Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[gdansk]

Why do you think 9950X3D2 will perform worse than 9800X3D?

Why wouldn't it? NUMA is gamer poison. It'll perform better at 720p with an entire CCD disabled like every dual CCD chip before it.

 

[deasd]

Dell clarifies Alienware Area-51 Desktop will be configured with 9850X3D instead of the rumored 9950X3D2. Looks like a last-minute change right before CES.

https://twitter.com/x/status/2009289235418009641

 

[Fjodor2001]

Why wouldn't it? NUMA is gamer poison. It'll perform better at 720p with an entire CCD disabled like every dual CCD chip before it.

Difference is that those dual CCD chips did not have X3D on both CCDs.

 

[Fjodor2001]

For people who already have 9800x3d or 9950x3d, the performance uplift is, IMO, too small to justify an upgrade.

Maybe Zen 4 owners, doing a CPU only upgrade or upgrading older systems (the best candidates).

With these new CPUs not released yet, and Zen 6 likely less than a year away, waiting to upgrade may be the best idea.

Same as currently for those that have CPUs that perform just below 9800X3D and 9950X3D. No need to upgrade to those CPUs either then, unless you want to do it for braging rights.

 

[gdansk]

Difference is that those dual CCD chips did not have X3D on both CCDs.

And why would that help? Cross CCD communication is still slow. It must be avoided. And any *game* that would benefit from having more than 8 cores is touching shared memory across multiple CCDs making it a bus benchmark. And the bus hasn't changed.

It's a self defeating combination for games. It's useful for non-gamers, or the mythical no-shared-memory message passing only game.

 

[Fjodor2001]

And why would that help? Cross CCD communication is still slow. It must be avoided. And any *game* that would benefit from having more than 8 cores is touching shared memory across multiple CCDs making it a bus benchmark. And the bus hasn't changed.

It's a self defeating combination for games. It's useful for non-gamers, or the mythical no-shared-memory message passing only game.

I did not say it helps with cross CCD communication. But it helps with perf for threads executing on the second CCD which now also has X3D (compared to if that CCD did not have X3D).

Cross CCD communication is not all that determines perf.

We also have this which I mentioned previously:

You also have the case with 9T-16T (note: not 9C/17T to 16C/32T).

Then with 9950X3D2, all of those threads can get a dedicated core (i.e. no SMT needed slowing down perf/thread) and also access to X3D cache at the same time.

With 9950X3D, threads 9-16 instead have to choose between:
a) Being on the CCD with X3D: Access to X3D cache, but no dedicated core and thus have to use SMT shared with threads 1-8.
OR:
b) Being on the CCD without X3D: Deciated core for threads 9-16 (and 1-8 of course), but no access to X3D cache.

Will be interesting to see what it all adds up to when benchmarked.

 

[gdansk]

But it helps with perf for threads executing on the second CCD which now also has X3D.

Yes, indeed. Games that have no shared memory and use n cores without communicating across chip boundaries will benefit.

For every other game it'll do better with one CCD disabled (I guess that's just a 9850X3D). It'll be very interesting to see how many games the 9950X3D2 is faster in.

 

[Fjodor2001]

Yes, indeed. Games that have no shared memory and use n cores without communicating across chip boundaries will benefit.

For every other game it'll do better with one CCD disabled (I guess that's just a 9850X3D). It'll be very interesting to see how many games the 9950X3D2 is faster in.

It will also help for (some) cases where there is cross CCD communcation.

There will basically always be some cross CCD communication. How much will differ per game.

What matters is the net result of the plus from those threads having access to X3D and also a dedicated core (instead of sharing one with SMT), minus any penalty from cross CCD communcation.

 

[gdansk]

It will also help for (some) cases where there is cross CCD communcation.

More cache works well thus far because it avoids trips to the IOd to access memory. But to coherently share memory across CCD, the IOd must arbitrate any RAW no matter how much cache is involved.

What portion of games do you think this will be a net benefit to?

 

[MS_AT]

It'll be very interesting to see how many games the 9950X3D2 is faster in.

I would guess the relative relation between 9950X3D2 and 9800X3D will mirror the relation relation between 9950x and 9700x.

 

[Joe NYC]

Yes, indeed. Games that have no shared memory and use n cores without communicating across chip boundaries will benefit.

For every other game it'll do better with one CCD disabled (I guess that's just a 9850X3D). It'll be very interesting to see how many games the 9950X3D2 is faster in.

If I were to make a guess, AMD will give 9950x3d2 extra ~100 MHz, and then, a well behaved game, that does not leave a single CCD will end up being marginally faster on 9950x3d2.

 

[eek2121]

I think the performance improvements are going to surprise a bunch of folks. Also, games aren’t the only thing that benefit from a larger cache.

Further, if large cache chips are normalized, software optimization can be done to take advantage of it.

That being said l, I just want to know if I get more performance in Factorio. 😉

 

[jdubs03]

It still hasn’t been added to CPU-Z. Whereas a 9850X3D has been. The announcement of it has been in shambles. Alienware backtracked on it.
Idk seems amateur to me.

If it does come out, what benefit does it really have over a standard X3D as of now? The 9850X3D is the new gaming champ.

 

[Thibsie]

Faster clock ?