Discussion Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

[Vattila]

Except for the details about the improvements in the microarchitecture, we now know pretty well what to expect with Zen 3.

The leaked presentation by AMD Senior Manager Martin Hilgeman shows that EPYC 3 "Milan" will, as promised and expected, reuse the current platform (SP3), and the system architecture and packaging looks to be the same, with the same 9-die chiplet design and the same maximum core and thread-count (no SMT-4, contrary to rumour). The biggest change revealed so far is the enlargement of the compute complex from 4 cores to 8 cores, all sharing a larger L3 cache ("32+ MB", likely to double to 64 MB, I think).

Hilgeman's slides did also show that EPYC 4 "Genoa" is in the definition phase (or was at the time of the presentation in September, at least), and will come with a new platform (SP5), with new memory support (likely DDR5).

What else do you think we will see with Zen 4? PCI-Express 5 support? Increased core-count? 4-way SMT? New packaging (interposer, 2.5D, 3D)? Integrated memory on package (HBM)?

Vote in the poll and share your thoughts!

 

[lobz]

Another reason why AMD would release Zen 4 earlier. Not just as a response to Alder Lake. If Intel is planning to quickly follow up with another Alder Lake revision in rapid succession. It would make Zen 4 underwhelming and return AMD to a value play CPU company. By releasing Zen 4 earlier. It would be on the market before Intel has a follow up to the current Alder Lake processors. Keeping Zen 4 relevant (market leading performance) for several months and protecting their margins.

People forget that Intel is on the 10nm process that was delayed for years. They are moving to 7nm silicon by next year if not late 2022. Intel was on 14nm and still is on the CPU's previous to Alder lake. That has been 6 years on 14nm.

We all know Zen 4 will be significantly better than Alder Lake. Zen 3 should have been on 5nm TSMC silicon. AMD has a silicon cushion because they have not been on the cutting edge of nodes offered by TSMC.

All is not lost for AMD. They simply have to adjust their product release dates on Zen 4.

What in the world have I just read?!

 

[Panino Manino]

Any idea what this can be?

https://twitter.com/x/status/1494144674998599681

 

[BorisTheBlade82]

Any idea what this can be?

https://twitter.com/x/status/1494144674998599681

Account hacked?

On a serious note: Quite a surprising statement by Charlie.

 

[moinmoin]

Account hacked?

On a serious note: Quite a surprising statement by Charlie.

Nah, Charlie has quite a lot of beef with different companies' customer support. And he's right that AMD's secrecy has been to the detriment to consumer friendliness lately (I largely agree with his stance on Microsoft's Pluton in Rembrandt for example).

 

[soresu]

Account hacked?

On a serious note: Quite a surprising statement by Charlie.

It's an idiotic statement which makes me wonder if either his grasp of the English language is terrible, or if he is just trolling by contract.

It's not incompetence to say you are going to do something and then don't.

Dishonesty? Yes.

Incompetence? No.

On another note I always did find the sudden disappearance of his forums odd - anyone know why that was?

 

[MadRat]

charlie could unleash storms one day then gush praise the next during the days of Aceshardware forum. It's nothing to worry about.

 

[DrMrLordX]

Charlie might be shorting Intel lol

 

[DisEnchantment]

Another attempt to represent the Z4 die, I left 2mm2 for scribe lines etc.


Using TSMC's official scaling values shown in the column "N5 scaling" above
In following diagram, Z3 layout from die shot (top), and estimated Z4 layout (below) from calculation above.

• L2 Cache bump 2.1x, L3 Cache unchanged
  • Design would be very cache heavy. ~53% of die is cache due to the doubled L2 even though L3 shrank a lot (36%) since it did not increase in capacity (similar percentage like Z3).
• Core+L2 = ~3.88mm2 vs 4.2mm2 of Z3.
  • Core 1.4x, FPU 1.4x MTr gain
• It looks like core being around 3.9mm2 +/- 1mm2 should be a reasonable estimation.
• GMI increasing up to 10% of die is still an acceptable tradeoff based on Naffziger's presentation, to accommodate the 2x SDP

AMD's claim density of N5 at 2x scaling will obviously be better, but unknown how each area would scale, but it should be better than this and this estimate is a very conservative one.



