Discussion Zen 7 speculation thread

[Joe NYC]

Sounds interesting, so what does this mean in practice. All Zen7 CPUs will have some V-Cache (or other type of similar cache?), but smaller in size than on X3D SKUs?

Not all Zen 7 SKUs. The only one shown with this design would be the dense server CCD.

It means no L3 in the main compute die, and L3 access always going across to the other die.

Going from current Zen6 Dense die (which does not have a V-Cache option), comparable Zen7 Dense SKU would have 2x L3 (~8MB), less of the very expensive die (A14) wasted on SRAM.

So a win-win for that SKU, but this is not transferable to other Zen 7 SKUs.

Also, is it not using traditional X3D tech, but instead something more similar to NVL-S bLLC?

It is using same hybrid bond. The similarity with NVL bLLC is that the extra cache is always part of the unit, unlike V-Cache being optional.

But a 2 die solution, where 1 die is expensive and 2nd die is cheaper may offer some additional rebalancing.

There may be opportunities for this type of division of labor in client, but nothing so far has been leaked as far a stacking.

 

[Joe NYC]

Because you are assuming all games only use max 8C, 4-ever? What about those who use more, now and in the future?

Ther is overhead in being on 2 CCDs and in splitting the L3. So, the overhead eats a lot of the upside that 9950x3d2 would offer.

When AMD goes to 12 fore CCD and 16 core CCD then there would not the one step back before taking 1-2 steps forward, as far as utilizing > 8 cores.

And if not targeting gamers, what would 9950X3D2 be targeting?

Same people who think they need 9950x. 9950x3d2 will deliver more multicore performance.

If 9800x3d outperforms 9700x (both 8 core) in workstation type tasks by ~11% (measured by Phoronix), 9950x3d2 should deliver similar performance increment over 9950x.

If the 9950x3d2 actually materializes.

 

[MS_AT]

Because you are assuming all games only use max 8C, 4-ever? What about those who use more, now and in the future?

Could you please list those games? I am not saying you are wrong, but you don't quantify the rate at which game become more effective at utilising more cores. I mean if games using 16+ cores become more popular in 5-6 years time, then buying 16 core CPU for gaming today is waste of money as it is much more likely in 5 years time you will be able to buy effectively better CPU for lower price.

The same thing applies to LLC. What is the average working set size of current games and how does it evolve? If it hovers at <= 96MB then getting twice, or thrice the cache will be wasteful. You can observe this with some benchmarks, where basically L3 size does not matter, because the code and working set size fit in L2 or L1 even.

Therefore if you claim these things (more cores and more than 100MB of LLC) will be useful for games, it would help your position if you could back this up with some data to spark a discussion on something more concrete.

 

[DrMrLordX]

Games can't really use more than a single CCD effectively.

Would having two uniform vcache CCDs help with thread migration penalties we've seen on earlier X3D CPUs (notably the 7950X3D)? It's a stretch but that's the only reason why I could see having the second CCD have vcache as being beneficial.

So they'll be lagging behind NVL-S 288 MB bLLC?

Probably not. If Nova Lake is still just Coyote Cove, it's not going to stack up all that well against Zen6, and bLLC will probably be slower/higher latency than L3.

That, if it's made, will lose to a single CCD with vcache. For games you want a single CCD with a slab of vcache. That's it. One ccd with one vcache slice please. That's all gamers need. That's it.

Probably so, yes.

 

[adroc_thurston]

Would having two uniform vcache CCDs help with thread migration penalties we've seen on earlier X3D CPUs (notably the 7950X3D)?

no, you're still doing slowass snoops between CCD so if your game actually uses 16 cores, you're outta luck there.

 

[Fjodor2001]

I mean if games using 16+ cores become more popular in 5-6 years time, then buying 16 core CPU for gaming today is waste of money as it is much more likely in 5 years time you will be able to buy effectively better CPU for lower price.

I guess it depends on how long you intend to keep your PC. Not everyone buys a new PC every 2-3 years.

There’ll likely also be a gradular shift, and not a jump directly from 8C to 16+ C being used. So 12C you might see more commonly used already in 2-3 years.

It’s a chicken and egg scenario. Having more cores widely available in PCs will drive game development companies to put more effort into using them in their games. Same for other SW too by the way.

 

[adroc_thurston]

It’s a chicken and egg scenario. Having more cores widely available in PCs will drive game development companies to put more effort into using them in their games. Same for other SW too by the way.

PCs do not drive gamedev targets at all.

 

[511]

PCs do not drive gamedev targets at all.

Idk of you have talked to any game devs but they do having talked to couple of game dev.

 

[MS_AT]

I guess it depends on how long you intend to keep your PC. Not everyone buys a new PC every 2-3 years.

Well future proofing is hard😉 If you bought 3950x for gaming cause it has 16 cores, would you be happy customer today? Or in 3 years time?

