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

Page 101 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Vattila

Senior member
Oct 22, 2004
809
1,412
136
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).

Untitled2.png


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! :)
 
Last edited:
  • Like
Reactions: richardllewis_01

DrMrLordX

Lifer
Apr 27, 2000
22,065
11,695
136
Wasn't the power limits on AM4 the reason they couldn't push the Zen chips harder or was that bunk info? Someone mentioned it on an IRC to me a year or two ago, but I never looked into it. AM4 came out long before Ryzen was announced, no?

AM4 first launched with A320 and Bristol Ridge in very limited markets. There were some A320s sold in the German and Japanese markets (and possibly China) that were never really designed for Ryzen and, if I recall correctly, had UEFIs that wouldn't even boot Ryzen 1xxx chips. As for power limits? The socket itself could move impressive amounts of power even on early x370 boards, but VRMs were all over the place (especially on A320 and B350 boards).
 

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
That would be an interesting fight:. AMD Zen 4 cores, 24 cores, 48 threads, all performance vs. Intel Cove next in an 8+16 p/e ratio with 32 threads. That's going to be quite interesting to see benchmarked.
I'd say that would rather be a slaughterhouse. 16/32 Zen4 vs 8/16+16 RTL on the other hand? Now THAT will be really interesting to see, where the strengths and weaknesses of both designs really lie.
 

DrMrLordX

Lifer
Apr 27, 2000
22,065
11,695
136
I'd say that would rather be a slaughterhouse. 16/32 Zen4 vs 8/16+16 RTL on the other hand? Now THAT will be really interesting to see, where the strengths and weaknesses of both designs really lie.

Raptor Lake may actually be in an 8P + 32E configuration at the top end. For consumers I don't see that as being much of a win, especially since AMD's 16c chips are still overkill for most desktop users. If AMD really can average +29% IPC on Zen4 while sustaining clocks, throwing out scads of E cores won't do Intel many favors. AMD has also made significant strides in interconnect bandwidth/latency to keep intercore communications between CCDs snappy enough that you don't really notice many problems when running software designed for monolithic CPUs. It's not clear at all how intercore latency will affect Alder Lake or Raptor Lake. Possibly not well at all.
 
  • Like
Reactions: Tlh97 and lobz

eek2121

Diamond Member
Aug 2, 2005
3,100
4,398
136
That would be an interesting fight:. AMD Zen 4 cores, 24 cores, 48 threads, all performance vs. Intel Cove next in an 8+16 p/e ratio with 32 threads. That's going to be quite interesting to see benchmarked.
I'd say that would rather be a slaughterhouse. 16/32 Zen4 vs 8/16+16 RTL on the other hand? Now THAT will be really interesting to see, where the strengths and weaknesses of both designs really lie.
An 8+16 part would beat a 5950x unless…
Raptor Lake may actually be in an 8P + 32E configuration at the top end. For consumers I don't see that as being much of a win, especially since AMD's 16c chips are still overkill for most desktop users. If AMD really can average +29% IPC on Zen4 while sustaining clocks, throwing out scads of E cores won't do Intel many favors. AMD has also made significant strides in interconnect bandwidth/latency to keep intercore communications between CCDs snappy enough that you don't really notice many problems when running software designed for monolithic CPUs. It's not clear at all how intercore latency will affect Alder Lake or Raptor Lake. Possibly not well at all.

Such a design would be probably be power limited in all core workloads. Gracemont may be efficient, but I doubt it is 32x as efficient.
 

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
An 8+16 part would beat a 5950x unless…


Such a design would be probably be power limited in all core workloads. Gracemont may be efficient, but I doubt it is 32x as efficient.
I don't understand you mentioning us here. The 5950 is not based on Zen 4.
 

DrMrLordX

Lifer
Apr 27, 2000
22,065
11,695
136
An 8+16 part would beat a 5950x unless…

5950X is the least of Intel's worries.

Such a design would be probably be power limited in all core workloads. Gracemont may be efficient, but I doubt it is 32x as efficient.

Alder Lake is already going to be power-limited in all-core workloads. Adding more e cores instead of p cores won't change that. Intel's real problem is that they'll be under pressure to throw more die area at increasing performance without being able to grow beyond 10ESF. And that's why Zen4 is such a threat. AMD doesn't have that limitation.
 

