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!

 

[eek2121]

That is incorrect. It is in stock shipped and sold by Amazon right this moment.

yes, Amazon started selling a few hours ago. They completely missed the launch date.

 

[eek2121]

The CPU will always hit 95C and limit the clock to maintain 95C at all times as power draw goes higher (unless you manually lower the number in bios). As to what clocks can you get with the mentioned cooler, I have no clue. My guess is that it should be somewhat lower than AIOs which AMD recommends (they do not recommend air coolers for 79xx).

This is correct. It is designed to hit 95C. There is nothing wrong with that.

 

[In2Photos]

If you can't watch the video he was able to achieve the following:

7700X, just under 20k in CB R23 Multi
-30 voltage offset provided a 200MHz ALL CORE boost (from 5100-5300), over 20k in CB R23 Multi
Set temp limit to 85C provided roughly the same CB R23 performance as the offset.
Set a power limit of 85W and was able to still maintain 5100 all core and nearly matched the stock CB score, but at 61C!!!

 

[Det0x]

*edit*
I'm getting old and slow

 

[moinmoin]

If you can't watch the video he was able to achieve the following:

7700X, just under 20k in CB R23 Multi
-30 voltage offset provided a 200MHz ALL CORE boost (from 5100-5300), over 20k in CB R23 Multi
Set temp limit to 85C provided roughly the same CB R23 performance as the offset.
Set a power limit of 85W and was able to still maintain 5100 all core and nearly matched the stock CB score, but at 61C!!!

View attachment 68346

View attachment 68347

It's interesting that his final pick is curve optimizer plus power limit. I'd use curve plus temp limit, wonder how a temp limit of 60°C would compare to stock and his final results.

 

[In2Photos]

It's interesting that his final pick is curve optimizer plus power limit. I'd use curve plus temp limit, wonder how a temp limit of 60°C would compare to stock and his final results.

Technically he employed all 3, but by reducing power he never hit the temp limit. My guess is that you could use the voltage offset with either temp or power to achieve similar results though.

 

[eek2121]

No per core undervolt? It seems like a better approach would be to see which core is running the hottest, and undervolt that core. Rinse and repeat until errors pop up or performance drops off too much.

Gamers Nexus noticed that only 1-2 cores were actually hitting 95C, the rest were well under that.

 

[inf64]

No per core undervolt? It seems like a better approach would be to see which core is running the hottest, and undervolt that core. Rinse and repeat until errors pop up or performance drops off too much.

Gamers Nexus noticed that only 1-2 cores were actually hitting 95C, the rest were well under that.

Also, we know from derbauer that IHS contact is not the greatest and further significant gains can be achieved with better IHS/die contact.

 

[In2Photos]

No per core undervolt? It seems like a better approach would be to see which core is running the hottest, and undervolt that core. Rinse and repeat until errors pop up or performance drops off too much.

Gamers Nexus noticed that only 1-2 cores were actually hitting 95C, the rest were well under that.

I'm sure that if you want to pull the maximum performance from the CPU you could spend the time to do this. The settings he used can be done in 15-20 seconds and are good enough for 99.9% of the people that will be looking for limiting power and temps. I used his 5000 series video on my son's 5600X, my first time using a Ryzen CPU, and was amazed at the results just from a few clicks in the BIOS. His CPU would only achieve a -20 offset, but still did wonders.

 

[Hitman928]

I always question the actual stability when people show large improvements like this. I used curve tracer on my 5900x to improve the perf/w, but was not able to get the same large increases in efficiencies some of the youtubers were showing. The thing is, since it's not a static overclock, it makes testing stability much more difficult as you have a very dynamic v/f behavior that you have to test. I was able to get similar results as the youtubers if I just used cinebench or even prime95 as my stability test, but then you're only really testing the all core v/f point. However, under lighter loads, the freqency and voltage can vary wildly and will do so jumping from core to core.

