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!

 

[Saylick]

Here is my estimate for top Ryzen 7000 series 16C/32T part, similar to the ones I made before for Zen1/Zen3.

Top 16C/32T SKU 170W: ~14% higher ST clock vs 5950X (4.95GHz->5.65GHz); ~22% higher all-core boost clock( 4.2GHz*->5.15GHz); ~13%** higher IPC vs Zen3 (geo-mean; ~12% integer, 15% in FP).

I really hope you are correct with your IPC guess. I thought it would be super easy for AMD to just deepen the buffers across the board from Zen 3 to Zen 4 to pick some low-hanging fruit (kind of like how Intel does it), but I'm not even sure we're getting that with Zen 4. If we aren't getting that at all and it's simply just 2x L2 and 1.5x L2 TLB as the main architectural improvements (not counting AVX512 here, I see that as more of a feature set improvement) then I'm not sure double digit IPC gains are possible. Usually when we see double digit IPC gains, pretty much everything across the board sees some kind of improvement, either it be deeper buffers, larger caches, or additional execution ports. Again, I hope you're right and I am wrong, but I guess we'll just have to wait and find out.

 

[inf64]

I really hope you are correct with your IPC guess. I thought it would be super easy for AMD to just deepen the buffers across the board from Zen 3 to Zen 4 to pick some low-hanging fruit (kind of like how Intel does it), but I'm not even sure we're getting that with Zen 4. If we aren't getting that at all and it's simply just 2x L2 and 1.5x L2 TLB as the main architectural improvements (not counting AVX512 here, I see that as more of a feature set improvement) then I'm not sure double digit IPC gains are possible. Usually when we see double digit IPC gains, pretty much everything across the board sees some kind of improvement, either it be deeper buffers, larger caches, or additional execution ports. Again, I hope you're right and I am wrong, but I guess we'll just have to wait and find out.

I'm pretty sure that AMD can do double digit IPC gains with Zen4, they have much larger transistor budget. Some of it was spent on 2xL2 and AVX512 but there should still be a lot of room left to improve the other parts of the core. They managed ~19% IPC uplift (verified by Ian in his own deep dive) on the same 7nm node as Zen2, so I think that 10-15% IPC uplift is an "easy" target for them on 5nm (even if they lengthened the pipeline a bit) .

 

[biostud]

Here is my estimate for top Ryzen 7000 series 16C/32T part, similar to the ones I made before for Zen1/Zen3.

Top 16C/32T SKU 170W: ~14% higher ST clock vs 5950X (4.95GHz->5.65GHz); ~22% higher all-core boost clock( 4.2GHz*->5.15GHz); ~13%** higher IPC vs Zen3 (geo-mean; ~12% integer, 15% in FP).

Gaming: 1.28x or 28% better performance vs 5950X (~12% ipc x ~14% higher effective clock)
View attachment 62140

ST workloads: 1.29x or 29% better ST performance vs 5950X (~13% ipc x ~14% higher effective clock)
View attachment 62141

MT workloads: 1.38x or 38% better MT performance vs 5950X (~13% ipc x ~22% higher effective clock)
View attachment 62142

*4.2GHz is commonly reported all-core boost for 5950X but there are workloads where it will not boost beyond 3.9GHz.
**Initially I wanted to say 15% higher avg IPC, but lowered it a bit in light of very high clock speed targets. 15% in FP gen-on-gen is a trend so far as shown in AT deep dive article here. I think that integer-specific IPC jump will be somewhat lower with Zen4 and this will drag the average down a bit

EDIT: If there is going to be a huge miss when it comes to my estimate, it's likely going to be gaming section. It's possible there is some sort of bottleneck(GPU?) that limits the performance so the scaling will be much lower, similar to how 12900KS has 21% higher ST performance vs 5950X but in games it's only 8% faster.

Also depends on if the specific MT workload is memory bandwidth limited. Going from DDR4 to DDR5 can see extra performance increase.

 

[carrotmania]

I still firmly believe that AMD still owes us long term customers. More than 10 years of darkness and misery with worthless AMD CPU's. I personally have a desk drawer full of my very old AMD CPU's from the golden era of AMD. Starting with the Athlon 1800+ to the dual Core AMD 64's.