There’ll likely also be a gradular shift, and not a jump directly from 8C to 16+ C being used. So 12C you might see more commonly used already in 2-3 years.

I don't deny it but unless you provide some input data it is hard to even make an educated guess as to the timeline when greater core counts will be popular and how this ties with the core count advancement of consoles.

Having more cores widely available in PCs will drive game development companies to put more effort into using them in their games. Same for other SW too by the way.

Well, but that won't happen because enthusiasts buy 600$ cpus but because 16c cpus will eventually cost 300 or 400$ so more people can afford them. Since PC games are tied closely to console games, console have to make the move first so the baseline becomes 8+ cores.

 

[adroc_thurston]

Idk of you have talked to any game devs but they do having talked to couple of game dev.

We're talking CPU baselines.
Graphics targets scale up and down relatively well. CPU stuff? not so much.

 

[511]

We're talking CPU baselines.
Graphics targets scale up and down relatively well. CPU stuff? not so much.

they do matter as well cause what hardware to target they have to do certain optimizations to make it fast on older hardware

 

[Joe NYC]

Well future proofing is hard😉 If you bought 3950x for gaming cause it has 16 cores, would you be happy customer today? Or in 3 years time?

3950x turned out to be a big fail as futureproofing for future games. So did 5950x, 7950x, and likely 9950x

I don't deny it but unless you provide some input data it is hard to even make an educated guess as to the timeline when greater core counts will be popular and how this ties with the core count advancement of consoles.

Forget more cores, how about fewer cores?

Suppose AMD sold 9600x3d with cock speed equal to 9800x3d. There would be a lot of ties in gaming tests.

 

[Fjodor2001]

3950x turned out to be a big fail as futureproofing for future games. So did 5950x, 7950x, and likely 9950x

Because no X3D.

 

[Joe NYC]

Because no X3D.

Even without x3d.

8 core N+1 gen is typically faster in gaming than 16 core N gen.

 

[adroc_thurston]

they do matter as well cause what hardware to target they have to do certain optimizations to make it fast on older hardware

Nonsense because Zen2 is still a very modern ISA target.

 

[Fjodor2001]

8 core N+1 gen is typically faster in gaming than 16 core N gen.

8 core N+1 gen is also typically faster than 8 core N gen. And:
16 core N gen is typically faster than 8 core N gen.

 

[511]

Nonsense because Zen2 is still a very modern ISA target.

which is same as Haswell rofl

 

[511]

Suppose AMD sold 9600x3d with cock speed equal to 9800x3d. There would be a lot of ties in gaming tests.

nice typo lol

 

[adroc_thurston]

which is same as Haswell rofl

Aka x86-64 v3 and we haven't moved from there because of Intel.
It's not AMD's fault Intel refused to ship newer SIMD ISAs until NVL.

 

[511]

Aka x86-64 v3 and we haven't moved from there because of Intel.
It's not AMD's fault Intel refused to ship newer SIMD ISAs until NVL.

Ahh yes they messed up AVX-512 but they are the biggest reason for it's existence not to mention NVL give us APX/AVX 10.2 and Zen 6 will be stuck with ISA of server GLC (except AMX) so the same thing applies to AMD as well

 

[maddie]

nice typo lol

Well its "F"PS after all, if you know what I mean.

 

[Thunder 57]

Ahh yes they messed up AVX-512 but they are the biggest reason for it's existence not to mention NVL give us APX/AVX 10.2 and Zen 6 will be stuck with ISA of server GLC (except AMX) so the same thing applies to AMD as well

But but but my APX! Won't make a difference for awhile. Kind of sounds like AMD64. AMD has had AVX-512 it was Intel who borked it up and needed a new version. So quit bitching about Haswell in a Zen 7 thread (that shouldn't exist yet).

 

[511]

But but but my APX! Won't make a difference for awhile. Kind of sounds like AMD64. AMD has had AVX-512 it was Intel who borked it up and needed a new version. So quit bitching about Haswell in a Zen 7 thread (that shouldn't exist yet).

i am not bitching about Haswell it's just that we are stuck with it for a while anyway APX is very nice you are free to look online here is something i found online a 1.5% perf improvement on Intel SDE

https://github.com/dotnet/runtime/pull/111072

 

[HurleyBird]

PCs do not drive gamedev targets at all.

Eh. They do, but not on multi-plats, and not so much on the CPU side, and of the CPU side, not so much on the MT side, and of the MT side, more "can I parallelize this task through some high level abstraction and will it make things faster?" than "how do I squeeze every drop out of every core?"

Most games will have some degree of implicit optimization for the developer's computer.

 

[adroc_thurston]

Eh. They do, but not on multi-plats, and not so much on the CPU side, and of the CPU side, not so much on the MT side, and of the MT side, more "can I parallelize this task through some high level abstraction and will it make things faster?" than "how do I squeeze every drop out of every core?"

Ergo, they don't.
Long live the console.