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!

 

[TESKATLIPOKA]

From Anandtech's article

This is an interesting idea. They save up one cache die and this move doesn't affect clocks.

 

[poke01]

You mean that Intel, who can actually SUPPLY chips to a wide range of product segments in the laptop market, will be affected by AMD's limited ability to supply the OEM market which historically hasn't been their focus until recently?

yep

 

[TESKATLIPOKA]

You mean that Intel, who can actually SUPPLY chips to a wide range of product segments in the laptop market, will be affected by AMD's limited ability to supply the OEM market which historically hasn't been their focus until recently?

Please, stop embarrassing yourself by bringing up the production capacity.
He was talking about performance and efficiency, and you know It.

Intel has a serious problem with U and P series.
7940HS is 34% faster in CB R23 than i7-1280P and Raptor will be barely faster, because It's the same chip only a bit higher clocked.

I know, we don't have the actual TDP for them only official 35-45W vs 28W and that U series would make a better comparison.
Still, at comparable TDP an 8C16T Phoenix should be faster than 6P8E20T i7-1370P.

 

[biostud]

10% higher clock +13% ipc, the 7800X3D should be around 20% faster than the 5800X3D.

 

[Kocicak]

Well, the slides with 7800X3D vs. 5800X3D gaming comparison are different in the press deck and presentation. In the press deck are 20-30 % improvements and in the presentation 10-25%. When Mrs. Su said that 7800X3D is on average by 15% better than 5800X3D, the crowd was silent, she had to cheer them up.

It seems that 7800X3D will not be better than 13900K in gaming, but that should not be a huge problem given the price difference.

 

[JayMX]

When Mrs. Su said that 7800X3D is on average by 15% better than 5800X3D, the crowd was silent, she had to cheer them up.
It seems that 7800X3D will not be better than 13900K in gaming, but that should not be a huge problem given the price difference.

However, for that kind of performance difference in GPUs, there is a price difference of several hundred dollars (4070Ti vs 4080, 15% performance difference, $400 price difference). And that will undoubtedly not be reflected in 7950X3D (vs 13900k).

 

[Exist50]

They say 14% in one of their press releases.

These new Ryzen 7000X3D are the fastest gaming processors in the world.4 With up to 14% faster performance than the previous generation.6 The AMD Ryzen 7000 Series processors with 3D V-Cache technology will be available for Socket AM5 starting February 2023.

Footnotes seem broken, however, so can't tell what they're comparing.

AMD Extends its Leadership with the Introduction of its Broadest Portfolio of High-Performance PC Products for Mobile and Desktop

— AMD Ryzen™ 7000X3D Series Desktop processors bring the power of AMD 3D V-Cache™ technology to gamers and creators — — AMD Ryzen 7000…...

ir.amd.com

 

[Exist50]

No mention of overclocking on the 7800X3D's page. https://www.amd.com/en/product/12731/
For contrast, the 7950X does explicitly mention it. https://www.amd.com/en/product/12151

 

[BorisTheBlade82]

I think it's even more complicated. Frequency vs cache sensitivity and scheduling decisions would require some really deep telemetry. Nothing like that on the market today, as far as I'm aware.

Totally agree with you here. It's going to be a nightmare. If I understand the statement in the THG article correctly, we will get Gaming chipset drivers that will be updated frequently in order to explicitly specify which strategy to apply for which Game.
The problem I have with this approach is that quite likely only AAA titles will be explicitly optimized and at some point in the future they will simply stop integrating new games.

 

[RnR_au]

No mention of overclocking on the 7800X3D's page. https://www.amd.com/en/product/12731/
For contrast, the 7950X does explicitly mention it. https://www.amd.com/en/product/12151

Not mentioned on the 7950X3D page either - https://www.amd.com/en/product/12741 - which could mean that you won't even be able to overclock the non-vcache die.

 

[tamz_msc]

Please, stop embarrassing yourself by bringing up the production capacity.
He was talking about performance and efficiency, and you know It.

Intel has a serious problem with U and P series.
7940HS is 34% faster in CB R23 than i7-1280P and Raptor will be barely faster, because It's the same chip only a bit higher clocked.

I know, we don't have the actual TDP for them only official 35-45W vs 28W and that U series would make a better comparison.
Still, at comparable TDP an 8C16T Phoenix should be faster than 6P8E20T i7-1370P.

The only problem Intel has with efficiency is the high idle power draw caused by the PCH consuming too much power in the Alder Lake-P CPUs. That should be fixed with Raptor Lake-P. Everybody knows that AMD's specs for TDP are meaningless when it comes to laptops, as these CPUs have a short duration power limit over and above TDP that lasts for much longer than tau on Intel CPUs, so no comparison at similar power levels is possible, especially between HS and P SKUs.

 

[Exist50]

Totally agree with you here. It's going to be a nightmare. If I understand the statement in the THG article correctly, we will get Gaming chipset drivers that will be updated frequently in order to explicitly specify which strategy to apply for which Game.
The problem I have with this approach is that quite likely only AAA titles will be explicitly optimized and at some point in the future they will simply stop integrating new games.

Yeah, I have to figure they'd use a whitelist of sorts. Maybe they could have some heuristics, but how else would they even know that a given app is a game? Would be interesting for someone to try out exe renaming and see if it has an effect.