It took me probably close to a week to finally be satisfied that my final curve tracer settings were stable and I had to adjust individual core offsets multiple times through the process. Maybe I just got unlucky in the silicon lottery, but I have a very strong suspicion its more that the numbers you see shown online aren't really stable, especially when they do the blanket, all core offsets as shown in the latest video. There will always be some margin AMD builds in that you can take advantage of, I'm just highly skeptical of such large gains being truly stable except maybe for golden sample silicon.

 

[In2Photos]

I always question the actual stability when people show large improvements like this. I used curve tracer on my 5900x to improve the perf/w, but was not able to get the same large increases in efficiencies some of the youtubers were showing. The thing is, since it's not a static overclock, it makes testing stability much more difficult as you have a very dynamic v/f behavior that you have to test. I was able to get similar results as the youtubers if I just used cinebench or even prime95 as my stability test, but then you're only really testing the all core v/f point. However, under lighter loads, the freqency and voltage can vary wildly and will do so jumping from core to core.

It took me probably close to a week to finally be satisfied that my final curve tracer settings were stable and I had to adjust individual core offsets multiple times through the process. Maybe I just got unlucky in the silicon lottery, but I have a very strong suspicion its more that the numbers you see shown online aren't really stable, especially when they do the blanket, all core offsets as shown in the latest video. There will always be some margin AMD builds in that you can take advantage of, I'm just highly skeptical of such large gains being truly stable except maybe for golden sample silicon.

As I mentioned previously I used the voltage offset in PBO on my son's 5600X. It was not stable at anything more than a -20. He has been using that system daily for gaming/Youtube watching since the first of the year. We had an issue with crashing early on that a BIOS update fixed. No issues since. I realize that a 5600X is easier to cool than a 5900X, but it is possible to see good results from a quick adjustment. In his case we saw temps decrease by 10-15C which allowed us to dial back his AIO to the point of being completely inaudible during normal use.

 

[Hitman928]

As I mentioned previously I used the voltage offset in PBO on my son's 5600X. It was not stable at anything more than a -20. He has been using that system daily for gaming/Youtube watching since the first of the year. We had an issue with crashing early on that a BIOS update fixed. No issues since. I realize that a 5600X is easier to cool than a 5900X, but it is possible to see good results from a quick adjustment. In his case we saw temps decrease by 10-15C which allowed us to dial back his AIO to the point of being completely inaudible during normal use.

How did you check stability to get to -20 offset?

 

[In2Photos]

How did you check stability to get to -20 offset?

Cinebench, Prime 95, and Heaven along with playing the games. -25 offset worked during benchmark testing, but not in a couple of games.

 

[moinmoin]

Technically he employed all 3, but by reducing power he never hit the temp limit. My guess is that you could use the voltage offset with either temp or power to achieve similar results though.

Good point! Hope somebody will test performance and power usage at low temp limits. How well would Zen 4 run at e.g. 40°C? And that with passive cooling?

I guess the most interesting part is at what point ST performance starts to get affected.

 

[inf64]

From the Optimumtech video, the results of 7950X tuning, great results:

 

[CakeMonster]

IIRC these kinds of improvements from tunings only showed up in professional workloads on Z3, and made no difference in games. Is that correct still and has anyone tested games with the same tunings?

 

[inf64]

IIRC these kinds of improvements from tunings only showed up in professional workloads on Z3, and made no difference in games. Is that correct still and has anyone tested games with the same tunings?

Tuned performance is the same or slightly better vs stock in games (at least for 7700X):

 

[LightningZ71]

Not really in their control. That is up to ssd and other providers to offer things that actually need pcie5 and that is where the problem lies. Even a RTX 4090 doesn't need it. On the ssd side, for NAND ssds the max bandwidth figures are mostly irrelevant for consumers as well. The number of consumers moving very large files from ssd to ssd is pretty limited. What consumers notice is copying tons of small files in a folder and that is were even these pcie5 ssds won't offer anything over a sata ssd. (in general IOPS at low QD for small files). That is were optane shined. A shame it's gone.