A///

Diamond Member
Feb 24, 2017
4,351
3,158
136
AM4 first launched with A320 and Bristol Ridge in very limited markets. There were some A320s sold in the German and Japanese markets (and possibly China) that were never really designed for Ryzen and, if I recall correctly, had UEFIs that wouldn't even boot Ryzen 1xxx chips. As for power limits? The socket itself could move impressive amounts of power even on early x370 boards, but VRMs were all over the place (especially on A320 and B350 boards).
Ah yes. Funny you mention that. I was one of those people wanting to buy and had access to x370 boards to play around with. There were some good boards realized later on, but I have to say not everything was nice. And I can't blame board makers. AMD hadn't been relevant for a long time and it would have been hard to invest that kind of capital into supporting a new processor line if AMD gave up the ghost soon. Zen+ was decent by all accounts, and it warmed board makers. Zen 2 was impressive, and Zen 3 was obviously incredible. We're about a year away from the supposed launch of Zen 4 and I'd say there's a lot riding on that, especially with the new AM5 socket. If AMD can bring new performance heights at or greater than Zen 3 gained performance against Intel's own products, which I believe would be Raptor Lake, then we're going to get some excellent pricing wars, IMO. There's only so much AMD can do with raised prices where Intel can take it on the nose for a while and drive prices down, especially if they're highly competitive processors, they being the Raptor Lakes in late '22. Intel has the means to produce and produce until they bleed in the nose, whereas AMD is limited to X WPM from TSMC. Calling that timeline interesting would be selling it short.

Of course I'll be there arguing with you about the whole 12-18 month timeline.


Edit: It had not occurred to me with the launch of Alderlake being roughly 2 months and some weeks away we should be beginning to get good leaks about performance within the next 4-5 weeks. I need to travel out of the country for work so hopefully my plane doesn't go missing!
 

DrMrLordX

Lifer
Apr 27, 2000
22,065
11,695
136
Ah yes. Funny you mention that. I was one of those people wanting to buy and had access to x370 boards to play around with. There were some good boards realized later on, but I have to say not everything was nice.

x370 Taichi was actually one of the best, if not the best x370 boards for raw power delivery. It could push 300W. It was available day one, albeit in extremely limited quantities.

Intel has the means to produce and produce until they bleed in the nose, whereas AMD is limited to X WPM from TSMC.

Intel has limits. If they intend to ship anything on 7nm/Intel 4 or beyond, those limits will strangle them.
 

eek2121

Diamond Member
Aug 2, 2005
3,100
4,398
136
I don't understand you mentioning us here. The 5950 is not based on Zen 4.
Perhaps I wasn’t being clear enough. I was implying that AMD would need a 40-50% gain in performance for Zen 4 to keep up with Raptor Lake.

5950X is the least of Intel's worries.



Alder Lake is already going to be power-limited in all-core workloads. Adding more e cores instead of p cores won't change that. Intel's real problem is that they'll be under pressure to throw more die area at increasing performance without being able to grow beyond 10ESF. And that's why Zen4 is such a threat. AMD doesn't have that limitation.

Alder Lake will not be power limited. It performs very well.
 

soresu

Diamond Member
Dec 19, 2014
3,230
2,515
136
Perhaps I wasn’t being clear enough. I was implying that AMD would need a 40-50% gain in performance for Zen 4 to keep up with Raptor Lake.
That remains to be seen.

Intel have spent years eroding any measure of goodwill on their future product promises - so I would not take any performance estimates for granted until it's literally weeks from silicon being in products.
 

eek2121

Diamond Member
Aug 2, 2005
3,100
4,398
136
What are you talking about?
Gracemont performs better than Zen 2 and worse than Zen 3. Throw 16 of them in a chip along with the Raptor Cove, and assuming Intel isn't power limited, the chip will be 40-50% faster than a 5950X. That means a 6950X will need to be 40%-50% faster to tie with the Intel chip. That is why AMD needs a 24 core SKU. Note that Raptor Lake is currently rumored to launch BEFORE Zen 4. If Raptor Cove is power limited, then AMD doesn't need to make up such a wide gap.

