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!

 

[uzzi38]

I think zen4 is a two-step launch, that isn't focused on IPC:

2022: New platform and new processors. The processors seems like "tick" shrinking of zen3 with some light chances with AVX512 and L2 cache, and we might even get vcache models at launch or maybe early 2023
2023: They launch zen4c on 4N with up to 32 cores on desktop, also with vcache models, and I think this is the main reason for the 170W TDP
2024: "tock" zen5

I wouldn't count on Zen4C coming to desktop. Not ruling it out entirely, but I'd say 95% chance it doesn't happen.

 

[JasonLD]

Process is 2x more efficient, so perf/watt improvement at same clock should be close to 100% if we discard IPC difference.

Process itself allows either 20% power reduction at same frequency or 15% more frequency at same power vs 7nm.

 

[Vope45]

You're absolutely right, even though you didn't intend to be so. All this is extremely disappointing for Intel, especially where it hurs Intel's bottom line the most. But you'll see over the course of the next 18 months 🙂

I really don't care about intel. 8-10% ipc increase seems a tad too little for me. Intel with its vast r/d will eventually bring out the big guns and when that happens amd cannot catch up.

 

[inf64]

I really don't care about intel. 8-10% ipc increase seems a tad too little for me. Intel with its vast r/d will eventually bring out the big guns and when that happens amd cannot catch up.

We keep hearing about those big guns, but somehow they are being postponed and slipping away. If I were intel, I'd be worried as AMD is growing pretty fast and has a superior products across the board.

 

[Leo9]

I really don't care about intel. 8-10% ipc increase seems a tad too little for me. Intel with its vast r/d will eventually bring out the big guns and when that happens amd cannot catch up.

Intel R&D is now only about 4.3x AMD, and AMD is fabless. Just 3 years ago Intel budget was almost 10x. The difference isn't that big anymore and it's only getting smaller.

Advanced Micro Devices Research and Development Expense (Quarterly) Analysis | YCharts

In depth view into Advanced Micro Devices Research and Development Expense (Quarterly) including historical data from 1973, charts and stats.

ycharts.com

Intel Research and Development Expense (Quarterly) Analysis | YCharts

In depth view into Intel Research and Development Expense (Quarterly) including historical data from 1972, charts and stats.

ycharts.com

 

[Det0x]

Some napkin math from me

So if we assume that Cinebench is among the worst case (low end of the scale) in terms of "IPC" increase from Zen3 to Zen4, ref this picture: (Cinebench r20 ST is + 13% while geomean is 19%)
View attachment 62850

Zen3 fmax 5050mhz VS Zen4 fmax 5850mhz = ~15% higher peak clockspeeds (sustain clocks probably see an even bigger increase)

ST Performance base (Zen3) x 1.10 IPC x increased clock speed by 15% = ~26% higher ST performance in "worst case apps" like Cinebnench R23

In games, we can probably/maybe quickly get "+20% IPC" or more considering doubling the L2 cache that Cinebench does not care about.

Gaming performance (Zen3) x 1.2 IPC x increased clock speed by 15% = ~38% higher performance in "games". (or more thanks to better sustained clocks)

On top of this we get the benefit from going from DDR4 to DDR5 with hopefully much higher infinity fabric clockspeeds, which has limited Zen3 performance in games (latency) from the get-go.

All in all, not bad i would say

More singlethread napkin performance math

ZEN 3:
CB: +13% IPC.
Average: +19% IPC.
Clock Speed: +4.5%.

Average ST Zen3 performance increase over Zen2 =
1 * 1.19 * 1.045 = a ~24% ST performance uplift

ZEN 4:
CB: +8-10% IPC.
Average: +11.5-14.5% IPC.
Clock Speed: +12.5-19%.

