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!

 

[soresu]

I am still wondering about the thick lid. Such a thick piece of metal generally isn’t good for thermal conductivity. The thermal conductivity of heat pipes can be orders of magnitude higher than any solid material (it is essentially passive phase change), so it is best to get the heat pipes as close to the die as possible.

I've been wondering about that.

Have they confirmed that it is in fact solid metal, and not something like a vapor chamber integrated to the heatspreader plate?

 

[jpiniero]

Robert Hallock confirmed both that the Ryzen 7000 series would be using up to 230 W PPT and that the CPU they demoed was in the 170 W TDP /230 W PPT class.

AMD is free to do and design products that they think are best and I hope they continue to be very successful, I just wished there was a bigger push for efficiency is all. I probably won't upgrade to Zen 4 anyway as I'm pretty happy with Zen 3 so we'll see what Zen 5 / whatever Intel has for desktops at that point brings.

If you think about it, they could have gotten more MT performance out of the 5950X if it could clock higher.

 

[jamescox]

I've been wondering about that.

Have they confirmed that it is in fact solid metal, and not something like a vapor chamber integrated to the heatspreader plate?

I haven’t seen any information. I was thinking integrated vapor chamber, but would that be economical for all models?

 

[Saylick]

Robert covers some of the stuff in that TPU interview, but there's a bunch more info he goes over in nuanced detail.

About the >15% ST uplift:
"The thing that AMD has always endeavored to do is to publish numbers that are highly reproducible and we tend to err on the conservative side. On purpose. We don't want to disappoint. We'd like to be realistic. It is worth saying that this is still early in the life of making these CPUs real. 3 months, 4 months, however long it takes to get to Fall. It doesn't seem like a long time, but it's a long time in terms of finalizing CPUs. So we guided greater than 15 because we definitely know it will be higher than 15, but how much higher will depend. And I will also say that there are multiple ways we're being conservative. As an example, the Zen 3 is a good analogy here. If you think back to our introduction of that architecture, we said on the whole that the 19% IPC uplift some were single digits, some were higher than that, but the geomean 19-20% and I feel that it is a widely accepted value. But if you look at Cinebench in particular, it was only 8, 9, 10%? So we used Cinebench at Computex, and I'm not promising that there is this massive IPC uplift across the board. I don't know yet, I haven't seen the numbers. All we've given is a tiny window into the world of what the chip can do. I am very optimistic about its performance, and there is a lot more to say, but you know, getting over 15% with a prototype part is not bad. More to come." [emphasis mine]

About the community on the 15%:
"I've seen some angst on that 15% number and I want to touch on that again. You know I can't go too deep into it. I want to say we were conservative on that number in 4 or 5 unique ways, and just give us the benefit of the doubt. I think we've earned it. We really try hard to make sure we deliver exactly as promised or often times a lot more than promised, so let's see how the summer goes before drawing conclusions. I think Zen 4 and the 7000 series will be extremely potent. It's a hellova upgrade. Give it time."

Edit: Here's AMD's Zen 3 IPC plot again. Not sure where CB R23 falls on this plot but....

 

[Henry swagger]

Wonder what pl2 will dragon range have

 

[Markfw]

PL2 is an Intel thing. This is an AMD thread. And desktop, not laptop.

 

[PJVol]

Not sure where CB R23 falls on this plot but....

I've tried to evaluate Zen2 -> Zen3 IPC and total (frequency gain included) performance uplift in CB R20 and CB R23, based on results of my 3600X and 5600X*, both ST and MT. For the IPC comparision, CPUs ran at 4300mhz OC, while total ST/MT uplift measured at stock settings, thus giving the following:
* R23 results for the 3600X were taken from reddit

	IPC ST	Total ST	IPC MT	Total MT
CB R20​	+11%​	+20%​	+12%​	+22%​
CB R23​	+10%​	+18%​	+13%​	+20%​

Regarding the Zen3 +19% so called Geomean, I tend to agree with Adored, that it's actually the biggest gain if we exclude games, and adding then "proper range of games" basically give whatever "geomean" you want )

 

[coercitiv]

and adding then "proper range of games" basically give whatever "geomean" you want

The same can be said about the rest of the workloads too: add the "proper" range of benchmarks to the list in order to tilt workload emphasis and you'll get whatever geomean you want. Arguing games are not that relevant for consumers is the same as arguing that most consumers do not use ANY rendering program for the lifetime of the CPU.

