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!

 

[Ajay]

It doesn't make fiscal sense for ASML to "ramp production" for something which has basically three customers in the whole world none of whom will need even three digit quantities of this equipment for years.

Sure, they'd all like to get EUV machines more quickly, but if you build the capacity to give them what they want quickly, then you have a bunch of capacity sitting idle for months/years until they need more. The device is so specialized and complex it isn't like they can turn around and start making something else with the same production equipment.

Its like making ultra high end supercars like Konigsegg. When they introduce a new model, they might get 50 people (not sure of the exact volume, but they ain't Ferrari or even Lamborghini) who want one. They would all prefer if they got it immediately, but Konigsegg might produce them over a year or two so just about all of them will have to wait to take delivery. If they geared up to produce all 50 in a week they'd have to lay off their employees and pay rent on an empty factory the rest of the time.

Sorry Doug, that's just a terrible analogy. ASML's book to bill ratio is simultaneously enviable and tragic. There huge sums of money being left on the table here. I don't know why they haven't expanded capacity - I haven't even looked it up yet. But it would be foolish for them to hamper production on the basis of exclusivity. Their products are born of real needs by real business partners - not persons who's vanity is matched by the size of their wallet. There has to be another reason.

 

[turtile]

Sorry Doug, that's just a terrible analogy. ASML's book to bill ratio is simultaneously enviable and tragic. There huge sums of money being left on the table here. I don't know why they haven't expanded capacity - I haven't even looked it up yet. But it would be foolish for them to hamper production on the basis of exclusivity. Their products are born of real needs by real business partners - not persons who's vanity is matched by the size of their wallet. There has to be another reason.

I have no idea of the real reason but my guess is that it is a long process to hire new, highly qualified employees for both production and support for the machines. I assume there is no room for error.

 

[wlee15]

Sorry Doug, that's just a terrible analogy. ASML's book to bill ratio is simultaneously enviable and tragic. There huge sums of money being left on the table here. I don't know why they haven't expanded capacity - I haven't even looked it up yet. But it would be foolish for them to hamper production on the basis of exclusivity. Their products are born of real needs by real business partners - not persons who's vanity is matched by the size of their wallet. There has to be another reason.

Backlogs aren't uncommon in capital intensive industries. For example when the 787 backlog fell to 546 planes or about a 3 year backlog Boeing responded by slashing production. ASML backlog in contrast looks to be just 2 years.

 

[maddie]

Backlogs aren't uncommon in capital intensive industries. For example when the 787 backlog fell to 546 planes or about a 3 year backlog Boeing responded by slashing production. ASML backlog in contrast looks to be just 2 years.

Two (2) years? What am I not seeing? Q4 19 - 8 shipped, orders either 49 or 41. The backlog line mean gradient is steeper than the shipped line one, meaning falling behind at increasing rate. Advanced node designs will become a battle for fab capacity even more than today.

 

[Ajay]

Backlogs aren't uncommon in capital intensive industries. For example when the 787 backlog fell to 546 planes or about a 3 year backlog Boeing responded by slashing production. ASML backlog in contrast looks to be just 2 years.

Four years, I think. Hmm, guess that's how Boeing got to be such a sh*tbag company. They practically owned the long haul market - now look where they are.

 

[LightningZ71]

Owned the long haul market?!?! There's this little hobby project in Europe called Airbus that might beg to differ. Their numbers for the A-330/340 are solid in several variants. They sold more passenger A-380s than Boeing built 747-8s. The A-350 looks to be able to replace the 777 in the market, and the new -8/-9 is struggling for customers as more and more are backing out, and the big three on the middle east hemorrhage orders.

The only reason that Boeing rulled the long haul market in the 80s and 90s was Airbus working to catch up after the success of the 767/777-200 in the market against less efficient A-340 models and early A-330s that had severe range shortcomings. Mc Donald Douglas severely missed their range targets with the MD-11, which left Boeing a captive market.

The ASML EUV machines are incredibly complicated, and uktra-precise machines. Building them is on the same or greater order of complexity as the composite and monocrystal fan blades in modern jet engines. Only a handful of actual facilities have the equipment to build them, the reject rate is high for supplies and final product, and final assembly for the customer is its own complicated and delicate process. This isn't something that you just "scale up" on a whim.