Average ST Zen4 performance increase over Zen3 =
1 *1.115 * 1.125 = a ~25% ST performance uplift on the low-end of the scale
1 *1.145 * 1.19 = a ~32% ST performance uplift on the high-end of the scale

Multithread performance increase is likely higher, hence the "greater than + 35% overall performance"

 

[Vope45]

We keep hearing about those big guns, but somehow they are being postponed and slipping away. If I were intel, I'd be worried as AMD is growing pretty fast and has a superior products across the board.

Well I guess you can't have both high ipc and high frequency. I'd rather have a bit of both.

Intel R&D is now only about 4.3x AMD, and AMD is fabless. Just 3 years ago Intel budget was almost 10x. The difference isn't that big anymore and it's only getting smaller.

Advanced Micro Devices Research and Development Expense (Quarterly) Analysis | YCharts

In depth view into Advanced Micro Devices Research and Development Expense (Quarterly) including historical data from 1973, charts and stats.

ycharts.com

Intel Research and Development Expense (Quarterly) Analysis | YCharts

In depth view into Intel Research and Development Expense (Quarterly) including historical data from 1972, charts and stats.

ycharts.com

Does that include xilinx r/d as well?

 

[Vope45]

https://www.reddit.com/r/GamingLeaksAndRumours/comments/jpmejf

This post over 2 year ago claimed most of the performance increase is because of infinity fabric 2.0. Claimed small ipc increase as well.

 

[uzzi38]

https://www.reddit.com/r/GamingLeaksAndRumours/comments/jpmejf

This post over 2 year ago claimed most of the performance increase is because of infinity fabric 2.0. Claimed small ipc increase as well.

It was nonesense from the beginning.

 

[lobz]

Well, I think that a lot of people here are in for a DisEnchantment (pun totally intended, though not just for him).
- has any of you seen or tested a Zen4 product?
- do you still remember the gaming king (Intel's words!) of Rocket Lake with its leaked and then later anmounced, incredible GeekBench ST IPC increase?
- do you still remember how that played out?
- do you still remember what overall IPC Intel has been shouting in the air before RKL launched?
- do you still remember how that product ended up performing?
- does any of you want AMD to unnecessarily hype ANY product again after poor Volta, I mean Vega?
- has any of you thought about the relation between raw IPC at a given clockspeed and a new uarch that could maybe, actually scale clockspeed and scale performance really well WITH clockspeed? Because that would be a territory that AMD hasn't threaded ( ) upon in a decade or so
- can any of you wait for the actual PRODUCT to launch before you make conclusions you'll be majorly embarrassed about in 3-6 monts?

Just askin'

 

[Markfw]

https://www.reddit.com/r/GamingLeaksAndRumours/comments/jpmejf

This post over 2 year ago claimed most of the performance increase is because of infinity fabric 2.0. Claimed small ipc increase as well.

A post that is over 2 years old is crap. And your post about Intels big guns is crap. They are in trouble, since they have almost no competing products right now, especially in the server area. And saying AMD will never catch up ? Dream on.

Do you work for Intel marketing or something ? As delusional as they are ?

 

[lobz]

I really don't care about intel. 8-10% ipc increase seems a tad too little for me. Intel with its vast r/d will eventually bring out the big guns and when that happens amd cannot catch up.

In what relation? I mean if actual IPC (but then again, at which clockspeed exactly?!) is your actual fetish, I can totally understand it being too little for you. Not sure about its general relevance though.
As for Intel's big guns and AMD not being able to catch up? Have you been sleeping since 2017?

 

[Vope45]

A post that is over 2 years old is crap. And your post about Intels big guns is crap. They are in trouble, since they have almost no competing products right now, especially in the server area. And saying AMD will never catch up ? Dream on.

Do you work for Intel marketing or something ? As delusional as they are ?

As much as I like amd. Zen's success is largely due to intel 10nm delays and intel culture. As there has been no report of of any drama plus intel's fabs getting on track, it is very possible for intel to win the long race. Normally it takes 4-5 years to realize one's vison, that coincides with lunar lake/ royal cove leaks thus far.

