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

Vattila · Oct 6, 2019

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!

Timorous · May 25, 2022

deasd said:
IPC is an ambiguous claim I never take that too seriously, if you see the Zen3vsZen2 perf summary that was released by AMD in 2020 you can see the R15 is boosted up to ~18% IPC while other items are far from 18%, and AMD claimed Zen3 is ~19% IPC uplift. If you take official '>15% ST' seriously then you fall into another trap whatever you trying to deprecate or hype.

You can be more specific. In that Zen 2 vs Zen 3 slide CB R23 1T was + 13% IPC. The > 15% for Zen 4 performance over a 5950X is also in CB R23 1T.

I have a feeling games will exceed this delta due to the L2 cache increase which does not impact CB performance at all but time will tell.

inf64 · May 25, 2022

So just to see if I get this correctly, a lot of people expect that Zen4 is a Zen3, shrank to 5nm with "only" 1MB L2 addition (2x) and "bolted" on AVX512? AMD had two years to just design a node shrank chiplet with bolted-on AVX512 and a slight L2 cache bump (and maybe a bit longer pipeline)? All that time invested and a whole node upgrade to create Zen3++ (basically)? Hmm ok.

edit: not to mention that core chiplet size is barely reduced Vs Zen3 (72 vs 80mm^2). I know that we were not going to see 0.5x reduction going to 5nm as only sram can scale that well, but that's still a huge chiplet on 5nm. Even if a lot of space was "wasted" on AVX512. 2x L/S pathways to accommodate AVX and 2x L2, it's incredible they didn't do much (if any) other improvements to the core.

Timmah! · May 25, 2022

deasd said:
Some eagleeye found that two core dies are tighter than before:

https://twitter.com/x/status/1529297282125791233

seems the new chiplets got the same treatment as rectified ceramic tiles, lol

mikk · May 25, 2022

Hitman928 said:
I've said it before as have others, MLID and these other 'leakers' are not reliable sources. Of course, next gen we'll start the cycle again and people will be posting their stuff like it is legitimate and maybe they'll get lucky and guess a few things right, but discussing these rumors without a huge amount of pessimism is just a waste of time, IMO.

MLID never was an AMD leaker to me because I don't think he has real sources (this is my impression), I was surprised he posted such IPC claims like he was 100% sure. He never can be trusted again anything AMD related and even when it comes to Intel he is done for many people, no matter if his Intel sources are real or not. Imho when it comes to such leaks they are much less accurate on AMD side, there is too much hype involved. Another good example are the Cinebench MT scores from this AMD oriented channel, like they were confirmed from AMD themselves. And for sure the next round with Zen 5 we will see more big claims and expectations from the same people.

DrMrLordX · May 25, 2022

Abwx said:
38% better perf for the 6C
33% for the 8C
32% for the 12C
45% for the 16C.

The huge gap for the 16C is due to TDP increased, at same power the improvement should be closer to the other SKUs.

Interesting! Of course those numbers aren't exactly verified yet, are they? Still it seems that most of Raphael's improvements come form higher sustained MT clocks than anything else.

uzzi38 said:
Maybe you should just wait for more useful universal tests to be performed, like SPEC and Geekbench?

Please God no. This is what we should ask for from a Raphael benchmark:

AMD Zen 3 Ryzen Deep Dive Review: 5950X, 5900X, 5800X and 5600X Tested

www.anandtech.com

A deep-dive with a variety of different benchmarks showing the strengths and weaknesses of the CPU. Not "just" Geekbench. Not "just" SPEC. Definitely not "just" SPEC.

Hans Gruber said:
I have made the point in this thread (months ago) that AMD needs to have Zen4 out yesterday which is ASAP.

Why?

nicalandia said:
Dr. Ian Cutress has said multiple times that SPEC and GeekBench are about the best benchmarks to test Overall IPC

Look at Dr. Cutress' actual CPU reviews (see above) and tell me how much SPEC and GeekBench factored into those. There's a reason why they stuck GB5 under "synthetic".

