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

[DisEnchantment]

Speculate at will

 

[dr1337]

Full speed CCD?

Not possible unless there is a 3rd GMI link on the IO die that AMD hid from everyone. From what I can tell reading the Epyc 9000 manual, full width mode uses 2 links for one CCD. So they could only give full speed to a single CCD by only having one of them currently.

Maybe there is a new cache die for dual chip parts that links the CCDs together, but I'd worry about even more heat issues and IDK if they could even pull that off anyways. The IO die would have to be somehow aware of the unified L3.

Knowing AMD and leakers, if anything about dual CCD parts being different is true, then I'd bet its a new driver for Ryzen Master that automatically allocates thread priority.

The main reason why I don't think the dual CCD chips will be special is that the reason V-cache scales so well for AMD is that they reuse the same dies for everything. Sure they could make a sole production line for a single product that might be 5% faster than it was otherwise, but then they lose their economies of scale. And with how EPYC is currently laid out they'd be stuck at 64c max for Turin-X.

 

[jdubs03]

Looks like the 9800X3D is coming in early November.

https://twitter.com/x/status/1839130656661782613

 

[branch_suggestion]

Looks like the 9800X3D is coming in early November.

https://twitter.com/x/status/1839130656661782613

Maybe AMD can start using their portfolio synergy and launch N48 at the same time.
Imagine a Microcenter bundle with 9800X3D+full N48 for $799, or for AM4/older Intel guys add a mobo and DDR5 for $999.
Now DIY mss is already solid, but the messaging and mindshare desktop brings can influence laptop and the like, so I think that can really convince people to change from the stereotypical Intel+NV mindset that is still very entrenched in many groups of consumers.

 

[igor_kavinski]

Maybe AMD can start using their portfolio synergy and launch N48 at the same time.
Imagine a Microcenter bundle with 9800X3D+full N48 for $799, or for AM4/older Intel guys add a mobo and DDR5 for $999.
Now DIY mss is already solid, but the messaging and mindshare desktop brings can influence laptop and the like, so I think that can really convince people to change from the stereotypical Intel+NV mindset that is still very entrenched in many groups of consumers.

Someone send this message directly to Lisa please.

@adroc_thurston ?

 

[tsamolotoff]

Does consumer sentiment ever influence ODM decisions laptopwise etc?

 

[cebri1]

https://videocardz.com/newz/amd-already-planning-ryzen-5-9600x3d-with-3d-v-cache

With the efficiency improvement this is going to be a great SFF chip.

 

[branch_suggestion]

Does consumer sentiment ever influence ODM decisions laptopwise etc?

Having a decent roadmap and ample supply with each product cycle does.
MDF rounds it out. AMD had the ball rolling at one point, then the PC market crashed and Intel flooded everything with RPL-282, AMD had to throttle laptop wafer allocations to keep the balance sheet okay.
On the GPU side RDNA2 was doing okay, and then RDNA3 happened and it is back to 0.
Now that DC money is flooding in we should see a proper push for market share again in laptop, but real volume comes with Kraken next year and HWK-R to further augment.

 

[MadRat]

Maybe there's a way to mathematically arrange four 4-core groups in the corners where they map caches out to two of the three 4-core co-processor groups, so you can quickly share perpindicular caches with the real penalty on the accesses to the core located at the diagonal. You would have much less of the worst sharing penalty 66% of the time. I would think CPUs are bundled in fours because they already do this at that scale.

That also leaves you space between all the cores for a unified cache. If accessing the unified cache matched the penalty for accessing the diagonal 4-core co-processor group, you eliminate the crosstalk traffic across perpindicular core groups for accessing that diagonal core group by redirecting requests to the central cache. Use math to balance all these penalties out for best results greater than 66% of the time.

 

[Josh128]

Again, I wouldnt get too excited about X3D. I very much doubt the "cool differentiators" are going to amount to much. I dont see clocks increasing much at all. AMD seems to have done the bare minimum outside of core redesign in this gen.

 

[fastandfurious6]

haters gonna hate 😎

 

[lightmanek]

Again, I wouldnt get too excited about X3D. I very much doubt the "cool differentiators" are going to amount to much. I dont see clocks increasing much at all. AMD seems to have done the bare minimum outside of core redesign in this gen.

Sometimes I wish people would listen to themselves before posting ...
Just done the bare minimum outside of core redesign... where core redesign is the biggest thing they could do to move architecture forward.

We all knew Zen 5 is core and Zen 6 IO revamp, why people are acting like this is somehow a letdown?

 

[igor_kavinski]

We all knew Zen 5 is core and Zen 6 IO revamp, why people are acting like this is somehow a letdown?

And since Zen 5 design supposed to be done by the Zen 2 design team, they DID get a huge perf uplift.

Over Zen 2

 

[lightmanek]

And since Zen 5 design supposed to be done by the Zen 2 design team, they DID get a huge perf uplift.

Over Zen 2

I know it's more a jest and completely justified when talking about gaming.

One could reply by saying that they did get a huge perf uploft over Zen 5, it's just that majority of users doesn't know how to use it 😉

My own opinion is that Zen5 will age like fine wine, it is a very good desig but it could have been so much better if it targeted N3 process (which it targeted elearly in development).

