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!

 

[DrMrLordX]

That is purely anecdotal experience without hard numbers.

Like I said, take it there if you want clarification. This is a Zen4 discussion thread, not a "Steam survey results are legit!!!!" discussion thread.

 

[A///]

Like I said, take it there if you want clarification. This is a Zen4 discussion thread, not a "Steam survey results are legit!!!!" discussion thread.

I won't bother wasting my time. But you should also address the original person who took it off topic.

Page 2 - Discussion - Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

Page 2 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[DrMrLordX]

I won't bother wasting my time.

You might learn something, though.

edit: start here:

Why the Steam hardware survey is a joke for what people try to use it for.

1) Me, from May You really don't want to look at the Steam numbers for the simple fact that they don't really bear any relation to the actual market. I say this based on the fact that the 2 products that hard numbers have ever been released for, the HD5700 and HD5800 cards, shortly after release...

forums.anandtech.com

When people start using Steam survey numbers as authoritative proof of anything, you'll inevitably get negative reactions on these forums for reasons which you'll better understand if you read the above thread; furthermore, playing the role of contrarian threatens to take us even further off-topic. Any other off-topicness is best left to the mods.

 

[A///]

Fair enough. You might learn something, though.

Otherwise, I'll let the mods deal with off-topicness elsewhere in the thread where necessary.

If I were claiming Steam were representative of all gamers, then I'd be wrong, but I never made such a statement. With 90M active users, it's not a bad platform to garner some data from. If you've got a J Peddie report to share, feel free.

Otherwise it would be like me stating that because Amazon shows the RX 5700 as their #1 best seller on their own market, that must mean the RX 5700 is the ultimate graphics card with high in-use numbers due to it being #1. When in reality it pertains to Amazon's market only, and only in their .com store, whereas results are different on their other domains as they operate differently. The RX 580, for example, is the #1 seller on their UK site. GTX 1660 Ti on their Canadian site.

 

[Thunder 57]

Show a source for your claims.

Guess we see things differently. I'm not saying deep dives don't count towards the enthusiast, but how many deep dives are published yearly versus regular reviews a few pages long with some benchmarks of a hardware device? Some of the companies I've worked for have offered deep dives into background ops, but those don't necessarily attract upper echelon enthusiasts. The yearly conferences they hold and sponsor do, however, satisfy those wants or needs.

My issue with the poster was his repeated claims which were corrected, given into, and then another claim was made without much basis. It would have been simple for him or her to link to last month's Steam hardware survey. While Steam is just a small segment in terms of online gaming and as an e-tailer themselves, they offer some interesting data for October 2019.

Steam Hardware & Software Survey

Right off the bat we get these results.

75% of users utilize NVidia graphics cards.
Of all graphics cards, 73.17% offer DX12 on W10
80.54% of users have Intel processors
Of all the processors reported, 51.21% use 4 cores (labeled as CPUs), less than 24% use other, which is for more than 4 cores.

You can see the breakdown data:

Steam Hardware & Software Survey

GFX card breakdown:

Steam Hardware & Software Survey

One could argue that a 1060 counts as enthusiast. I really don't. That's entry level decent graphics. You could argue on price, because of NVidia's sly shift towards a higher price that the 1060 is an enthusiast card now. I could also argue the opposite on the CPU front, that the 10980XE that was just released doesn't qualify for HEDT based on price because Intel slashed 50% over a generational shift which is really a 2 year old processor with boosted clocks, and that it should only be compared with the 3950X because of the pricing bracket.

The above is an argument people have been using, not necessarily on AT, over the past week. The reality is if Intel slashed another $800 off, are we supposed to compare it to a sub $200 processor? Or compare it for what it is? You wouldn't compare the 2080ti to the rumored 5500 XT, now, would you? No.

That is purely anecdotal experience without hard numbers.

Querying against the site shows threads from 2011 where people criticized it based on data prior to 2011. I'm sorry, if the site has a bias against it that's fine, but suggesting it has junk data because Anand's users have some decade old grudge against it isn't a valid reason to exclude it. If it isn't allowed, then I'd like some hard data showing individual units of video cards being sold. It would only be fair, wouldn't it?

Provide a better source if you have one. Perhaps a breakdown of sales data from 2010 to 2019. Major GPU makers AMD/ATI and NVidia, the model names of cards, the amount sold each quarter or year, and some data that show capture data of equipment still in use.

