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!

 

[Zucker2k]

Gotta give Nosta his props

 

[Abwx]

Gotta give Nosta his props

It s the exception that make the rule...


On another note Computerbase reviewer said that they have a lot more info about the 7000 series but that they are tied by a NDA until september 27.

 

[TESKATLIPOKA]

That was a typo but we re talking of the same slide, this one :

Even if the slide notes say that its a 5950X the slide itself say "5000 series", wich imply that it s an average of the whole lines, if it was only about the 5950X then there would be 49% somewhere in the other slide :

Ok, RPL-013 and RPL-014 is the same, so just a typo.

 

[inf64]

Clock for clock Zen 4 vs Golden Cove in ST portion of GB5 (12900KS @ 5.4-5.5Ghz Vs 7600X @ ~5.4Ghz according to .gb5):

Micro-Star International Co., Ltd. MS-7D30 vs Micro-Star International Co., Ltd. MS-7D69 - Geekbench

browser.geekbench.com

 

[Abwx]

Ok, RPL-013 and RPL-014 is the same, so just a typo.

I was talking of my mention of 59% instead of 49%...

 

[nicalandia]

Clock for clock Zen 4 vs Golden Cove in ST portion of GB5 (12900KS @ 5.4-5.5Ghz Vs 7600X @ ~5.4Ghz according to .gb5):

Micro-Star International Co., Ltd. MS-7D30 vs Micro-Star International Co., Ltd. MS-7D69 - Geekbench

browser.geekbench.com

When I saw AMD Pitting the 7600X against the 12900K, you know they just beat ADL at every meaningful performance measurements, Gaming and productivity.

 

[rainy]

There's also no guarantee that the IMC on a Zen 4 will even work at DDR5-6000, since it is officially rated for 5200.

You've clearly missed that: https://www.techpowerup.com/297736/...at-applies-zen-4-ddr5-6000-sweetspot-settings

Btw, limiting access to your profile wasn't a great idea because I was still able to find this "gem" of yours:

Frankly, Alder Lake is so far ahead of Zen 3 that Zen 4 is unlikely to do anything more than catch up to that 2021 chip.

 

[TESKATLIPOKA]

I was talking of my mention of 59% instead of 49%...

I know, but the footnotes are likely also wrong. RPL-013 and RPL-014 are the same thing.

 

[inf64]

When I saw AMD Pitting the 7600X against the 12900K, you know they just beat ADL at every meaningful performance measurements, Gaming and productivity.

Basically they have beaten the Golden Cove, on average. GC was around 11% faster than Zen 3 and Zen 4 edges it barely with 13% jump from Zen 3. AMD still has an ace in their sleeve with Vcache though, that can bring the average up by a few % points. On the other hand, Chips and Cheese had access to Raptor Lake and found no IPC increases in their own workloads https://chipsandcheese.com/2022/08/23/a-preview-of-raptor-lakes-improved-l2-caches/

"I ran some Champsim simulations with the different cache sizes, and expect IPC gains to be minor (simulations showed under 1% IPC improvement, though I feel the setup tends to underestimate gains). But focusing on IPC gains at fixed clock speed misses the larger picture. L2 misses decreased by 14.55% and 16.05% on the simulated P-Core and E-Core configurations, respectively. That means less data has to be pulled across the chip’s ring bus, saving power. High performance CPUs are often power limited, so using less power to move data around means that power could be spent hitting higher clocks. At higher clocks, Intel’s ring bus often doesn’t keep pace with core clock. Catching more accesses within L2 will minimize IPC loss as clock speed increases. And of course, there are bandwidth benefits to getting higher L2 hitrate. All of that comes together to give Raptor Lake some nice potential for performance and power efficiency increases, even if Intel doesn’t change the core architecture. "

 

[nicalandia]

I am putting up a chart to confirm that AMD has just said. 7600X is equal to 5800X3D in gaming.

 

[Markfw]

I have a question. With all the power improvements, like up to 49% performance at the same power (142 watts for 5950x, post 8403), how is it that the 7950x is supposed to draw 230 watts ? And some poster calculated in cb23 it would be 153 watts ? Something does not add up here. Yes, I can wait a month for the real benchmarks, which I am sure will include power, but so far, AMD has been right on with their ads on promised features.

Anybody ??

 

[Det0x]

Rumors say that der8auer will make a cpu delidding tool for AM5 👍

Will probably buy one when it become available

 

[Zucker2k]

Clock for clock Zen 4 vs Golden Cove in ST portion of GB5 (12900KS @ 5.4-5.5Ghz Vs 7600X @ ~5.4Ghz according to .gb5):

Micro-Star International Co., Ltd. MS-7D30 vs Micro-Star International Co., Ltd. MS-7D69 - Geekbench

browser.geekbench.com

What's the score, if we take out the AES score, as has been the practice on this forum, iirc?

Also, that score is a bit low for a 12900KS, and it's running on Windows 10 vs Windows 11 for the 7600x.

Here's a 12900k scoring better in both single core and multicore test.

ASUS System Product Name - Geekbench

Benchmark results for an ASUS System Product Name with an Intel Core i9-12900K processor.

browser.geekbench.com

One observation is looking at the multicore test, the 13700k is looking to top the 7900x, while the 7900x should most likely win the single core test due to higher clock.

 

[In2Photos]

I have a question. With all the power improvements, like up to 49% performance at the same power (142 watts for 5950x, post 8403), how is it that the 7950x is supposed to draw 230 watts ? And some poster calculated in cb23 it would be 153 watts ? Something does not add up here. Yes, I can wait a month for the real benchmarks, which I am sure will include power, but so far, AMD has been right on with their ads on promised features.