Not that I'm saying amd is doom, it's that intel's current delays are not the delays of the past.

 

[Abwx]

Process itself allows either 20% power reduction at same frequency or 15% more frequency at same power vs 7nm.

What is the use to argue without even knowing what you re talking about..?..

AMD Zen 4 in 5 nm: Genoa bringt 96, Bergamo mit Zen 4c sogar 128 Kerne

Nach Instinct MI200 hatte AMD zum Event am Abend noch eine Überraschung im Gepäck: Infos zu Zen 4 für Server in Form von Genoa und Bergamo.

www.computerbase.de

 

[uzzi38]

As much as I like amd. Zen's success is largely due to intel 10nm delays and intel culture. As there has been no report of of any drama plus intel's fabs getting on track, it is very possible for intel to win the long race. Normally it takes 4-5 years to realize one's vison, that coincides with lunar lake/ royal cove leaks thus far.

Not that I'm saying amd is doom, it's that intel's current delays are not the delays of the past.

Zen's success I agree with. CNL-S would been way better than both Zen and Zen+.

Zen 2 and on? No way. Zen 2 would have been extremely competitive against even a really good 5GHz boosting ICL-S (which was still supposed to be 8 cores). On the server and mobile roadmaps AMD would have had even stronger positions than the desktop.

Alder Lake is the first really competitive part from Intel and I'm thankful for it. But lets not pretend that process is the only reason why AMD were successful before it.

Guessing anything 2025/2026 in 2022 is setting yourself up for failure. We can't even reliably put a pin on where MTL is landing, what makes you think saying anything about 2025+ is even remotely sensible?

EDIT: Also, "Intel's current delays are not the delays of the past"? CCG would disagree and say they're on track (and they generally are doing very well, which is a good thing). DCG would laugh awkwardly and try and cover up both ICL-SP and SPR.

 

[moinmoin]

As much as I like amd. Zen's success is largely due to intel 10nm delays and intel culture. As there has been no report of of any drama plus intel's fabs getting on track, it is very possible for intel to win the long race. Normally it takes 4-5 years to realize one's vison, that coincides with lunar lake/ royal cove leaks thus far.

Not that I'm saying amd is doom, it's that intel's current delays are not the delays of the past.

Intel is seriously behind in performance per watt. Pat himself admitted that and said mid last year (so not factoring in delays since then):

We’re in a clear path to process performance-per-watt parity in [2024]

Intel Promises ‘Performance Leadership’ By 2025 With New Roadmap | CRN

Intel is promising a return to "process performance leadership" by 2025 with a new roadmap that is shaking up node naming conventions.

www.crn.com

 

[Karnak]

What is the use to argue without even knowing what you re talking about..?..

Do you know what you're talking about? Or you know what AMD did use to make that statement?

Pretty sure it'll be something like a 64C Milan EPYC vs. an upcoming 96C Genoa EPYC. But without knowing it's pointless. ofc it's not the process itself. For that we know everything directly from TSMC already.

 

[Tuna-Fish]

This post over 2 year ago claimed most of the performance increase is because of infinity fabric 2.0. Claimed small ipc increase as well.

That post is fanfic, and everyone who knows anything about tech should have noticed it on day one.

Specifically:

Sorry.I have zero info on PCI-E 5.0. I wouldn't necessary write it off, but I wouldn't expect it either. Will be determined later in development.

Two years ago, all the decisions about PCI-E 5.0 were definitely set in stone.

 

[uzzi38]

That post is fanfic, and everyone who knows anything about tech should have noticed it on day one.

Specifically:



Two years ago, all the decisions about PCI-E 5.0 were definitely set in stone.

They also memory validation, but that's not actually done early in the product lifetime.

Actual validation takes place nearer the end. You set targets you expect the IP to be capable of in the beginning, but actual validation can only happen when you have silicon.