In fact we can take this a step further: show me a CPU that is great at gaming and isn't great at most other general consumer workloads.

 

[inf64]

I've tried to evaluate Zen2 -> Zen3 IPC and total (frequency gain included) performance uplift in CB R20 and CB R23, based on results of my 3600X and 5600X*, both ST and MT. For the IPC comparision, CPUs ran at 4300mhz OC, while total ST/MT uplift measured at stock settings, thus giving the following:
* R23 results for the 3600X were taken from reddit

	IPC ST	Total ST	IPC MT	Total MT
CB R20​	+11%​	+20%​	+12%​	+22%​
CB R23​	+10%​	+18%​	+13%​	+20%​

Regarding the Zen3 +19% so called Geomean, I tend to agree with Adored, that it's actually the biggest gain if we exclude games, and adding then "proper range of games" basically give whatever "geomean" you want )

Dr Ian Cutress verified the average 19% IPC uplift ,independently, in SPEC2017 ST, and other workloads. AMD is basically saying that R23 is on the lower end of the scale as we had diminishing returns even with Zen3 (Zen1->Zen2 was ~18% in R23, while Zen2->Zen3 was around 14%). If the trend continues with Zen4, it could be "only" 10% faster per clock than Zen3 in R23, while the geo-mean is between 13 and 15%. That would be still great given how high the Zen4 can clock, and how AMD has given the chip room for massive clock speeds by removing the old TDP limit on AM4.

 

[Timorous]

It will be interesting to see how this shakes out.

IPC uplift is looking lower than people expected from what I can tell but at the same time the clock speed uplift is looking greater than people expected so I am interested to see how it nets out.

In terms of performance I kind of expect around +20% on average for ST workloads. MT workloads depends on the SKU. I think for stuff like the 8c part with a 105W TDP MT workload uplift will be similar to ST uplift or maybe a small amount higher (less IO die power consumption allows for more core power consumption so higher all core clocks). For the 16c part though with a much higher TDP vs the 5950X I expect around 40% more performance.

Gaming will be interesting. On average I expect it will best the 5800X3D due to higher clock speeds and double the L2. There will be some titles though that are still much faster on the X3D. Stuff like Stellaris (In terms of tic rate) or the racing sims (ACC, iRacing) and a few other games with lots of background computation. The standard AAA stuff will probably do better on Zen 4 though.

The performance delta though may make it a tough sell for gaming only builders who already have AM4. The platform cost of an entry level Zen 4 part will be higher than just grabbing a 5800X3D and calling it done. Where Zen 4 will make sense is for people who want more productivity performance than they could get with the 5800X3D, that is where Zen 4 will be far more rounded and I would not be surprised if the 6c part is faster in productivity than 8c zen 3 most of the time.

 

[uzzi38]

Wonder what pl2 will dragon range have

As is everything in the laptop space, it's entirely down to what the OEM wants.

 

[coercitiv]

Gaming will be interesting. On average I expect it will best the 5800X3D due to higher clock speeds and double the L2.

And DDR5. What worked for ADL will also work fro Zen4, especially considering we're getting faster memory kits (bw & latency) and AMD already stated they are confident in Ryzen's ability to run high memory clocks.

Pricing for the whole package though.... we'll see.

 

[Kedas]

He also talked about TDP, that there wasn't a profile 'fused' yet (since they haven't tried/decided that yet) somewhere between 105 and 170 but certainly not maxed out.
I think they were shooting for an 125W TDP setting similar like the 12900k
In that case they get 45% more speed for the same power package compared to 12900k. (Obviously it's 5nm, gives +20% over 7nm)

He also referenced to it as a 'prototype part', sounds like they have to figure out a lot despite being in full 'die production' already...
It may be possible that he was using an previous stepping.