Never mind that we don't even have silicon to prove that, but the PL numbers we've seen indicate that Alder Lake will be a power hog that will not be able to sustain max clocks and "perform very well" without pulling ~220W or more.

125W PL1, 241W PL2. That is from the Gigabyte leak. Raptor lake apparently has a lower PL2 (I would have to dig up the Igor's lab article).
 

A///

Diamond Member
Feb 24, 2017
4,351
3,158
136
Intel has limits. If they intend to ship anything on 7nm/Intel 4 or beyond, those limits will strangle them.
Yep, if Intel can get their process down to limit waste and understand what makes them more money, they'd focus on that. It's going to be interesting to see Intel crawl back and attempt to regain their lost market space from AMD. How they do it is up to anyone's guess since we don't know what their next 2-3 products will perform after Adlerlake, and we know very little about Zen 4 and beyond. Almost makes it exciting to be a consumer again. Someone posted an interesting outlook here of Intel possibly shifting some work to TSMC in future so they can focus on pumping out products for certain markets. I'd be very impressed if AMD can maintain a large IPC gallop while maintaining normal power limits each generation.

The only market we won't see much price competition in would be video cards unless Intel one day decides to use their own foundries for their video cards, provided their video cards are worth using.
 
  • Like
Reactions: Tlh97

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
Gracemont performs better than Zen 2 and worse than Zen 3. Throw 16 of them in a chip along with the Raptor Cove, and assuming Intel isn't power limited, the chip will be 40-50% faster than a 5950X. That means a 6950X will need to be 40%-50% faster to tie with the Intel chip. That is why AMD needs a 24 core SKU. Note that Raptor Lake is currently rumored to launch BEFORE Zen 4. If Raptor Cove is power limited, then AMD doesn't need to make up such a wide gap.
This is the Richie Rich heights of math built on imaginary foundations that I adore so much. I'm not accusing you of being an Intel or any kind of fan, because that's irrelevant. I'm just saying you're probably being mislead, because your calculation reminds me of how people here have visioned Rocket Lake being a GaMiNG KiNG, all based on Intel's ever so misleading marketing (they literally hyped RKL up as the gaming saviour FOR MONTHS!) and leaked GeekBench numbers. I remember a still very active member calculating 30% IPC increase in GeekBench over Comet Lake, and deciding that AMD was done.

I don't really blame you, since Pat Gelsinger is giving interviews talking like a troll from wccftech, while talking as much about competitors as about his own company and using phrases like "trust me, we are have things cooking still". I almost thought he was gonna say "trust me dude" to that journalist.

So yeah all in all I get why it can resonate with enthusiasts without a shed of critical thinking, but I also think it's a very-very dangerous game for him to play. "We have 2 cores now, and you know, AMD has only one!"... the more conservative shareholders must be very happy reading these 🤣🤣🤣
 

andermans

Member
Sep 11, 2020
151
153
76
Interesting thing though is that in most cases one is power limited in MT benchmarks. If you have 50% more cores each core gets 33% less power. If the power/performance curve was linear that would mean performance stays equal. Of course the power/performance curve isn't linear, but we don't have real power/performance curves for golden cove and gracement.

If I take Zen3 power/frequency from Anandtechs Zen3 deep dive, I'd estimate that with 33% less power per core from the official TDP, Zen3 will clock 10-20% lower. At 10% we're talking 1.5x * 0.9 = 1.35x a.k.a 35% more performant with 50% more cores, and at 20% that would result in 20% more performant with 50% more cores.

Of course that assumes certain power/frequency curves, and assume performance scales linear with frequency (which it won't. One way because latency is better at lower clocks, on the other hand with more cores you're more likely to be memory bound, though that effect may be limited in cinebench).

At 225-250W (unlocked) the scaling might be better because you're way at the inefficient point of those cores. Anyway, don't look to hard at the numbers since I used the wrong power/frequency curves, this was all for illustration purposes because I was curious about the magnitude of this effect myself.
 
  • Like
Reactions: Tlh97

Abwx

Lifer
Apr 2, 2011
11,557
4,349
136
but we don't have real power/performance curves for golden cove and gracement.

We have something that is enough for estimations.
Thoses curves are the processes typical perf/power curves, no matter what is the uarch.

27.png