I should point out that I currently have AMD Ryzen 5600. That processor was released in April 2022 and I am getting grief from some AMD fanboys in this thread. It should have been released in early 2021.

AMD doesn't owe you a thing. They had a massively reduced RnD budget because Intel screwed them over, not in Core2, cos one or two lines of CPU doesn't break a company, but in taking away, illegally, their customers. This has been proven in a court of law. They are still working with a reduced budget, and are having to cover the sectors of business that will make them profitable long term, so quit with the "I say they should have had 'x' product out when I demand".

So, you SAY you have an AMD machine, but every post you make it anti-AMD. Go cry to Intel, they are the reason technology stalled for a decade, and are still the reason that AMD is having to pick and choose their customers. But I bet that just washes over you, huh?

 

[eek2121]

https://www.tomshardware.com/news/amd-corrects-socket-am5-for-ryzen-7000-power-specs-230w-peak-power-170w-tdp

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

Nobody stated the chip was at 230W PPT. I am inclined to think the chip was at 125W (170W ppt)

We won’t know until more details drop, but N5 should allow for those clocks at very low power levels.

I keep asking myself the same question for Bergamo, Where in the heck are they going to place 16C/32T Chiplets on the same substrate?

View attachment 62146

So The Dragon Range 24C and any Higher than 16 Core Ryzen CPU will use the Zen4c cores that allow 16C/32T per chiplet.

It is also possibly to redesign the package and flip the chips 90 degrees. Then AMD could have 16 Zen4 + 16 Zen4c (3 chiplets).

They also could move to N4 or N3.

I'm pretty sure that AMD can do double digit IPC gains with Zen4, they have much larger transistor budget. Some of it was spent on 2xL2 and AVX512 but there should still be a lot of room left to improve the other parts of the core. They managed ~19% IPC uplift (verified by Ian in his own deep dive) on the same 7nm node as Zen2, so I think that 10-15% IPC uplift is an "easy" target for them on 5nm (even if they lengthened the pipeline a bit) .

The focus was likely on power management and frequency because they were moving to a new process. There clearly is some IPC uplift, but I suspect that was not the core focus.

 

[coercitiv]

nvm, I see it's already been discussed...

 

[Hans Gruber]

AMD doesn't owe you a thing. They had a massively reduced RnD budget because Intel screwed them over, not in Core2, cos one or two lines of CPU doesn't break a company, but in taking away, illegally, their customers. This has been proven in a court of law. They are still working with a reduced budget, and are having to cover the sectors of business that will make them profitable long term, so quit with the "I say they should have had 'x' product out when I demand".

So, you SAY you have an AMD machine, but every post you make it anti-AMD. Go cry to Intel, they are the reason technology stalled for a decade, and are still the reason that AMD is having to pick and choose their customers. But I bet that just washes over you, huh?

I take it that this is your alt account.

 

[uzzi38]

https://www.reddit.com/r/Amd/comments/uy7ep8/_/ia3ohke

So AMD just aren't telling us what power it was running at. I guess I'll try asking in the HotHardware interview just to be safe, but this really isn't helpful.

:/

 

[moinmoin]

Yeah, definitely. It was brought up a few days ago by Skyjuice, who's articles you should probably be reading (seriously, follow him on Twitter).

Anyway, the Gigabyte documentation is old, and so the values aren't really up to date, but you can still see a 120W group here too.

https://media.discordapp.net/attachments/807442045068640296/979383271769112586/AMD-AM5-Socket-TPD.png

Thanks for reposting that table:

It's interesting how 170W TDP deviates from the other levels, requiring liquid cooling. I wonder if that may be an indication for a new higher end overclocking platform, not necessarily chips that actually require that TDP. It could be the inverse counterpart to ECO mode, opening up the TDP limit for MT heavy use cases. Think 7950X being a 120W TDP chip, but you want to make the most of its MT capability? Get a board including a liquid cooling solution that allows to use its 170W TDP LIQUID OC mode on it. That's how I'd handle it anyway. New market for board OEMs, less fiddling for consumers wanting to get the most out of MT without going manual OC.