All this while, the one thing bothering me is how AMD arrived at that 2x scaling vs TSMC's own numbers, it is not a bogus claim for sure.

 

[DisEnchantment]

Charlie might be shorting Intel lol

If you have access to info like he does, it should be possible he is using that for other purposes than writing articles on semiaccurate.

 

[Mopetar]

That kind of information might be good for some occasional quick and profitable moves, but most people are better off just investing in a market account and getting a solid annual return year over year.

The biggest opportunity for profit is small companies that can see a big spike in a short time. Something like AMD or Intel isn't going to make big moves like that unless something truly catastrophic has happened.

 

[DisEnchantment]

That kind of information might be good for some occasional quick and profitable moves, but most people are better off just investing in a market account and getting a solid annual return year over year.

The biggest opportunity for profit is small companies that can see a big spike in a short time. Something like AMD or Intel isn't going to make big moves like that unless something truly catastrophic has happened.

Folks like Charlie and Dylan provide consultation to many Investment firms with the kind of info they have. They are Analysts (SemiAnalysis and Stonearch Networking Services) or at least they consider themselves to be and they are not considered leakers.
It is very frequent to hear Charlie providing transcripts of calls with Susquehanna (quick google should help you). Others like David Schor just write some articles and not really providing other services afaik.

 

[Thibsie]

Folks like Charlie and Dylan provide consultation to many Investment firms with the kind of info they have. They are Analysts (SemiAnalysis and Stonearch Networking Services) or at least they consider themselves to be and they are not considered leakers.
It is very frequent to hear Charlie providing transcripts of calls with Susquehanna (quick google should help you). Others like David Schor just write some articles and not really providing other services afaik.

So? State clearly what you're getting at...

 

[DisEnchantment]

So? State clearly what you're getting at...

Only if you sub me and pay 1K, that's how much Charlie is charging

 

[eek2121]

Only if you sub me and pay 1K, that's how much Charlie is charging

You can buy a student membership (you don't have to be a student) for $100/year.

 

[DisEnchantment]

You can buy a student membership (you don't have to be a student) for $100/year.

I am subbed to him, but you have to take full membership, if you take student you can see only half the articles .
Anyway, sorry for OT.

I saw again some LinkedIn Profiles few days ago about AMD employees developing something on N4. But I don't wanna link anything.
Morons at click bait tech sites caused lots of issues by dumping those profile info as news.

Seems N3 issues could potentially impact AMD badly.
This is a really severe problem because if Apple does not move to N3, N5 capacity will be a huge issue if TSMC does not repurpose F18P4/5/6 for N5.
This should have been taken it into account for sure, but it is a risk to consider.
It could be a massive revenue loss for TSMC if they have F18P4/5/6 idling, waiting for the process to get fixed. Like 4-5B USD lost revenue per Q.
N5/N4 basically fabbed in the same F18P1/2/3.
If TSMC shifted F18P4/5/6 to N5 should help a lot. F18P4 anyway was supposed to do N3 and N5 both.
Someone should ask TSMC during the next earnings call.
But AMD would have known and TSMC would have informed them, last update in Oct they repeated H1'23. So they knew, since risk production was already started since last Dec not last minute surprise 1 year from now.

On the other hand, even if N3 delayed by 2Q it should still meet AMD's H2'23 schedule if not Apple's.
Thing is if AMD's chiplet approach can come to the rescue here, 65 mm2 chiplets should be able to yield enough if its only yield TSMC is having issues with.
STX would be affected if it was originally planned for N3.
The big boost in density from N3 would be necessary if Z5 core is going to be a big core as was alluded to by Clark.

 

[jamescox]

It's an idiotic statement which makes me wonder if either his grasp of the English language is terrible, or if he is just trolling by contract.

It's not incompetence to say you are going to do something and then don't.

Dishonesty? Yes.

Incompetence? No.

On another note I always did find the sudden disappearance of his forums odd - anyone know why that was?

I wonder if Threadripper 5000 just isn’t going to be released. It seems like it has been rumored to be coming several times but has not materialized. Demand for Milan might just be too high and supply may still be constrained. This isn’t just a matter of higher prices for Milan. I don’t know if Milan processors are really that much higher profit than Threadripper.