Add in a layer of security as they keys to making these machines and leading performance fan blades is absolutely considered no less than a state secret, and it gets to be an even larger headache.

 

[Ajay]

It's looking like that graph from an earlier post is erroneous or outdated.

TSMC Places Massive EUV Tools Order to Boost Capacity

When it comes to EUV installations, TSMC is ahead of the industry.

www.tomshardware.com

But money is not the only concern when it comes to EUV tools. ASML is the single company that produces and installs EUV scanners and it has a relatively limited production and installation capacity. After all the adjustments of its manufacturing process, the company believes it can shrink cycle time for a single machine to 20 weeks, which would result in per annum capacity of 45 to 50 systems.


Throughout three quarters this year, ASML has shipped 23 EUV scanners and intends to sell a little less than the 35 systems it originally planned for 2020. So far, ASML has shipped 83 commercial EUV tools (which includes NXE:3350B, NXE:3400B, and NXE:3400C machines sold from Q1 2015 to Q3 2020) to all of its customers.

 

[DisEnchantment]

It's looking like that graph from an earlier post is erroneous or outdated.

TSMC Places Massive EUV Tools Order to Boost Capacity

When it comes to EUV installations, TSMC is ahead of the industry.

www.tomshardware.com

Graph is from David Schor, he is reliable, but it is outdated. In my opinion, probably one of the very few reliable (and knowledgable) people when it comes to these things. Other things I heard now is that there is some politics involved with ASML being pressurized by foreign governments for priveleged access. When I find the article I will link here.

 

[Ajay]

Graph is from David Schor, he is reliable, but it is outdated. In my opinion, probably one of the very few reliable people when it comes to these things.

Schor is solid as a rock, but that graph if OFN.

 

[Doug S]

Sorry Doug, that's just a terrible analogy. ASML's book to bill ratio is simultaneously enviable and tragic. There huge sums of money being left on the table here. I don't know why they haven't expanded capacity - I haven't even looked it up yet. But it would be foolish for them to hamper production on the basis of exclusivity. Their products are born of real needs by real business partners - not persons who's vanity is matched by the size of their wallet. There has to be another reason.

I'm not talking about exclusivity (I guess using high end supercars was a bad analogy in retrospect, but there aren't a lot of industries that are similar)

Building the extra capacity needed to satisfy the orders more quickly will 1) cost more up front and 2) mean down the road they have long periods where they have excess capacity that's wasted.

So who makes up for all that extra money they have to put in up front, and the money they lose by having idle capacity down the road? Either their three customers do via higher prices, or their shareholders do by making less profit (in reality situations like this almost always end up being a combination of two taking the hit)

ASML could raise their prices to fund bringing additional capacity online, but even then having a full order book doesn't mean those orders will materialize unless they are paid for up front. It would not be shocking for Intel to push back a lot of its orders if their 7nm process takes 5 years to deliver like 10nm did. If ASML rushes to fill their orders and suddenly Intel tells them "we need to delay 15 delivers to 2025" and suddenly their order book is empty and they have a lot of excess production capacity.

 

[Ajay]

@Doug S EUVL isn't going to be disappearing anytime soon. Fabs need either more, or faster machines to get cycle times down. They need better uptimes, lower maintenance, etc.

 

[DisEnchantment]

https://twitter.com/x/status/1340294511218585601

Zen3+ on 6nm. Seems like we will see something else before we see Zen4. Or they might coexist.
Doing them both would make more sense from fab capacity standpoint, imo. Zen4 for the high end, and Zen3+ for the lower end.
Laptops SKUs are among the least profitable for AMD due to price pressure from ODMs, so it makes sense to stick with cheaper 6nm.
Question is, what architectural changes does Zen3+ have? Is it just 6nm port w/ DDR5? Or just AMD pressurised into a yearly cadence? Which I imagine resulting from all these Apple/ARM talk lately. But it is definitely a good thing.

Personally, I am looking forward to more aggressive designs from AMD, doing a single CCD for all means they optimize for the least common denominator. That notion of one CCD fits all is not helping vs the competition.
Redesign/Replacement of all that IFOP circuitry has to happen as well, that thing consumes almost half of the EPYC TDP.
For a 12 CCD Genoa CPU it could possibly burn in excess of 140 Watts, which is totally nuts if AMD is to keep the 280W TDP of their CPUs.