This trend is awful since 20 years ago we got <40W Radeon R9700 paired with 70W Athlon XP. Now you run a 300W GPU with 150+W CPU. We are heading towards purposedly air-conditioned "computer rooms" again. It's the 70s again.

The industry is successfully normalizing expectation for higher performance that is only achievable at the cost of more and more energy.

That's mostly fine with desktop, but I find this particular development infuriating in the mobile space.

 

[Hitman928]

https://www.reddit.com/r/Amd/comments/uy7ep8/_/ia3ohke

So AMD just aren't telling us what power it was running at. I guess I'll try asking in the HotHardware interview just to be safe, but this really isn't helpful.

:/

Because the link didn't work well for me for some reason:



Another reply by Robert Hallock:



So Ryzen 7000 will have up to a 170W TDP / 230W PPT and the demo CPU was using something lower than 230W. Definitely AMD is willing to push further out on the f/v curve and is OK with getting worse efficiency with Zen4 Ryzen which is disappointing to hear. I'll wait for final product to judge, but it seems AMD is happily following Intel's lead in pushing power to the edge to get any bit of performance they can.

Edit: Robert Hallock and AMD technical marketing in general needs to do better.

 

[carrotmania]

I take it that this is your alt account.

I'll add that to the list of things you're wrong about

 

[Paul98]

I am happy to hear this, gives them more room at the higher end, especially if they want to add more cores and high clocks, plus they now have that GPU integrated. I expect most designs won't be max power, but good to have the headroom for top end chips.

 

[deasd]

This is a trend, maybe we will soon seeing a ~400watts CPU.
Anyway the AMD's marketing material is a mess.... 'Prototype sample' ....what does it mean? What I could understand is Engineering Sample....

 

[nicalandia]

Another reply by Robert Hallock:

So Ryzen 7000 will have up to a 170W TDP / 230W PPT and the demo CPU was using something lower than 230W. Definitely AMD is willing to push further out on the f/v curve and is OK with getting worse efficiency with Zen4 Ryzen which is disappointing to hear.

So you are telling me that at "Pre Production" 16C sample it's hitting 40%+ Performance boost on MT while not even pushed to the limits? Well color me Impressed because the Final/Retail product will kick some serious butt in MT.

All the while people cry about lower than Zen3 efficiency and the competition will have worst efficiency..

 

[nicalandia]

This is a trend, maybe we will soon seeing a ~400watts CPU.
Anyway the AMD's marketing material is a mess.... 'Prototype sample' ....what does it mean? What I could understand is Engineering Sample....

Prototype is an Engineering Sample. Not the first time they use the Word "Prototype" They used that on the 5900X3D Prototype that never saw the light of day as a production sample.

 

[Karnak]

I am happy to hear this, gives them more room at the higher end, especially if they want to add more cores and high clocks, plus they now have that GPU integrated. I expect most designs won't be max power, but good to have the headroom for top end chips.

Yeah remember AM5 is just not there for Zen4. Probably even Zen6 in the future. I'am 100% certain that a 16C AM5 CPU like we're getting with Zen4 for now is not the maximum for this socket. They'll go for 24C, maybe even a 32C. And for that it doesn't make any sense to limit the power draw at a similar low level as AM4 because even the 5950X is held back by that 142W limit pretty hard without ruining the efficiency (ofc it get's worse the higher you get, but it's not at a '20% higher power draw means only 5% higher performance' level).

 

[uzzi38]

AMD Answers Our Zen 4 Tech Questions, with Robert Hallock

AMD just announced their new Zen 4-based Ryzen 7000 desktop processors. We sat down with AMD's Robert Hallock and asked him a load of technical questions to learn more about what to expect from the new platform, its highlights, and capabilities.

www.techpowerup.com

And here's a technical interview with TPU with a couple of juicy details in there

 

[Hitman928]

So you are telling me that at "Pre Production" 16C sample it's hitting 40%+ Performance boost on MT while not even pushed to the limits? Well color me Impressed because the Final/Retail product will kick some serious butt in MT. All the while people cry about lower than Zen3 efficiency while the competition will have worst efficiency..