Oh, they aren't gone. They may not be in volume producti on, but you can still readily get them, especially the little accelerator nvme cards. There are lower cost PCIe 3.0 x4 cards out there with cheap asmedia 2 into 1 x4 plx chips on them that give you a pair of m.2 x4 slots. Thro2 in a pair of dirt cheap 32GB optane drives, setup a raid 0, and you have 64GB of storage that you can't hope to saturate. I've seen them used as swap file space for someone that would often overflow his 128GB of ram and the improvement was night and day compared to killing his SSD. I've seen it used as SSD tier cache in a MS storage spaces tiered storage arrangement with a bunch of regular HDDs and it was typically as responsive as a refular SSD of that size, sometimes faster.

 

[inf64]

BTW, here is Overclockers UK video showing a tuned 7950X with PBO2 hitting 6Ghz on single core (on 3 cores actually) :

Mind you this is all automatic stuff with PBO2. Impressive.

 

[pakotlar]

BTW, here is Overclockers UK video showing a tuned 7950X with PBO2 hitting 6Ghz on single core (on 3 cores actually) :

Mind you this is all automatic stuff with PBO2. Impressive.

View attachment 68362

Effective clock is 5.8Ghz, not 6. Still very nice.

 

[CakeMonster]

Tuned performance is the same or slightly better vs stock in games (at least for 7700X):

Thanks, its reassuring that gaming isn't worse at least. I might actually try it then, even though gaming is my main priority and I usually don't load fully often enough to worry about power.

 

[inf64]

Effective clock is 5.8Ghz, not 6. Still very nice.

Core effective clock is not the same as maximum clock some of the cores hit :


So yes, some of the cores did hit 6Ghz.

Thanks, its reassuring that gaming isn't worse at least. I might actually try it then, even though gaming is my main priority and I usually don't load fully often enough to worry about power.

HW Unboxed reviewed the 7700X, impressive gaming results ( performs better in a number of games that have dual CCD scheduling issues):

 

[Det0x]

AMD Ryzen 7000 heatspreader and cooling analysis – temperatures, hotspots and problems

AMD Ryzen 7000 heatspreader and cooling analysis - temperatures, hotspots and problems | igor´sLAB

Today’s article will focus exclusively on the heatspreader of the Ryzen 7000 CPUs and, of course, on cooling, the optimal application of thermal paste and the question of where the nasty hotspots…

www.igorslab.de

By the way, I also stopped at 87 degrees on the IHS because I would like to continue using the silicon, which was already sizzling away here at about 107 degrees. This would also prove where the delta with the up to 20 Kelvin difference actually comes from:

Substrate -> Oxide -> Solder -> Oxide -> 2.8mm IHS -> Oxide -> 0.05mm Phase Changer

 

[Hitman928]

Cinebench, Prime 95, and Heaven along with playing the games. -25 offset worked during benchmark testing, but not in a couple of games.

Based upon my testing, this isn't sufficient as Cinebench and Prime will only keep the cores near their lowest all core boost frequency. Heaven will let the cores bounce around but doesn't require much CPU work and there's no real way to catch errors unless it gets really unstable, same applies to games. You could game fine for hours and not really be stable. To each their own and maybe I'm just old school on this, but truly testing stability on modern CPUs is very difficult and time consuming and personally, I don't feel comfortable on a machine that I haven't put through a gauntlet to ensure it's stable.

 

[In2Photos]

Based upon my testing, this isn't sufficient as Cinebench and Prime will only keep the cores near their lowest all core boost frequency. Heaven will let the cores bounce around but doesn't require much CPU work and there's no real way to catch errors unless it gets really unstable, same applies to games. You could game fine for hours and not really be stable. To each their own and maybe I'm just old school on this, but truly testing stability on modern CPUs is very difficult and time consuming and personally, I don't feel comfortable on a machine that I haven't put through a gauntlet to ensure it's stable.

What do you define as stability? If my son is able to do all the things he wants to use the PC for without crashing, that, to me, is stable, is it not? He's been using the PC for 9 months now. Heaven was included in testing to make sure the GPU and PSU were also up to the task. So how was this not sufficient? While it may not be sufficient for your specific use case it is what worked for us. How was your 5900x not stable after using benchmarks?