Hans Gruber · May 25, 2022

DrMrLordX said:
Interesting! Of course those numbers aren't exactly verified yet, are they? Still it seems that most of Raphael's improvements come form higher sustained MT clocks than anything else.

Please God no. This is what we should ask for from a Raphael benchmark:

AMD Zen 3 Ryzen Deep Dive Review: 5950X, 5900X, 5800X and 5600X Tested

www.anandtech.com

A deep-dive with a variety of different benchmarks showing the strengths and weaknesses of the CPU. Not "just" Geekbench. Not "just" SPEC. Definitely not "just" SPEC.

Why?

Look at Dr. Cutress' actual CPU reviews (see above) and tell me how much SPEC and GeekBench factored into those. There's a reason why they stuck GB5 under "synthetic".

I will answer the why question that you quoted. Because Raptor Lake comes out later this year. If that is more powerful than Zen 4. AMD will be 15 months out from Zen 5. After Raptor Lake Intel will be on 7nm silicon. When you are #1 you get to charge a premium. By not releasing Zen 4 ASAP, there is no premium. AMD fanboys here do not accept that AMD is in 2nd position right now behind Alder Lake.

poke01 · May 25, 2022

Hans Gruber said:
After Raptor Lake Intel will be on 7nm silicon.

Which is around TSMC 5nm and 3nm. Intel will have reached node parity with Meteor Lake.

igor_kavinski · May 25, 2022

Why does everyone keep saying Zen 4 has AVX-512? All I see is support for BF16 which is used for AI. What am I missing?

JoeRambo · May 25, 2022

inf64 said:
Even if a lot of space was "wasted" on AVX512. 2x L/S pathways to accommodate AVX and 2x L2, it's incredible they didn't do much (if any) other improvements to the core.

They did, no doubt about it. 5-10% IPC improvement does not come from 1.5x L2 TLB and 1MB of L2 alone that's for sure. Other structures were increased as well, just GB leak tells us that it is same Zen3 with ~same counters + AVX512 stuff.
But one is free to hope that GB leak is old, AMD is sandbagging etc.

coercitiv · May 25, 2022

igor_kavinski said:
What am I missing?

You're missing the AVX-512, obviously.

Tuna-Fish · May 25, 2022

igor_kavinski said:
Why does everyone keep saying Zen 4 has AVX-512? All I see is support for BF16 which is used for AI. What am I missing?

The gigabyte leak (a ransomware group infiltrated Gigabyte servers and stole a whole bunch of stuff, then posted it online after they didn't get paid) contained a bunch of AMD documentation that listed how to detect presence of various instruction sets, from which can be inferred that AVX-512F, VL, BW, CD, IFMA, DQ, VPOPCNTDQ, BITALG, VNNI, VBMI, VBMI2 and BF16 are always present for all Zen 4 models. This documentation could of course be incomplete or out of date, as it was dated for August 2021, and turning off instruction sets is one of those things that can be done late in the release cycle or even post-release (usually because serious errata is found in something new).

DrMrLordX · May 25, 2022

Hans Gruber said:
Because Raptor Lake comes out later this year.

AMD is not worried about Raptor Lake, even if it is faster.

poke01 · May 25, 2022

DrMrLordX said:
AMD is not worried about Raptor Lake, even if it is faster.

Yes AMD won't be because RPL needs more watts/power to be faster. AMD will be fine till Lunar Lake.

DrMrLordX · May 25, 2022

poke01 said:
Yes AMD won't be because RPL needs more watts/power to be faster. AMD will be fine till Lunar Lake.

AMD is more worried about server and workstation. Mostly server. Enjoy your scraps! That's all we'll get from them on desktop.

deasd · May 25, 2022

DrMrLordX said:
AMD is not worried about Raptor Lake, even if it is faster.

We don't even know the exact launch date of Raptor... this's also what I was wondering

eek2121 · May 25, 2022

Timmah! said:
I dont think that he meant all-core turbo to be 5,2 to 5,5 during gaming. Just that iregardless of what core or core or cores are assigned to the game, all are capable to boost that high.