About the >15% they just don't know yet exactly since they haven't decided on SKU profiles. It will also depend on what intel does by then I guess.

Also the instruction set is not exclusive to AMD if I understood correctly, so that would mean AVX512 or less.

One other thought if all CPU's you can buy will have video decoding/encoding hardware do they still need to be present on the GPU cards? (intel has it already a long time)

 

[PJVol]

The same can be said about the rest of the workloads too: add the "proper" range of benchmarks to the list in order to tilt workload emphasis and you'll get whatever geomean you want.

I also wonder why they didn't.
And yet, that 19% is the highest uplift among non-gaming workloads presented. And I don't think games are irrelevant, just 3rd party reviews showed a bit different numbers.
For example "the real world" testing scenario

Is Zen 3 Worth It for Gaming? Ryzen 5600X vs. 3600 vs. Core i5-10400F

In order to wrap up our 6 core/12 thread CPU testing, we had to add the Ryzen 5 5600X to the mix. We're going back to the...

www.techspot.com

where 5600X is ~ 10% faster than 3600 in 1080p tested in 16 games. And that is total uplift.
And from their 5600X review, in a "CPU limited" scenario ~ 24% faster.

 

[DrMrLordX]

So blender is a good measurement of true IPC gains based on uArch design instead of new IO or memory interface?

It's more of an fp benchmark. It's kind of a cache benchmark as well, but. Definitely fp. You can't use it to model integer performance.

230W peak! Looks like those 5.5+ GHz frequencies don't come for free.

Eww.

View attachment 62155

Well that's . . . interesting. Makes it hard to get an idea of Raphael's perf/watt from the demo. Guess we'll need more data.

View attachment 62154

That chip may not be at launch. 3950X wasn't.

 

[coercitiv]

And btw, I didn't say games are irrelevant

I wasn't putting words in your mouth, just taking the argument to the extreme in order to make a point (let's remove games, let's remove rendering, let's remove X,Y, Z etc). Personally I don't agree with AdoredTV on this, his hindsight justification does not stand up against results obtained in SPEC as @inf64 already showed.

You can think of the AMD marketing IPC graph as a way to visualize what 19% IPC means to the average consumer, in the sense that AMD may evaluate the IPC gain using a proprietary benchmark suite and then show the audience how 19% gains look like when distributed among a spectrum of known workloads and benchmarks. The fact that they include many games also helps the audience understand variance, not necessarily the weight given to gaming in the final score. Remove all but one game from the spectrum and they would actually induce false expectations regarding gaming performance, as this would suggest a steady, uniform gain.

Moreover, if the paragraph above is correct, removing the games and calculating a new average from that particular set would actually skew the results greatly even for non-gaming workloads. Obviously this is just one possible hypothesis, but I hope it's clear that trying to extract this much information from a marketing slide can result in equally inaccurate and deceitful conclusions.

 

[PJVol]

Dr Ian Cutress verified the average 19% IPC uplift ,independently, in SPEC2017 ST, and other workloads

I'd like to see some other clock-for-clock testing results with the more common set of benchmarks, just for the sake of completeness ) (if such exist).
Quick googling didn't work. If you know of any, post a link pls.

If the trend continues with Zen4, it could be "only" 10% faster per clock than Zen3 in R23

That would be great, but my estimates are still more pessimistic, I think ~ 5-8% in R23

trying to extract this much information from a marketing slide can result in equally inaccurate and deceitful conclusions.

Yep, fully agree with this. Just wanna see what conclusions one can come to based on "educated" guesswork )

 

[inf64]

I'd like to see some other clock-for-clock testing results with the more common set of benchmarks, just for the sake of completeness ) (if such exist).

Another popular benchmark is Geekbench, you can find it in the same deep dive article by AT. Adjusted for clock speeds, Zen3 is almost exactly 19% faster per clock than Zen2 in ST portion of the test:


1655 x 4.7 / 4.9 ~= 1587 pts for 5950X running at the same ST boost clock as 3950X
1587/1340=1.184 or 99.5% of 19% IPC average that SPEC2017 showed

 

[lobz]

If Zen 4 was 37% faster in ST, they would have given a number better than 15%. There's some wiggle room, sure, but I think hoping for more than twice the given number is extremely unrealistic.

dude if you really must go down the road of this sacred leakerbusting, at least try to be bothered to admit, that the 37% figure was mentioned as an 'as high as' metric, so at the high end of what's possible, not as a general 37% increase in ST workload performance. You might wanna re-watch his video about that.

