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!

 

[Kocicak]

The fact that they didn't give a price is odd. I hope they are not waiting to see the public's response to determine where to price it.

It is in line with what I wrote, some of the 3D CPUs are probably not clearly winning overall, but just in perf/price ratio compared to Intel CPUs, so once it is clear how 13900KS performs and how much it costs, they will determine the price so that their products make sense and will be attractive to customers with Intel offerings in mind.

 

[exquisitechar]

I think the X3D CPUs and a tuned 13900K will probably perform similarly on average, each having significant wins in different games depending on how much the extra cache helps. Hope the pricing is good. A sanely priced 7800X3D with average DDR5 would be a not-so-expensive way to get top gaming performance.

 

[inf64]

From the PCworld's video: no direct OC will be possible but AMD will support both PBO and curve optimizer on all 3 X3D parts. I guess with good cooling and undervolting, it will be possible to push the clocks a bit further, hopefully on the Vcache die as well.

 

[JM Popaleetus]

From the PCworld's video: no direct OC will be possible but AMD will support both PBO and curve optimizer on all 3 X3D parts. I guess with good cooling and undervolting, it will be possible to push the clocks a bit further, hopefully on the Vcache die as well.

I feel like...other than to try sub-ambient coolers, set WRs or simply just because, PBO+200 and undervolting is all you can really do on Ryzen 5000/7000 anyway.

The boost algorithm already has these chips so close to the limit. Maybe you can eek out an extra 1% fine tuning each core individually, but daily driving all core OCs are a thing of the past.

 

[deasd]

The differentiate CCDs inside Zen4 X3D is a much more interesting solution than any other big-little design, I never thought CPU could work like this way. The last question is if AMD could schedule both 3D and non-3D cores flexibly, game with cache-intensive code use 3D cores, and non-3D high-frequency cores kicks in while switching back to less cache-intensive situation.

It would be a surprise for me if the different frequency of both CDDs inside 7900X/7950X was already a well developed plan by AMD which would be used in Zen4 X3D. Just wow.

 

[nicalandia]

It is in line with what I wrote, some of the 3D CPUs are probably not clearly winning overall, but just in perf/price ratio compared to Intel CPUs, so once it is clear how 13900KS performs and how much it costs, they will determine the price so that their products make sense and will be attractive to customers with Intel offerings in mind.

You are out of your mind if you think that the 13900KS will be faster than the 7950X3D and that AMD will be the Budget solution.

 

[nicalandia]

I suspected as much, this is really impressive


Also the Xeon Phi-HBM could be used without DIMM RAM,

 

[Kaluan]

I think the X3D CPUs and a tuned 13900K will probably perform similarly on average, each having significant wins in different games depending on how much the extra cache helps. Hope the pricing is good. A sanely priced 7800X3D with average DDR5 would be a not-so-expensive way to get top gaming performance.

This may come as a shocker to people, but you can tune Ryzen 7000 CPUs as well, and from what we've heard from the PCWorld video, 7000X3D won't have the same restrictions as 5000X3D anymore either.
Check Zen 4 builder's thread, particularly @Det0x 's results.

7700X is only 2-4% behind 13900K (haven't seen direct well done big game averages, but you can extrapolate based on stuff like Jarrod's or HUB's 13600K launch day and 7600X/7700X/13700K big-game-average head-to-heads).
If someone imagines 7000X3Ds will be only 5% (or slightly more) than their vanilla cohorts then IDK what to tell them, that's just crazy talk.

 

[Mopetar]

We dont know if the 5GHz boost on 7800x3d is cause of product segmentation or because v-cache related limitation. If its cause of v-cache, i think its pretty possible same limitation will concern 7950x as well.

Could be a little bit of both. Is the v-cache applied to all chiplets on the wafer before it's diced, or individually?

All silicon has some variation in performance characteristics, so I wouldn't be surprised if the same is true here.

I'll be interested in how good the CPU is at getting the threads that can use the extra cache on those cores. Same for putting other threads on the faster cores if we get a fast/slow mix.