Why so confrontational? And for someone who demands a source from everyone who posts, you have yet to show any that I have seen.

To my knowledge, pure gamers make up a small portion of DIY sales of the already small DIY sector. No idea on OEMS but Dell's been selling Ryzens in their gaming computers. No idea how long,t hough.

Where's your source for that? Others can play that game, too.

I won't bother wasting my time. But you should also address the original person who took it off topic.

Page 2 - Discussion - Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

Page 2 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

"I won't bother wasting my time". Translation: "I might actually be wrong and don't want to accept that".

 

[A///]

Why so confrontational? And for someone who demands a source from everyone who posts, you have yet to show any that I have seen.

I'm not being confrontational. I posted a link to the Steam survey. That represents Steam users. The problem is, Anand's forum doesn't like it because it isn't representative of all global vendor sales. That data is accessible if you subscribe to bodies that track such sales data, and there are quite a few. They're like Nielsen for hardware. If Epic ran a survey, you could find a flaw with the data in that too. The current flaw with Steam's data is people claim, have not proven, but claim it submits improper data. They do claim there is not total number of respondents, which is true, but few surveys of this nature give out an exact number or ballpark figure.

I can get why people disliked it 10 years ago or even a few years ago when AMD cards were on a roll either due to better performance than Nvidia and then mining, respectively. If you know a bunch of 200 series cards are outselling NVIdia but barely any show up on Steam, you would easily presume the data is flawed and not subjective to real world data, because it isn't. However, Steam is a storefront and they track their own internal data, not external data like many would expect them to. Though I can't imagine why. Steam's usage was a fraction a decade ago compared to now, or even a few short years ago. If cards were being bought from AMD's lineup a few years ago in great numbers and used mainly for mining, I can't imagine why you would expect to see those numbers on a gaming storefront.

J Peddie trackers global shipments from like 28 countries. That's about as close to a real source as you can get without hacking into each vendor and getting their sales data. If a J Peddie report suggests 5,000 units of NVidia and AMD were shipped this quarter for 2019, of which 3800 were NVidia and the rest AMD, and of which knocked off individual models, then you could get a really good idea of what's being sold. Though I can't recall if J Peddie tracks individual unit sales or mixes it with prebuilt gaming PCs that have those installed, such as an Alienware, for example.

 

[DrMrLordX]

With 90M active users, it's not a bad platform to garner some data from.

Yes it is. It always has been.

Dec. Steam Hardware Survey - AMD maintains DX11 dominance at near 80% of the market.

AMD - 77.33% nvidia - 22.51 Here is a link to the past month's survey: http://forums.anandtech.com/showthread.php?t=2124878&highlight=steam Looks like again minimal adjustment to the DX11 marketshare numbers. It's interesting how much getting to market first has cemented AMD's holding the...

forums.anandtech.com

Best quote is from Borealis7:

we've done this to death. the example was shown that AMD can outsell nV 2:1 and still lose market share. props to whoever made that one, i cant remember who it was though...

absolute numbers are key here. steam is limited survey in a limited community inside a limited market. kthxbye.

As you can see from later in the thread, mods tried for a time to keep things civil in Steam survey threads, but I think they eventually gave up. Nobody makes those posts anymore, and discussing Steam survey results around here usually raises hackles. Steam data is too easy to manipulate to push a false narrative.

 

[A///]

Yes it is. It always has been.

As you can see from later in the thread, mods tried for a time to keep things civil in Steam survey threads, but I think they eventually gave up. Nobody makes those posts anymore, and discussing Steam survey results around here usually raises hackles. Steam data is too easy to manipulate to push a false narrative.

I posted a link to the Steam survey. That represents Steam users. The problem is, Anand's forum doesn't like it because it isn't representative of all global vendor sales. That data is accessible if you subscribe to bodies that track such sales data, and there are quite a few. They're like Nielsen for hardware. If Epic ran a survey, you could find a flaw with the data in that too. The current flaw with Steam's data is people claim, have not proven, but claim it submits improper data. They do claim there is not total number of respondents, which is true, but few surveys of this nature give out an exact number or ballpark figure.