 

[nicalandia]

Robert covers some of the stuff in that TPU interview, but there's a bunch more info he goes over in nuanced detail.

About the >15% ST uplift:
"The thing that AMD has always endeavored to do is to publish numbers that are highly reproducible and we tend to err on the conservative side. On purpose. We don't want to disappoint. We'd like to be realistic. It is worth saying that this is still early in the life of making these CPUs real. 3 months, 4 months, however long it takes to get to Fall. It doesn't seem like a long time, but it's a long time in terms of finalizing CPUs. So we guided greater than 15 because we definitely know it will be higher than 15, but how much higher will depend. And I will also say that there are multiple ways we're being conservative. "As an example, the Zen 3 is a good analogy here. If you think back to our introduction of that architecture, we said on the whole that the 19% IPC uplift some were single digits, some were higher than that, but the geomean 19-20% and I feel that it is a widely accepted value. But if you look at Cinebench in particular, it was only 8, 9, 10%? So we used Cinebench at Computex"

Guys we have just learned from AMD that the >15% performance boost was on Cinebench R23 and that the least favorable benchmark possible. And it's true, go on and check what is the performance boost from Zen2 to Zen3 on Cinebench R23. In the Single Digits(about 9%)...

 

[coercitiv]

Guys we have just learned from AMD that the >15% performance boost was on Cinebench R23 and that the least favorable benchmark possible.

We knew that from day one.

 

[nicalandia]

We knew that from day one.

Perhaps you and informed posters, but people that don't know any better got histerical and were saying OMG just >15% boost ?? Its the end of the world.

 

[AtenRa]

Guys we have just learned from AMD that the >15% performance boost was on Cinebench R23 and that the least favorable benchmark possible. And it's true, go on and check what is the performance boost from Zen2 to Zen3 on Cinebench R23. In the Single Digits(about 9%)...

Actually ZEN 3 Cinebench 1T IPC was 13% for R20 and 18% for R15

 

[Anhiel]

Everything is relative. The values may differ from test suit to test suit and suit toward certain type more than the other.
But test results within a test suit should idealy show the same curves as any other test, just shifted or scaled.
Also the ratio between one SKU to another should be still the same. So as long as tests are done consistently, any ratio you get from any good test suit is equally good for evaluation.

Note how the average is lower for applications compared to games so this is more real world accurate for most usages. So it would be more accurate to have evaluations be divided into more fine grained groups like for normal application, scientific/math, and game. Ofc this entails lots of testing...

Back onto the efficiency, if you look at the Ryzen 6k series on N6 these appear to get 4.67x compared to Zen. As I said before Zen4 needs 4.6x to keep up with efficiency evolution. So at least for laptops Zen4 should be able to match or better that value.
Desktop on the other hand appear to be ~89.4% as efficient (ofc worse) compared to laptops (5900HX vs 5950X). Since N6 is just a density improvement node the gains could be slightly multiplied by N5P which offer +7% Pef or -15% power compared to N5.

Ryzen 6k brings about ~+9% IPC and if Zen4 is on N5P as suggested by AMD by saying it being an improved N5 node we are looking at a possible balance pool of 15+7=22%(23.05% multiplied) performance or 30+15=45% (49.5% multiplied)power plus Ryzen 6k's 9% . This makes the released information look even more odd. ~13% perf already went into higher clockspeed. So given the perf/power ratio of 1:2 we still got 9% perf or 19% power left to balance out plus Ryzen 6k's 9%. Beyond these values diminishing returns will result in higher power consumption.

 

[nicalandia]

Actually ZEN 3 Cinebench 1T IPC was 13% for R20 and 18% for R15

It was less for R23 and CB R23 is what AMD is using as reference for the most recent info on performance