Still not a fan of the marketing listing the total cache. Why not count the TLB, BTB, and micro-op caches in there as well?

 

[utahraptor]

I hope this does not end up requiring specific coding from each game publisher with regards to speed chiplet or cache chiplet thread selection. That would be a nightmare as almost no past titles would have the optimization and only some new ones and for how many years? One of two? If this can’t be fully automated then it could be very hit or miss. I guess it will come out in the benchmarks. At the end of the day If it’s faster across past and current games one day one that would seem to indicate it won’t require secret sauce in each game engine.

 

[eek2121]

You are out of your mind if you think that the 13900KS will be faster than the 7950X3D and that AMD will be the Budget solution.

I would honestly be shocked if Intel won any gaming benchmarks.

Side note: It is trivial to pin gaming workloads to a specific CCD. Just look and see if Direct3D is being actively used. From there, false positives would need to be filtered out (apps that use DirectX in creative ways, for example), along with games that perform better at high frequencies.

I do wish we had a way to easily pin these workloads ourselves, however. I would love to see an app that allows for custom scheduler profiles to be set up on an app-to-app basis.

 

[Timmah!]

I would honestly be shocked if Intel won any gaming benchmarks.

Side note: It is trivial to pin gaming workloads to a specific CCD. Just look and see if Direct3D is being actively used. From there, false positives would need to be filtered out (apps that use DirectX in creative ways, for example), along with games that perform better at high frequencies.

I do wish we had a way to easily pin these workloads ourselves, however. I would love to see an app that allows for custom scheduler profiles to be set up on an app-to-app basis.

Does not Process Lasso do that?

 

[inf64]

Does not Process Lasso do that?

I don't think the free version has such options. Maybe paid version has it.

 

[Kocicak]

You are out of your mind if you think that the 13900KS will be faster than the 7950X3D and that AMD will be the Budget solution.

I wrote that 7950X3D may be trading blows with 13900KS and it may be a few percent better.

I also wrote that 7800X3D may not be better than 13900K.

13900KS will be the competitors best gaming CPU and AMD is obviously waiting for its price and performance to set the prices, because none of the new AMD 3D CPUs will be able to wipe the floor with it.

 

[Hail The Brain Slug]

I would honestly be shocked if Intel won any gaming benchmarks.

Side note: It is trivial to pin gaming workloads to a specific CCD. Just look and see if Direct3D is being actively used. From there, false positives would need to be filtered out (apps that use DirectX in creative ways, for example), along with games that perform better at high frequencies.

I do wish we had a way to easily pin these workloads ourselves, however. I would love to see an app that allows for custom scheduler profiles to be set up on an app-to-app basis.

I was just thinking earlier this morning, it would be nice if Ryzen Master was updated to allow manipulating the scheduling, perhaps on a per-app basis. Could be as simple as select an app and choose whether V-Cache or Speed CCD is preferred. Also would be interesting if there were some other blanket options like "Prefer scheduling on VCache CCD" or "Prefer scheduling on speed CCD" for if you just want to ignore the custom scheduling behaviors.

If someone from AMD browses here, take notes.

 

[Vattila]

I do wish we had a way to easily pin these workloads ourselves, however. I would love to see an app that allows for custom scheduler profiles to be set up on an app-to-app basis.

You can set the affinity (i.e. select the logical processors to use) for a program by using simple Windows console commands, which you can put in a BAT file (people still know how to automate simple tasks with a BAT file these days, right?) or in a program shortcut file. For example, see:

Set affinity with start /AFFINITY command on Windows 7 - Stack Overflow

 

[Mopetar]

I wrote that 7950X3D may be trading blows with 13900KS and it may be a few percent better.

I also wrote that 7800X3D may not be better than 13900K.

You could also have written that you're the queen of Spain. It wouldn't be any less reasonable than anything else you've had to say in the last few pages.

There are still games where the 5800X3D beats everything else on the market, including the 13900K. Even if we ignore any improvements that Zen 4 brings, the 7800X3D is getting an 11% clock speed bump. In most large game sample averages, that puts it ahead on average.