Yeah, I see it. I saw it when it was posted back then, too. I never said Steam was representative of all sales figures. I even said this which you chose to ignore.

If I were claiming Steam were representative of all gamers, then I'd be wrong, but I never made such a statement. With 90M active users, it's not a bad platform to garner some data from. If you've got a J Peddie report to share, feel free.

And in the post where I mentioned Steam, I even stated this very clearly, which people ignored.

My issue with the poster was his repeated claims which were corrected, given into, and then another claim was made without much basis. It would have been simple for him or her to link to last month's Steam hardware survey. While Steam is just a small segment in terms of online gaming and as an e-tailer themselves, they offer some interesting data for October 2019.

It was never my intention to suggest Steam captured all sales data. Because I never made such a statement. I said that Steam is:

A) A small segment of online gamers
B) Offer some data

The argument you're providing is in regards to hard sales data. I know AMD outsold NVidia for quite a few years more than just 2:1, but I understand why you're being conservative with your ratio. And that data simply isn't shown on Steam. And I get that. I never suggested it otherwise. So I don't understand why I'm being ganged up on when I was succinct with what I said from the beginning, and never intentionally implied Steam was representative of all sales data. Is that clear now?

 

[DrMrLordX]

Yeah, I see it. I saw it when it was posted back then, too. I never said Steam was representative of all sales figures. I even said this which you chose to ignore.

No, it's that it isn't representative of anything useful. You needn't be as specific as "all sales figures". It's generally inapplicable to anything.

 

[soresu]

I'm just going to address the pink elephant in the room - Zen4 and die stacking.

Whether it will be just memory, or memory and logic dies - I think something wicked this way comes....

 

[A///]

Whatever. I'm done with this. You do you, I'll do me. We're both right, or wrong, doesn't matter.

 

[Markfw]

You might learn something, though.

edit: start here:

Why the Steam hardware survey is a joke for what people try to use it for.

1) Me, from May You really don't want to look at the Steam numbers for the simple fact that they don't really bear any relation to the actual market. I say this based on the fact that the 2 products that hard numbers have ever been released for, the HD5700 and HD5800 cards, shortly after release...

forums.anandtech.com

When people start using Steam survey numbers as authoritative proof of anything, you'll inevitably get negative reactions on these forums for reasons which you'll better understand if you read the above thread; furthermore, playing the role of contrarian threatens to take us even further off-topic. Any other off-topicness is best left to the mods.

I wasn't even talking about DRM (his first reply to my statement that steam numbers are no good)

I was talking about % Intel vs AMD and the other numbers. The steam survey does not count
1) workstation users
2) accounts and the live
3) People like me that use them for DC, and for playing games USING a CD
4) moms that don't use a phone or tablet
And data centers, business users and countless other computer users I have forgotten to mention.

 

[Thunder 57]

I wasn't even talking about DRM (his first reply to my statement that steam numbers are no good)

I was talking about % Intel vs AMD and the other numbers. The steam survey does not count
1) workstation users
2) accounts and the live
3) People like me that use them for DC, and for playing games USING a CD
4) moms that don't use a phone or tablet
And data centers, business users and countless other computer users I have forgotten to mention.

And there is other data to back that up. Links:

AMD Decimates Intel With 82% Share At Germany's Largest Retailer, Mindfactory.de

AMD has been breaking records left, right and center and if you thought this the highest the company can go then I have news for you. The company just achieved a new record for its sales at Germany's largest retailer: Mindfactory.de. Clawing even more market share from Intel, the company clocked...

wccftech.com

60% of European Users Prefer AMD CPUs According To EHA Survey

An independent survey by the European Hardware Association found that users who prefer AMD CPUs now make up more than 60% of all respondents. This is a number that is up from 40% in 2018 and can be attributed to AMD's highly successful Zen architecture as well as their process leadership on the...

wccftech.com

In the US, Amazon as of now has the top 10 best selling CPU's being AMD. These are updated hourly, so that may change.

 

[DrMrLordX]

I'm just going to address the pink elephant in the room - Zen4 and die stacking.

If AMD wants to go down that route, they'll have to deal with cooling issues. Zen2 already has some of those now wrt hotspots. Should be interesting to see how (of if) they address the problem with Zen3. The combination of hotspots and die stacking could be technically difficult (and expensive) to overcome. You could kiss overclocking goodbye.