Though if these optimizations are W11-exclusive, it'll be fun watching some of the outlets who've defended sticking with W10 suddenly proclaim W11 is a necessity.

 

[Exist50]

It's been talked about at length, but this naming scheme is absurd.

 

[BorisTheBlade82]

@Exist50
To add even more to this mess: Not only will they have to decide on a per-game-level. They will have to decide on a per-thread-of-game level.
Is it even possible to address specific threads as in: This is a cache sensitive thread of Game x and that is a frequency sensitive thread of the same game?

 

[Timorous]

Unbalanced cache on the dual CCD X3D products.

Seems like a precursor to having something akin to big little with Zen 5 where you have 8 Zen5 cores with v-cache and 16 Zen 5c cores without for the 8950X(3D).

Something I suggested yonks ago. So long as games are dumped onto the high cache CCD it should work fine.

 

[BorisTheBlade82]

@Timorous
Generally yes. But for 7000X3D the problem is more severe: There is no universally better core, but two kinds with specific advantages and disadvantages. That makes it a lot harder for the scheduler than just BIG.little.

 

[Exist50]

@Exist50
To add even more to this mess: Not only will they have to decide on a per-game-level. They will have to decide on a per-thread-of-game level.
Is it even possible to address specific threads as in: This is a cache sensitive thread of Game x and that is a frequency sensitive thread of the same game?

Even if it's possible, I doubt there's any benefit to per-thread CCX scheduling. As things stand, gaming shows a strong preference for all significant threads to be bound to the same CCX / L3 domain. I think the quick-and-dirty solution would be to simply bind whitelisted games/apps to the V-cache CCX and let anything else schedule freely (likely favoring the high frequency CCX).

I don't have any data to back this up, but I imagine it would often be better for a game to incorrectly bind to the non-Vcache CCX than to split its threads between the two.

 

[lopri]

100%

If it was referring to Windows 11 is only supported, AMD's site notes both Win10 and Win11.

https://www.amd.com/en/products/apu/amd-ryzen-9-7950x3d

And Ubuntu. XD

 

[Exist50]

If it was referring to Windows 11 is only supported, AMD's site notes both Win10 and Win11.

https://www.amd.com/en/products/apu/amd-ryzen-9-7950x3d

And Ubuntu. XD

The CPU will surely work in all the typical OSs, but will these "optimizations" apply universally? The fact that they've spoken of Windows drivers seems to imply otherwise.

 

[coercitiv]

It's been talked about at length, but this naming scheme is absurd.

Who's affected by this? I don't like it either, having to retrain to prioritize for the third digit feels so random. But the average consumer doesn't care, just like they shouldn't care that half of Raptor Lake is using Adler Lake cores.

 

[coercitiv]

AMD fails twice in a row with their keynotes, first with RDNA3 and now this. They have such conflicting information in those graphs that I feel they're not worth my time. Either the product is weak, or the presentation of the product is extremely weak.

Meh.

 

[BorisTheBlade82]

Even if it's possible, I doubt there's any benefit to per-thread CCX scheduling. As things stand, gaming shows a strong preference for all significant threads to be bound to the same CCX / L3 domain. I think the quick-and-dirty solution would be to simply bind whitelisted games/apps to the V-cache CCX and let anything else schedule freely (likely favoring the high frequency CCX).

I don't have any data to back this up, but I imagine it would often be better for a game to incorrectly bind to the non-Vcache CCX than to split its threads between the two.

Yeah, I think you are right. Hadn't thought about inter-thread communication.

 

[Exist50]

Who's affected by this?

Anyone who buys old tech thinking it's new. I.e. anyone who doesn't follow this stuff as closely as we do.

But the average consumer doesn't care, just like they shouldn't care that half of Raptor Lake is using Adler Lake cores.

The average consumer would definitely notice the difference in performance, features, battery life, etc.

As for Raptor Lake, it's looking like only the desktop is true rebrands, but either way, at least higher number == better, and there're no significant feature differences. Of course, still not ideal, but not nearly this bad.

 

[Timorous]

Even if it's possible, I doubt there's any benefit to per-thread CCX scheduling. As things stand, gaming shows a strong preference for all significant threads to be bound to the same CCX / L3 domain. I think the quick-and-dirty solution would be to simply bind whitelisted games/apps to the V-cache CCX and let anything else schedule freely (likely favoring the high frequency CCX).

I don't have any data to back this up, but I imagine it would often be better for a game to incorrectly bind to the non-Vcache CCX than to split its threads between the two.

If I was AMD I would have profiles in Ryzen Master to assign threads to the ideal CCD for certain workloads such as gaming.

That would mean CS:GO can be assigned to the high frequency CCD and MSFS can be assigned to the high cache CCD.

For stuff that scales beyond 8c it will be trickier to manage.

 

[coercitiv]

The average consumer would definitely notice the difference in performance, features, battery life, etc.

We were talking about the naming scheme. When it comes down to performance / features / battery life, we would need to see the balance between perf numbers & prices.

As for Raptor Lake, it's looking like only the desktop is true rebrands, but either way, at least higher number == better

In the AMD naming scheme it's the same, higher number == better.

The really annoying part about AMD plan is the first digit, since they claim it signifies the year. If they run the first digit out of sync with the desktop platform, nasty things will happen in a couple of years. But then again, maybe they ditch the new scheme in less than that