If we consider the average uplift to be closer to 20%, then even if you take a small, lopsided selection of games, the 7800X3D is still likely to come out ahead. You'd probably need to limit yourself to UserBenchmark before you're going to find a set of numbers that've been sufficiently tortured to make the 13900K come out ahead on average.

Unless your definition of trading blow is the 13900K tossing a few light jabs to the body while it eats uppercut after uppercut until its getting coloring books for Christmas, I don't think your analysis is even close to the mark. You would have to go back to selling Manhattan island for a chest of beads to find a worse trade.

 

[Kaluan]

Seems AMD chose to go with games showing slightly lower performance increases in the actual CES presentation for both 7800X3D and 7950X3D, probably to not oversell/temper expectations. As opposed to some slides allegedly shown to the tech press beforehand:

7800X3D v 5800X3D CES:

7800X3D v 5800X3D tech press:

7950X3D v 13900K CES:

7950X3D v 13900K tech press:


Most are different titles. Either way, more data for us.

 

[Timmah!]

You could also have written that you're the queen of Spain. It wouldn't be any less reasonable than anything else you've had to say in the last few pages.

There are still games where the 5800X3D beats everything else on the market, including the 13900K. Even if we ignore any improvements that Zen 4 brings, the 7800X3D is getting an 11% clock speed bump. In most large game sample averages, that puts it ahead on average.

If we consider the average uplift to be closer to 20%, then even if you take a small, lopsided selection of games, the 7800X3D is still likely to come out ahead. You'd probably need to limit yourself to UserBenchmark before you're going to find a set of numbers that've been sufficiently tortured to make the 13900K come out ahead on average.

Unless your definition of trading blow is the 13900K tossing a few light jabs to the body while it eats uppercut after uppercut until its getting coloring books for Christmas, I don't think your analysis is even close to the mark. You would have to go back to selling Manhattan island for a chest of beads to find a worse trade.

To be fair, whether ones consider 20 percent performance difference big or small, its rather subjective. Perhaps to him is 20 percent "trading blows" - after all its IMO safe to say that in majority of cases those 20 percent would not be very noticeable.
That said, i would take 7950x3D over 13900KS for sure personally. But mostly cause i am not fan of big.little approach, significantly better power consumption and ability to replace it with Zen5/6. Definitely not because i would get 240 frames instead 200 in some random game.

 

[Exist50]

Seems AMD chose to go with games showing slightly lower performance increases in the actual CES presentation for both 7800X3D and 7950X3D, probably to not oversell/temper expectations. As opposed to some slides allegedly shown to the tech press beforehand:

7800X3D v 5800X3D CES:
View attachment 73990
7800X3D v 5800X3D tech press:
View attachment 73991
7950X3D v 13900K CES:
View attachment 73992
7950X3D v 13900K tech press:
View attachment 73993

Most are different titles. Either way, more data for us.

Might have been an attempt to avoid a repeat of the 7900 XTX situation, where they grossly overpromised.

 

[Joe NYC]

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.

Or go with 7800x3d and stick with Windows 10.

 

[Joe NYC]

@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?

Unlikely.

The set of games that benefit from > 16 threads is quite low, so no reason to jump from the "gaming CCD" to the other CCD at all, in vast majority of games.

 

[Joe NYC]

@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.

"Severe"? Very strange, undefined metrics?

How about this metrics: Penalty for guessing wrong. In case of 7950x3d, the penalty is nearly zero. In case of Alder Lake, Raptor Lake? This is where you can use the word severe.

 

[Joe NYC]

@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.

"Severe"? In a strange, undefined metrics?

How about this metrics: Penalty for guessing wrong. In case of 7950x3d, the penalty is nearly zero. In case of Alder Lake, Raptor Lake? This is where you can use the word severe.

 

[Joe NYC]

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.

The cringe Space Cadet from Microsoft and cringe astronaut, I thought that this was the worst AMD keynote...

But then came the Mi300 tease, and all was forgiven...