 

[moinmoin]

Question is, what architectural changes does Zen3+ have? Is it just 6nm port w/ DDR5?

Just a reminder that Zen+ didn't have architectural changes either, just changed node and microcode. So the same happening to a Zen3+ makes sense.

What do we know about Barcelo, Dragon Crest and Pollock?

 

[Zepp]

Pollock | 4.5W | Zen | Vega | 14nm

It seems strange to me for them to continue with Zen1 and 14nm for low power pollack. I understand the possible need to continue utilizing GloFo nodes but why not go Zen+/12nm? Is GloFo's 14nm just better suited to such low power SKUs than their 12nm?

 

[scineram]

It’s the same thing.

 

[Thunder 57]

Just a reminder that Zen+ didn't have architectural changes either, just changed node and microcode. So the same happening to a Zen3+ makes sense.

What do we know about Barcelo, Dragon Crest and Pollock?

And better cache latencies.

 

[soresu]

What do we know about Barcelo, Dragon Crest and Pollock?

Barcelo is embedded Cezanne maybe?

From the spec list I'd infer different market SKU's for Cezanne and Van Gogh.

 

[moinmoin]

And better cache latencies.

Not really. Consumer Ryzen 1000 were affected by a bug which was already fixed with Zen 1 Threadripper and Epyc. Fixing that didn't really need Zen+, but AMD kept it for that.

 

[Thunder 57]

Not really. Consumer Ryzen 1000 were affected by a bug which was already fixed with Zen 1 Threadripper and Epyc. Fixing that didn't really need Zen+, but AMD kept it for that.

Still worth noting since most of us plebs would have got Ryzen and not TR or EPYC. IIRC even Raven Ridge had the fix. But for many Zen+ provided that extra bit of performance.

 

[uzzi38]

Barcelo is embedded Cezanne maybe?

From the spec list I'd infer different market SKU's for Cezanne and Van Gogh.

Sounds to me like Barcelo is to Cezanne as Lucienne is to Renoir.

Barcelo is easier to justify though IMO because Rembrandt is rumoured to be rather large. Iirc Patrick Schur said 208mm^2. Selling Cezanne at the same time actually does make sense there.

 

[uzzi38]

Oh and as an aside, it does seem like Warhol is a straight Vermeer refresh akin to the Zen 2 XT skus.

I was hoping it would be AM5 but unfortunately not, seems they're focusing on clocks. Wondering what is possible though given it is possible - even if rare - to get Zen 3 chips clocking above 5GHz already.

For example:

https://twitter.com/x/status/1340906673137840128

 

[exquisitechar]

Sounds to me like Barcelo is to Cezanne as Lucienne is to Renoir.

Barcelo is easier to justify though IMO because Rembrandt is rumoured to be rather large. Iirc Patrick Schur said 208mm^2. Selling Cezanne at the same time actually does make sense there.

Exactly my thoughts.

Oh and as an aside, it does seem like Warhol is a straight Vermeer refresh akin to the Zen 2 XT skus.

I was hoping it would be AM5 but unfortunately not, seems they're focusing on clocks. Wondering what is possible though given it is possible - even if rare - to get Zen 3 chips clocking above 5GHz already.

For example:

https://twitter.com/x/status/1340906673137840128

It seems so, which is pretty disappointing, considering that it seems like we'll be waiting a while for Zen 4.

 

[xilli_fiberbit]

The CES is not too far now and AMD should give updated roadmap during this event

I don't uderstand what is the codename Trento for ? An EPYC between Rome and Genoa ?

https://twitter.com/x/status/1340443273643450368

 

[moinmoin]

The CES is not too far now and AMD should give updated roadmap during this event

I don't uderstand what is the codename Trento for ? An EPYC between Rome and Genoa ?

https://twitter.com/x/status/1340443273643450368

Sounds like Trento is the codename for the Zen 3 variant specific to the Frontier supercomputer. It has been mentioned before that Frontier doesn't use stock Epyc chips so isn't Milan.

 

[Gideon]

Sounds like Trento is the codename for the Zen 3 variant specific to the Frontier supercomputer. It has been mentioned before that Frontier doesn't use stock Epyc chips so isn't Milan.

i wonder what'll be the main difference, DDR5?