They really want to supply the server market first. If a company specs Milan processors for a new system, gets evaluation systems in for testing, makes the decision to go with Milan, and then can’t get the processors for the build, that is really bad from a reputation standpoint. Companies need to have confidence that AMD can supply the demand. This is difficult with shortages and Intel not being particularly competitive. It wouldn’t surprise me if they just gave up on Threadripper 5000, especially if Zen 4 is coming sooner rather than later. A 16 core Zen 4, even with only dual channel DDR5, might actually be serious competition for many Threadripper 5000 processors. It also might be a short lived product given the situation, cancellation seems like a possibility.

 

[soresu]

I wonder if Threadripper 5000 just isn’t going to be released. It seems like it has been rumored to be coming several times but has not materialized. Demand for Milan might just be too high and supply may still be constrained. This isn’t just a matter of higher prices for Milan. I don’t know if Milan processors are really that much higher profit than Threadripper.

Yh, for sure supply constraint really isn't something that AMD can just snap their fingers and fix.

I would say that demand for Milan is probably more than for Threadripper - demand for TR may determine whether that platform continues at all for Zen4 onward.

I think that they may just unify TR branding with the AM5 socket tbh.

 

[eek2121]

I wonder if Threadripper 5000 just isn’t going to be released. It seems like it has been rumored to be coming several times but has not materialized. Demand for Milan might just be too high and supply may still be constrained. This isn’t just a matter of higher prices for Milan. I don’t know if Milan processors are really that much higher profit than Threadripper.

They really want to supply the server market first. If a company specs Milan processors for a new system, gets evaluation systems in for testing, makes the decision to go with Milan, and then can’t get the processors for the build, that is really bad from a reputation standpoint. Companies need to have confidence that AMD can supply the demand. This is difficult with shortages and Intel not being particularly competitive. It wouldn’t surprise me if they just gave up on Threadripper 5000, especially if Zen 4 is coming sooner rather than later. A 16 core Zen 4, even with only dual channel DDR5, might actually be serious competition for many Threadripper 5000 processors. It also might be a short lived product given the situation, cancellation seems like a possibility.

Eh, they will drop it once Zen 4 drops. I can’t see them spending time and money building the chips, testing them, and then not releasing them unless sales numbers for previous gen was way down.

 

[moinmoin]

I think that they may just unify TR branding with the AM5 socket tbh.

If there'll ever be >16 cores chips on AM5 those could use the TR branding maybe.

As for the higher end I'd be all in favor of relinquishing the TR brand and having one single compatible Epyc platform, with the workstation oriented boards offering additional boost features and going by a new Epyc Workstation (neé Threadripper Pro) brand.

 

[Thibsie]

There will be a small Epyc socket already.
I'd keep TR personally.

 

[moinmoin]

There will be a small Epyc socket already.

Any news on that? Or is Charlie's teaser still all we have to go by?

 

[DrMrLordX]

Seems N3 issues could potentially impact AMD badly.

Why? They won't be first to market on N3. Intel and Apple will be.

 

[Joe NYC]

It's an idiotic statement which makes me wonder if either his grasp of the English language is terrible, or if he is just trolling by contract.

It's not incompetence to say you are going to do something and then don't.

Dishonesty? Yes.

Incompetence? No.

On another note I always did find the sudden disappearance of his forums odd - anyone know why that was?

His forum used to have a "Games" section, where people discussed (obviously) computer games.

I posted a review of a popular new game that was not exactly "politically correct" for which I received a suspension from his hand-picked "moderator".

Since then, Charlie has turned into a full-fledged Covidiot. So, he clearly limits his "truth to power" reporting to "truth to benevolent power" - or an oxymoron. Something to keep in mind about his reporting.

 

[Joe NYC]

If there'll ever be >16 cores chips on AM5 those could use the TR branding maybe.

As for the higher end I'd be all in favor of relinquishing the TR brand and having one single compatible Epyc platform, with the workstation oriented boards offering additional boost features and going by a new Epyc Workstation (neé Threadripper Pro) brand.

I don't think AM5 socket will have the features that a true HEDT would need. The only good option I see is using the "small" Genoa socket to be also for Threadripper.

But that would make too much sense, so I am not counting on it.

 

[Joe NYC]

Why? They won't be first to market on N3. Intel and Apple will be.

If Zen 5 is to launch in 2023, and if the plan was for it to launch on N3, it would be impacted.

Same for RDNA4.