 

[uzzi38]

Btw, I don't think it was mentionned here, but in the Q&A AMD clarified that both N4 and N3 Zen 5 based products will launch by the end of 2024. Also: "You'll see a diversity in nodes for most products going forwards".

External chiplets with UCIe is looking like a 2025 deal at best, really needs a second generation standard.

Just a couple of interesting snippets that have been overlooked.

 

[turtile]

I really don't care about intel. 8-10% ipc increase seems a tad too little for me. Intel with its vast r/d will eventually bring out the big guns and when that happens amd cannot catch up.

All that matters for AMD is the efficiency of the core for servers. AMD just claimed that it's greater than 25%. That is a significant jump. I figured IPC would be higher too because I didn't think 5nm could hit the massive clock speed increases. In addition, AMD now has 50% more cores per socket and a larger power budget so that it can get around 70% to 90% more performance per socket.

It seems like AMD decided to forego greater IPC increases and made a design to push the clocks so that they can keep the chiplet smaller on the expensive 5nm node. They basically used fewer transistors to get a bigger efficiency increase.

 

[naad]

Okay, so more info out at FAD.

As a server first uarch since its inception zen4 does its job, large perf/watt improvements and significant v/f curve improvements, IPC was measured with specint/fp so around 10% is a good estimate since no large changes are happening on the execution unit/OoO parts of the core, I still stand that memory bound, both latency and throughput applications will see closer to 20% than to 10% gain

Desktop against Raptor will look pretty close, Intel will definitely have the advantage in the mid and low range due to E-core spam, but I doubt AMD is worried about that too much considering their projected revenue growth.

All in all I'm satisfied, but Phoenix APU looks way more juicier than anything else, I have doubts even Meteorlake will put up much of fight.

 

[lobz]

As much as I like amd. Zen's success is largely due to intel 10nm delays and intel culture. As there has been no report of of any drama plus intel's fabs getting on track, it is very possible for intel to win the long race.

yeah sure... just as much as the winning team's success after a series of seasons are because of the losing team's failures... super useful summary

but more seriously: are you kidding? ever since pixel pat has stepped up to be the face, public business wise he's been doing THE EXACT SAME thing quarter over quarter when it comes to culture and drama as his predecessors in the past 6 years. Everything is always on track in every Q call, then things get very quietly slipped farther and farther away, more or less publicly. The only thing pixel pat has brought to the table so far is the even more blunt, baseless and senseless bashing of any of his competition, which he seems to take way too personally (lifestyle company& rearview mirror and close combat in a phone booth? tf? 🤦🤦🤦). Because logically you can be sure that he had absolutely nothing to do with either Alder Lake or Raptor lake, and I mean literally nothing whatsoever.

Care to reiterate?

 

[lobz]

All that matters for AMD is the efficiency of the core for servers. AMD just claimed that it's greater than 25%. That is a significant jump. I figured IPC would be higher too because I didn't think 5nm could hit the massive clock speed increases. In addition, AMD now has 50% more cores per socket and a larger power budget so that it can get around 70% to 90% more performance per socket.

It seems like AMD decided to forego greater IPC increases and made a design to push the clocks so that they can keep the chiplet smaller on the expensive 5nm node. They basically used fewer transistors to get a bigger efficiency increase.

to make things worse, the core count is over 50% and the perf is even more than your estimates, but let's see if intel can finally launch SPR in the first place to make a verdict.

 

[RnR_au]

Do you know what you're talking about? Or you know what AMD did use to make that statement?

Hmm idk... I think that slide is just talking about the silicon.

N5
* +15% performance over N7, or 30% power saving
* 1.84x density over N7

N5p
* +7% performance or 15% power saving over N5

So in total +22% perf or 45% power saving.

And we know that AMD has a special N5 process so they may be getting a little better than N5p in all aspects; power, perf and density - at least enough to get close to their whole numbers on the slide.

The flaw in the slide is that it doesn't mention 'or' between the perf improvements and the power saving. But yeah marketing