On topic of Adored TV - as he shows Lisa Su holding that prototype CPU in hand on Computex, he speculates those chiplets could actually hold more cores than 8 or that they might be space for another chiplet(s) - to support his former claim there will be more cores. Too bad at that particular time and place she says in the video literally "as you can see, Ryzen 7000 include up to 2 core chiplets, each with 8 Zen4 cores...." I presume this is why some people refuse to watch his stuff.

Exist50 said:
Even if you ignore a pretty poor track record with any company but Intel (which frankly is plenty enough reason to criticize), he's gotten a number of things wrong with Intel too, including "current" information. He clearly has some info source, but it's mixed with so much BS that he's not terribly useful. I know we're still waiting for proper Zen 4 numbers, and I'm even optimistic, but quite clearly the early claims of >30% single thread from various leakers were way off base.

I have heard claims he (and others, such as MLID) leaked stuff before, so I did some digging on twitter. Every single one of the supposed leaks came after twitter posts he had nothing to do with. He was just taking the tweets (and also posts in places like chiphell) and turning them into videos.

inf64 said:
So just to see if I get this correctly, a lot of people expect that Zen4 is a Zen3, shrank to 5nm with "only" 1MB L2 addition (2x) and "bolted" on AVX512? AMD had two years to just design a node shrank chiplet with bolted-on AVX512 and a slight L2 cache bump (and maybe a bit longer pipeline)? All that time invested and a whole node upgrade to create Zen3++ (basically)? Hmm ok.

edit: not to mention that core chiplet size is barely reduced Vs Zen3 (72 vs 80mm^2). I know that we were not going to see 0.5x reduction going to 5nm as only sram can scale that well, but that's still a huge chiplet on 5nm. Even if a lot of space was "wasted" on AVX512. 2x L/S pathways to accommodate AVX and 2x L2, it's incredible they didn't do much (if any) other improvements to the core.

If that is the case, don’t you find it amazing that Intel is being beaten by this chip?

Hans Gruber said:
I will answer the why question that you quoted. Because Raptor Lake comes out later this year. If that is more powerful than Zen 4. AMD will be 15 months out from Zen 5. After Raptor Lake Intel will be on 7nm silicon. When you are #1 you get to charge a premium. By not releasing Zen 4 ASAP, there is no premium. AMD fanboys here do not accept that AMD is in 2nd position right now behind Alder Lake.

Alder Lake does NOT beat Zen 3 in the vast majority of scenarios. Just look at AnandTech bench vs. the 5950x. The 5950x still has plenty of wins, despite using half the power.

Hans Gruber · May 25, 2022

DrMrLordX said:
AMD is not worried about Raptor Lake, even if it is faster.

Right, nothing to worry about. The Zen4 server chips will still be golden. What will AMD do if Meteor Lake arrives ahead of schedule and before Zen 5 and the Intel 7nm is far ahead of the TSMC 7nm process. What happens if 3nm TSMC is further delayed. Nothing to worry about?

That is why I said Zen 4 needed to be on the market yesterday. That was months ago. If Zen 4 is released only a few weeks before Raptor Lake, the process lead and performance crown mean nothing. If Zen 4 was out today, they would have months as #1 again.

nicalandia · May 25, 2022

igor_kavinski said:
Why does everyone keep saying Zen 4 has AVX-512? All I see is support for BF16 which is used for AI. What am I missing?

As per the Gigabyte Leaked data on Zen4(which has been spot on so far), Chips and Cheese found that Zen4 will have AVX512 and will use quite a large amount of Die size area

jpiniero said:
Speaking of which, do we have any idea how big the Bergamo chiplet is?

As per the Gigabyte Leaked data on Zen4(which has been spot on so far) the Zen4 CCD will be 72 mm^2 it's a bit longer than Zen3(due to AVX512 and 1 MiB of L2)

jpiniero · May 25, 2022

nicalandia said:
As per the Gigabyte Leaked data on Zen4(which has been spot on so far) the Zen4 CCD will be 72 mm^2 it's a bit longer than Zen3(due to AVX512 and 1 MiB of L2)

That's the Zen 4 die, not the Zen 4c one.

nicalandia · May 25, 2022

jpiniero said:
That's the Zen 4 die, not the Zen 4c one.

