CPU package has two chiplets of 8 cores each, which isn't so distant.

Is there any solid info regarding increased number of cores per CCD in Zen 5? It would be weird, for both Ryzen and Epyc, if they stay at 8.

Also, I want to believe that the 3D cache die can extend L2 too, not just L3. Just a crazy thought. Too crazy.

Sure i agree it's not a given, but it seems to be the most probable path imo.
All the alternative paths that you mentioned are known and already debated in most sites forums 2 years back before Zen3 launch regarding what path AMD will choose.
Like you said the use that makes most sense for a low power, small die size design is in mobile (and in some server cases) and this will come anyway.
The challenges regarding universal increase in single thread performance that arise if you increase core count in CCD or if you increase SMT to 3-way etc are more difficult to solve than just add a second compatible low power/small die size CCD that you already have anyway, and keep the high performance CCD at 8C/16T.
I'm not saying that what you say isn't potentially coming eventually, i just think that in the immediate future timeframe (2024) it makes more sense for AMD in client space at least to choose that path economically and based on what competition is offering.
But i agree that this isn't set in stone and that AMD have other alternatives also!

 

[Bill_Bright]

Top performance, performance crown, and the trend of market share - number of sales.

LOL

"Top performance" and "performance crown" are just 2 ways of saying the same thing - so only count as one criteria.

But it also brings up another "blurry" point. No CPU provides the "best performance" for every task. For example, some CPUs are better at gaming while others are better at CAD/CAE tasks. And others still are better for file servers. Even among the most popular games, no single CPU (or CPU brand) is the best performer.

"Trend in market share" and "number of sales"? No way. Those only indicate which company has done the better job at marketing and pricing (and perhaps distributing) their product.

I mean McDonald's sells over 1/2 billion Big Macs every year - by a very large margin, more than any other. And there is no denying, it tastes good (if you add bacon! ). But does dominating global market share make the Big Mac the best burger in the world?

 

[ARF]

Yes, intel used 14nm for several years and this allowed AMD to catch.

AMD will have problems because from 2022 to 2025-2026 it will use the same process, while intel will very aggressively push the more cores agenda.

View attachment 262734
Intel Next-Gen Desktop CPUs Confirmed: 13th Gen Raptor Lake This Year, 14th Gen Meteor Lake In 2023, 15th Gen Arrow Lake In 2024, 16th Gen Lunar Lake In 2025 (wccftech.com)

Wccftech has updated this roadmap - now it looks much worse for Intel ?

Intel Client & Server CPU Roadmap Updates: Meteor Lake In 2023, 20A & 18A Powered Xeons & Core Chips Beyond 2024 (wccftech.com)

 

[AusWolf]

1. How do you expect stagnation based on a roadmap that only mentions process nodes?
2. Why worry about stagnation? I mean, sure, innovation is good because we have shiny new toys, but stagnation is good too, because we can keep our shiny new toys longer.

 

[ARF]

Because if Mooré's law is indeed dead, we have a major problem in our society - as we know they can't shrink those transistors for forever and ever, because they enter the quantum effects pretty fast and we are not ready with it just yet.

Stagnation in PCs is really bad - we have powerful RTX 4090 and the vast majority have crappy TN 1080p panels with 6-bit colour.

We need this to end and to move to 2160p with 10-bit OLED HDR panels for all PC monitors sizes.

 

[cvaldes]

Stagnation in PCs is really bad - we have powerful RTX 4090 and the vast majority have crappy TN 1080p panels with 6-bit colour.

I'm not aware of any graphics cards that have built-in displays (other than a couple of models that have some sort of stats panel for showing temperature, load, etc.). Even the displays in notebook computers are separate line items on the BOM.

And I've never heard of a 4090 in a notebook computer.

Owners of 4090 graphics cards are free to use whatever display they'd like. If some of them are professional esports competitors and a 1080p TN panel gives them the edge to win, that's fine too. It's their money.