 

[igor_kavinski]

My own opinion is that Zen5 will age like fine wine, it is a very good desig but it could have been so much better if it targeted N3 process (which it targeted elearly in development).

Agree.

 

[Josh128]

Sometimes I wish people would listen to themselves before posting ...
Just done the bare minimum outside of core redesign... where core redesign is the biggest thing they could do to move architecture forward.

We all knew Zen 5 is core and Zen 6 IO revamp, why people are acting like this is somehow a letdown?

I'll tell you exactly why it is a letdown. Zen 5 = + ~5% gaming over Zen 4, same frequency. Therefore, it follows that Zen 5 3D would logically be ~5% faster than 7800X3D unless something we dont know about was changed in the design that allows for significantly higher frequency or throughput. But looking at how similar in physical packaging Zen 5 is to Zen 4, I dont see how that could possibly be the case.

 

[SteinFG]

I'll tell you exactly why it is a letdown. Zen 5 = + ~5% gaming over Zen 4, same frequency. Therefore, it follows that Zen 5 3D would logically be ~5% faster than 7800X3D unless something we dont know about was changed in the design that allows for significantly higher frequency or throughput. But looking at how similar in physical packaging Zen 5 is to Zen 4, I dont see how that could possibly be the case.

5.5+ GHz could make it worth. 7800X3D is stuck at 5.0
But yeah, if 9800X3D is <5.5 GHz, it's gonna be sad

 

[fastandfurious6]

something tells me vcache will reap the benefits of zen5

9800x3d > 7800x3d by ~~~~+20%+~~~~

 

[Hail The Brain Slug]

I'll tell you exactly why it is a letdown. Zen 5 = + ~5% gaming over Zen 4, same frequency. Therefore, it follows that Zen 5 3D would logically be ~5% faster than 7800X3D unless something we dont know about was changed in the design that allows for significantly higher frequency or throughput. But looking at how similar in physical packaging Zen 5 is to Zen 4, I dont see how that could possibly be the case.

I recall reading something about decreased L3 latency by 3.5 cycles? I'd think making the L3 lower latency would potentially amplify the benefits of the additional cache.

 

[StefanR5R]

I'll tell you exactly why it is a letdown.

Newsflash: These are not the 1980s anymore. CPU computing speed increases are only a minor factor in video game system performance now.

 

[DrMrLordX]

It's still e-waste generation. AMD shouldn't enable OEMs like that but I understand they don't want to lose a deal to Intel because of some fuses.

Probably the worst thing in recent memory that AMD has done, even if it doesn't affect the vast majority of us.

 

[Josh128]

Newsflash: These are not the 1980s anymore. CPU computing speed increases are only a minor factor in video game system performance now.

Newsflash. The selling point of X3D cpus is specifically video game performance. You only have to go back to 2022.

 

[jpesk2]

I'll tell you exactly why it is a letdown. Zen 5 = + ~5% gaming over Zen 4, same frequency. Therefore, it follows that Zen 5 3D would logically be ~5% faster than 7800X3D unless something we dont know about was changed in the design that allows for significantly higher frequency or throughput. But looking at how similar in physical packaging Zen 5 is to Zen 4, I dont see how that could possibly be the case.

One of the theories as to why Zen5 struggles to outpace Zen4 is that there isn't enough memory bandwidth for it to stretch its legs. Maybe the 3d cache will let zen5 stretch its legs a bit and maybe the clock regression for the X3D chips won't be there. If they just fixed the clock regression it could come in at 15-20% faster than zen4X3d, if the x3d cache mitigates some of the memory bandwidth problems it could go even higher.
Who knows? I do think Zen5 is a good chip, it's just not worth upgrading if you have a zen4.

 

[adroc_thurston]

Imagine a Microcenter bundle with 9800X3D+full N48 for $799, or for AM4/older Intel guys add a mobo and DDR5 for $999.

Forcing bundles into the channel is no bueno.

 

[StefanR5R]

I'll tell you exactly why it is a letdown. [games]

Newsflash: These are not the 1980s anymore. CPU computing speed increases are only a minor factor in video game system performance now.

Newsflash. The selling point of X3D cpus is specifically video game performance. You only have to go back to 2022.

Uhm, system performance, including for a range of game titles, did indeed improve (at least measurably) when Ryzen 5000 increased last-level cache per CCX from 16 to 32 MB, and Ryzen 5000X3D increased it once more to 96 MB. (Still down from Intel Broadwell's 128 MB of last level cache.) Ryzen 7000 and 9000 as well as 7000X3D and 9000X3D merely replicate this. Furthermore, Ryzen 7000(X3D) introduced DDR5 support, and Ryzen 9000(X3D) merely replicates this.

So did you expect some other drastic system level changes from Zen 5/ Zen 5X3D, and feel let down because there are none?
Or did you expect generational increases of CPU core computing speed to bring relevant improvements to games, and feel let down because that's not the case?

 

[adroc_thurston]

Still down from Intel Broadwell's 128 MB of last level cache

those are very different things

The selling point of X3D cpus is specifically video game performance.

they're good webservers too