 

[H T C]

If AMD wants to go down that route, they'll have to deal with cooling issues. Zen2 already has some of those now wrt hotspots. Should be interesting to see how (of if) they address the problem with Zen3. The combination of hotspots and die stacking could be technically difficult (and expensive) to overcome. You could kiss overclocking goodbye.

It's why i think Forrest Norrod meant with a "clock reduction" back @ his Oil and Gas Presentation, a while back. This being the case, in order for new products to be upgrades over the "old ones", a significant rise in IPC is needed to overcome the higher clock deficit.

For a scenario where a 4 GHz Zen 3 CPU outperforms a 4.5 GHz Zen 2 CPU , it would require 12.5% IPC increase just from the clock deficit alone + whatever else AMD comes up with from other changes: this suggests @ least 20% IPC increase for THIS specific scenario. AMD may not have to reduce clocks this much or it could be forced to reduce clocks even further: we just don't know, @ this point.

 

[DrMrLordX]

It's why i think Forrest Norrod meant with a "clock reduction" back @ his Oil and Gas Presentation, a while back. This being the case, in order for new products to be upgrades over the "old ones", a significant rise in IPC is needed to overcome the higher clock deficit.

For a scenario where a 4 GHz Zen 3 CPU outperforms a 4.5 GHz Zen 2 CPU , it would require 12.5% IPC increase just from the clock deficit alone + whatever else AMD comes up with from other changes: this suggests @ least 20% IPC increase for THIS specific scenario. AMD may not have to reduce clocks this much or it could be forced to reduce clocks even further: we just don't know, @ this point.

Not to make this too much of a Zen3 thread (we have one of those), but TSMC has said that 7nm+ should provide +10% performance at isopower. And yes that's assuming "isodesign" as well, take that with a grain of salt. So I'm guessing that AMD hopes to get a few more clocks out of Zen3 (rather than less) without jacking TDPs through the roof. But you know, increased density so . . . they might have to give up that performance just to fight even worse hotspotting. And then let's assume Zen4 goes to 5nm. I haven't seen any information released from TSMC on how they expect performance @ isopower to change moving from 7nm+ to 5nm.

Anyway, jist of it is that I could see Zen3 being a clockspeed sidegrade at a lower TDP to keep temps in check. Maybe they could, you know, claim higher clocks by letting the boost algorithm hit higher MT speeds than Zen2 without significantly (or at all) increasing max boost clocks past 4.7 GHz which is the peak for Matisse at the moment. Moving on to Zen4 with hypothetical die stacking . . . that's where things could get ugly. If they tried to keep clocks stable from Zen2->Zen3->Zen4, they'd have to see essentially greater reductions in power @ isoperformance than increases in density to reduce hotspotting just for the chance to make die stacking actually a workable solution. Just moving from 7nm to 7nm+, you can get 15% power reduction @ isoperformance with 20% increase in density, so even if you do hold clocks stable, hotspotting could actually get worse. Obviously, AMD will have to drop clocks/voltage eventually. Unless they finally start using new, novel packaging techniques to do a better job of spreading out heat from die hotspots.

 

[moinmoin]

Not to make this too much of a Zen3 thread (we have one of those), but TSMC has said that 7nm+ should provide +10% performance at isopower. And yes that's assuming "isodesign" as well, take that with a grain of salt. So I'm guessing that AMD hopes to get a few more clocks out of Zen3 (rather than less) without jacking TDPs through the roof. But you know, increased density so . . . they might have to give up that performance just to fight even worse hotspotting. And then let's assume Zen4 goes to 5nm. I haven't seen any information released from TSMC on how they expect performance @ isopower to change moving from 7nm+ to 5nm.

Anyway, jist of it is that I could see Zen3 being a clockspeed sidegrade at a lower TDP to keep temps in check. Maybe they could, you know, claim higher clocks by letting the boost algorithm hit higher MT speeds than Zen2 without significantly (or at all) increasing max boost clocks past 4.7 GHz which is the peak for Matisse at the moment. Moving on to Zen4 with hypothetical die stacking . . . that's where things could get ugly. If they tried to keep clocks stable from Zen2->Zen3->Zen4, they'd have to see essentially greater reductions in power @ isoperformance than increases in density to reduce hotspotting just for the chance to make die stacking actually a workable solution. Just moving from 7nm to 7nm+, you can get 15% power reduction @ isoperformance with 20% increase in density, so even if you do hold clocks stable, hotspotting could actually get worse. Obviously, AMD will have to drop clocks/voltage eventually. Unless they finally start using new, novel packaging techniques to do a better job of spreading out heat from die hotspots.