Anybody ??

I thought the 230w was the socket limitation for AM5. The 7950x is rated at 170w TDP.

 

[maddie]

What exactly were you disagreeing with? You offered a different point of view (I didn't touch on the current economic environment at all), but I found no disagreement.

You said:

But as Zen 3 already has shown if faced with the question of supplying low margin mass market or high margin specialist markets AMD will pick the latter, happily leaving the former to Intel which is still more competent at creating true mass market products in high quantity.

I replied:

I'm going to disagree here.

I suggest that the only reason that AMD went for margin at the expense of marketshare was their constrained production capacity. Once any business can produce at will they will be forced to go for volume. There is no other option if you want to grow revenue.

In other words there is no longer any "either or" choice for AMD.

 

[Markfw]

I thought the 230w was the socket limitation for AM5. The 7950x is rated at 170w TDP.

A bunch of others, exist50 and Zucker2k notably are of the mind it will use 230. I am even not sure of the 170 based on that calculation of 153 watts for cb23.

 

[FangBLade]

You've clearly missed that: https://www.techpowerup.com/297736/...at-applies-zen-4-ddr5-6000-sweetspot-settings

Btw, limiting access to your profile wasn't a great idea because I was still able to find this "gem" of yours:

Intel supporters are having hard times, great that they are revealed so we can create a list of them and filter them.

 

[Exist50]

I have a question. With all the power improvements, like up to 49% performance at the same power (142 watts for 5950x, post 8403), how is it that the 7950x is supposed to draw 230 watts ? And some poster calculated in cb23 it would be 153 watts ? Something does not add up here. Yes, I can wait a month for the real benchmarks, which I am sure will include power, but so far, AMD has been right on with their ads on promised features.

Anybody ??

AMD did not say +49% performance at the same power vs the 5000 series. They have a chart with numbers at various TDPs. Moreover, the last 60W will get comparatively little performance, as always.

 

[Exist50]

A bunch of others, exist50 and Zucker2k notably are of the mind it will use 230. I am even not sure of the 170 based on that calculation of 153 watts for cb23.

We have confirmation twice over. 170W TDP, 230W PPT. Consistent across this presentation and comments directly from AMD's chief of marketing.

 

[inf64]

Geekbench ST analysis between 12900KS and 7600X, AES excluded.

Text Compression	1635​	2060​	0.79
Image Compression	2049​	2046​	1.00
Navigation	1683​	1679​	1.00
HTML5	1833​	1766​	1.04
SQLite	1812​	1709​	1.06
PDF Rendering	1943​	1653​	1.18
Text Rendering	1960​	1820​	1.08
Clang	1794​	1928​	0.93
Camera	1822​	1638​	1.11
N-Body Physics	1986​	2193​	0.91
Rigid Body Physics	1966​	1960​	1.00
Gaussian Blur	1949​	2323​	0.84
Face Detection	2155​	2121​	1.02
Horizon Detection	1764​	1720​	1.03
Image Inpainting	3582​	3853​	0.93
HDR	3629​	3928​	0.92
Ray Tracing	2625​	2507​	1.05
Structure from Motion	1784​	1580​	1.13
Speech Recognition	1864​	1959​	0.95
Machine Learning	1605​	1787​	0.90
Average			0.99
12900KS ran at slightly higher peak clock (5461 vs 5394; 0.987x factor)
IPC excluding AES: 0.99 x 0.987 = 0.977 or ~ 2% lower IPC than Zen 4

This fits perfectly with AMD's 13% number, as Golden Cove has ~11% higher IPC than Zen 3 according to Ian's deep dive article and ComputerBase average ST chart (normalized for clockspeed)

 

[TESKATLIPOKA]

A bunch of others, exist50 and Zucker2k notably are of the mind it will use 230. I am even not sure of the 170 based on that calculation of 153 watts for cb23.

Abwx calculated 153W based on AMD's presentation about >35% higher performance withh 25% better efficiency compared to 5950x. He used the standard 105W TDP (142W PPT) and real power consumption of 130W.

65W TDP -> 88W PPT
105W TDP -> 142W PPT
170W TDP -> 230W PPT

 

[FangBLade]

What's also impressive is Zen 4 die size, it is much smaller than Intel, combined with 5nm yields will be excellent.

 

[gdansk]

I have a question. With all the power improvements, like up to 49% performance at the same power (142 watts for 5950x, post 8403), how is it that the 7950x is supposed to draw 230 watts ? And some poster calculated in cb23 it would be 153 watts ? Something does not add up here. Yes, I can wait a month for the real benchmarks, which I am sure will include power, but so far, AMD has been right on with their ads on promised features.

Anybody ??

Over and over AMD marketing is confusing you. But even in their own slides the truth is there. It's ~35% more perf/watt than Zen 3 at 170W TDP (230W PPT) and ~37% more perf/watt when both are at 105W TDP (142W PPT). Why would they include that? Because that's the actual generational uplift in MT.

They also show it's ~74% more perf/watt in the Genoa-like power limits for which Zen 4 was actually optimized.

 

[nicalandia]

I know that they are different games. But the AMD has been spot on the Average. So the 5800X3D is about 5% ahead of the 12900K on average and the 7600X is also about 5% ahead.... Impressive

 

[Abwx]

I know, but the footnotes are likely also wrong. RPL-013 and RPL-014 are the same thing.

There are a lot of things that are badly presented, FI there are graphs of perfs that werent scaled accordingly, they just changed the names of the loads and kept the same comparison bars , you can see the Cinebench bar at 48% lower than the other test at 45%...