True, I was thinking about Genoa...

I've been trying to find that out and it's been bugging me. How are they going to fit 16 Cores on a die that has to be under 80 mm^2. I believe that they are going to cut L2/L3 in half and will use 5nm dense libraries.

inf64 · May 25, 2022

eek2121 said:
If that is the case, don’t you find it amazing that Intel is being beaten by this chip?

I'm not sure they will get a clear cut win if that's really the case (Zen4= Zen3 on 5nm with AVX512 and cache bump). We'll have to wait and see I guess.

I think it's time for me to create another (computerbase based) chart with my own benchmark scores projection. I'll try to do it today.

Heartbreaker · May 25, 2022

Hans Gruber said:
I will answer the why question that you quoted. Because Raptor Lake comes out later this year. If that is more powerful than Zen 4.

Raptor Lake is just a Rocket Lake refresh on the same process on Rocket Lake, it's unlikely to move the bar much.

maddie · May 25, 2022

Hans Gruber said:
Right, nothing to worry about. The Zen4 server chips will still be golden. What will AMD do if Meteor Lake arrives ahead of schedule and before Zen 5 and the Intel 7nm is far ahead of the TSMC 7nm process. What happens if 3nm TSMC is further delayed. Nothing to worry about?

That is why I said Zen 4 needed to be on the market yesterday. That was months ago. If Zen 4 is released only a few weeks before Raptor Lake, the process lead and performance crown mean nothing. If Zen 4 was out today, they would have months as #1 again.

What happens if the sky falls tomorrow?

Any sensationalist can imagine a horrible future. It's up to us to filter the malicious and foolish predictions from real world possibilities.

Could AMD even produce significant quantities starting several months ago? Your statement reads as divorced from reality. I want "X" now, so it must be possible and it must happen now. As we are witnessing painfully now, what we want and what happens in the real world can be reverse mirror images. I won't expand on that.

eek2121 · May 25, 2022

inf64 said:
I'm not sure they will get a clear cut win if that's really the case (Zen4= Zen3 on 5nm with AVX512 and cache bump). We'll have to wait and see I guess.

I think it's time for me to create another (computerbase based) chart with my own benchmark scores projection. I'll try to do it today.

Intel struggles to get a clear-cut win even today. They resort to obscene amounts of power to be competitive with the 5950x. A lot of people are expecting more out of Raptor Lake, but the reality of it is, doubling the E-cores takes power (around 50W) away from the P-Cores. If Intel decides to prioritize P-Cores, then E-cores are not going to run at full clocks. Adding 8 additional e-cores to a power limited chip is not going to provide the gains most people are expecting. Gracemont will be using 100w of the power budget, leaving 130 or so watts for the p cores. This is down from 180w with alder lake.

This is one of the issues with Intel's approach. They should be focusing on either speeding up Gracemont, or designing a new core entirely, not continually iterating on a large, inefficient, power-hungry design. Oh, and let's not forget that AMD has a process advantage. Let's also not forget that AMD can potentially add stacked cache to Zen 4 (possibly on the IOD).

To finish my comment, I wanted to add that there are likely a ton of improvements to the chip we aren't aware of. We've heard nothing of power consumption (idle power usage), IF improvements, DDR5 speeds/overclocking, etc.

biostud · May 25, 2022

Why would AMD disable AVX512?

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

Senior member

Golden Member

Diamond Member

Golden Member

Diamond Member

Lifer

Platinum Member

Golden Member

Lifer

Golden Member

Diamond Member

Golden Member

Lifer

Golden Member

Lifer

Senior member

Diamond Member

Platinum Member

Diamond Member

Lifer

Diamond Member

Diamond Member

Diamond Member

Diamond Member

Diamond Member

Lifer