I feel the placement of hotspots will be a more and more integral part of silicon design (there were already a couple patents that mention cooling as part of the silicon and 2.5D/3D design). It's already affecting the boost algorithm, since it's considering package power consumption, temperature and current. I can imagine the boost algorithm going to also respect ways/outlets to cool hotspots if it's not already doing that, preventing hotspots to accumulate in close areas and affecting each others.
Going from there I'd think using denser nodes as well as chip stacking are a challenge but the framework to handle that may be already there.

 

[Azix]

I actually built my current system (CPU and motherboard at least) when ddr3 was about to go out. I hope I don't upgrade right before ddr5 becomes mainstream.

 

[DisEnchantment]

I think DDR 5 adoption rate could be faster than that of DDR4.

 

[ksec]

My Take on Zen Roadmap.

Zen 3 will have new L3 cache structure that shares more between Core compared to current CCX design. ˜15+% IPC improvement, 7nm EUV. I am expecting an even smaller die size as cost optimisation .

Zen 4 will be 5nm shrink and I/O Shrink. Possibly Improved Infinity Fabric, DDR5, PCI-E 5, Smaller I/O Die, may be FD-SOI from GF, or completely move to TSMC 6nm for I/O Die. Along with some minor tweak on IPC and ClockSpeed improvement. The I/O improvement alone would be extremely welcome for HPC customers. Not to mention the power reduction from Node Shrink.

Zen 5 will see 5nm+, may be 16 Core per die, or minor IPC improvement.

But the trend is AMD will basically be doing IPC, ClockSpeed, Node, I/O cost optimisation in tick tock fashion like old Intel.

 

[DrMrLordX]

Zen 5 will see 5nm+

There's a TSMC 5nm+?

 

[uzzi38]

There's a TSMC 5nm+?

N5P is the actual term for it, and yeah, it exists.

 

[uzzi38]

My Take on Zen Roadmap.

Zen 3 will have new L3 cache structure that shares more between Core compared to current CCX design. ˜15+% IPC improvement, 7nm EUV. I am expecting an even smaller die size as cost optimisation .

Zen 4 will be 5nm shrink and I/O Shrink. Possibly Improved Infinity Fabric, DDR5, PCI-E 5, Smaller I/O Die, may be FD-SOI from GF, or completely move to TSMC 6nm for I/O Die. Along with some minor tweak on IPC and ClockSpeed improvement. The I/O improvement alone would be extremely welcome for HPC customers. Not to mention the power reduction from Node Shrink.

Zen 5 will see 5nm+, may be 16 Core per die, or minor IPC improvement.

But the trend is AMD will basically be doing IPC, ClockSpeed, Node, I/O cost optimisation in tick tock fashion like old Intel.

Zen 3: Yep. Agreed there. Other stuff as well is almost certain, as pretty much every new arch from here on out will likely increase cache sizes (have to to lower thermal density, will be a major problem in the future)

Zen 4: No point in shrinking I/O afaik, but could be wrong there. Definitely not to 6nm though. It'll likely use GF's 12LPP+ instead (or whetever it's called, their 12nm+ that's slated for late 2020). Agreed on the smaller IPC bump and clocks. Power reduction isn't much, but density improvements is the real bonus from 5nm over 7nm.

Zen 5: Doubt cores per die increasing. Like I said before, thermal density is a real issue. This is where the big arch jump will happen again, so big IPC improvement. As for how, heck if I know.

 

[DrMrLordX]

N5P is the actual term for it, and yeah, it exists.

Hmmm interesting. Thanks for the tip on the name:

https://www.eetasia.com/news/article/TSMC-Adds-a-N5P-process

So N5P uses the same design rules as 5nm, meaning AMD could go from 7nm+ -> 5nm -> N5P/"5nm+" without skipping a beat. It looks like, in terms of design rules, 6nm is a dead end.

 

[kapulek]

https://twitter.com/x/status/1201939134928977920