Well, we don't know if it was using 5 W less or 50 W less so it's a reply that doesn't actually clarify much other than it was almost assuredly above 170 W. Either way, I don't expect performance to increase much even though the CPU was using less than it's full power target as you are so far into the inefficient part of the curve that small increases in performance take overly large amounts of power increases. That is, unless AMD marketing is feigning incompetence to hardcore sandbag everyone but I really doubt this is the case. Don't get me wrong, Zen 4 will still be a very strong CPU and most likely the best money can buy in almost every way upon release and potentially even after RPL drops, but I would rather see them use their node advantage to really work on IPC and keeping power efficiency as the dominant feature over their competition, rather than follow Intel with ever faster, more power hungry cores.

 

[ondma]

The reason AMD would increase the TDP on the AM5 socket. To better handle 24core 48thread and 32core 64thread CPU's that AMD will inevitably release during the Zen 4 era at some point.

They may well release higher core count chips, but I see no indication of that from the statement that you quoted. He is referring to increasing performance of existing chips that were power limited.

 

[Hitman928]

AMD Answers Our Zen 4 Tech Questions, with Robert Hallock

AMD just announced their new Zen 4-based Ryzen 7000 desktop processors. We sat down with AMD's Robert Hallock and asked him a load of technical questions to learn more about what to expect from the new platform, its highlights, and capabilities.

www.techpowerup.com

And here's a technical interview with TPU with a couple of juicy details in there

I saw that the 15% ST number was them being conservative (which I think isn't much of a surprise, we just want to know how conservative it is), that 5.5 GHz was 'easy' for them (not sure what that really means), and a confirmation that the touted AI acceleration refers to AVX-512 NN/DL features. Did I miss anything else significant?

 

[nicalandia]

Did I miss anything else significant?

Yeah, Silver Plating..


In the photos, the compute dies look gold-plated?
"They are not gold-plated, it's a process called "back-side-metallization" that we use to solder the dies to the heatspreader. Depending on how that's manufactured, it can refract light in different colors [like the surface of a DVD], and in this case it refracts in gold"

What he "Forgot" to mention is that it is Silver Plating (Thru Physical vapor deposition)

Backside Metallization
High performance computing will generate high volume heat during calculation. Backside metallization are with deposition of metal layers on the bare backside silicon to enhance its thermal conductivity.

Winstek

www.winstek.com.tw

Edit. it's actually Silver Ag instead of Gold Au, which is still impressive as it's very expensive. That speaks well for the High Performance nature of Zen4

 

[Hitman928]

Yeah, Gold Plating..

Robert Hallock is Full of it....

In the photos, the compute dies look gold-plated?
"They are not gold-plated, it's a process called "back-side-metallization" that we use to solder the dies to the heatspreader. Depending on how that's manufactured, it can refract light in different colors [like the surface of a DVD], and in this case it refracts in gold"

What he "Forgot" to mention is that it is Gold Plating (Thru Physical vapor deposition)

Backside Metallization
High performance computing will generate high volume heat during calculation. Backside metallization are with deposition of metal layers on the bare backside silicon to enhance its thermal conductivity.

View attachment 62159

Winstek

www.winstek.com.tw

Ag is silver, not gold (gold is Au). I'm assuming you got confused because of the gold coloring in the cross-section drawing. There are also different metalization techniques using different combinations.

 

[nicalandia]

Ag is silver, not gold (gold is Au). I'm assuming you got confused because of the gold coloring in the cross-section drawing. There are also different metalization techniques using different combinations.

It has a yellow color tone to it. Perhaps Titanium Dioxide layer behind the super thin Silver layer?

 

[leoneazzurro]

that 5.5 GHz was 'easy' for them (not sure what that really means)

That there is quite probably headroom to increase clock even more (at the expense of power budget, of course). That is, he's hinting at easy overclock.

 

[Hitman928]

It has a yellow color tone to it. Perhaps Titanium Dioxide layer behind the super thin Silver layer?

Yeah, not a chemist and I don't know how all the different metals and alloys reflect light, especially when very thin, but I do know you get some unexpected colors sometimes when applying very thin layers of metals like this.