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

[DisEnchantment]

Speculate at will

 

[adroc_thurston]

Yeah, I mean, they've been hanging around +15% for multiple microarchitecture generations on top of a pretty solid baseline. That's not bad, especially since they've done it on a reasonably quick cadence. (But it was kind of funny watching this thread whip itself into a frenzy about how +25% or higher was a done deal and +40% was plausible...)

How do I get 50% moar skt score off a 96c Turin at 16% moar PPC?

 

[Markfw]

Yeah, I mean, they've been hanging around +15% for multiple microarchitecture generations on top of a pretty solid baseline. That's not bad, especially since they've done it on a reasonably quick cadence. (But it was kind of funny watching this thread whip itself into a frenzy about how +25% or higher was a done deal and +40% was plausible...)

And nobody has addressed the efficiency yet. Even at 16% perf at the same wattage, means perf/watt is even better that Zen 4, and that blew Intel away

 

[SarahKerrigan]

How do I get 50% moar skt score off a 96c Turin at 16% moar PPC?

And yet here we are.

 

[adroc_thurston]

And yet here we are.

And now we gotta figure it out!

 

[tamz_msc]

You know it's bad when you have to include Geekbench 5 AES separately in the IPC comparison. I don't even trust that the actual IPC uplift is +16%, unless verified independently after launch.

 

[adroc_thurston]

You know it's bad when you have to include Geekbench 5 AES separately in the IPC comparison. I don't even trust that the actual IPC uplift is +16%, unless verified independently after launch.

They didn't put y-cruncher or w-prime or anything like that.

 

[poke01]

That's not bad, especially since they've done it on a reasonably quick cadence.

This makes Apple look even better.

 

[SarahKerrigan]

And now we gotta figure it out!

Eh, what the heck. Since it's SIR - better L3 eviction strategy can make a serious impact. SMT improvements (increase in replicated vs competitively shared structures?) Lower latency to remote chiplet L3?

 

[adroc_thurston]

And nobody has addressed the efficiency yet. Even at 16% perf at the same wattage, means perf/watt is even better that Zen 4, and that blew Intel away

That's not the issue.
To get 50% moar socket perf on a 96C Turin assuming modest clock bumps, you need >30% PPC.

better L3 eviction strategy can make a serious impact

Not that much.

SMT improvements (increase in replicated vs competitively shared structures?)

Doesn't look like it given how nonexistent perf uplift from SMT on is on Turin-D.

Lower latency to remote chiplet L3?

It's SIR n-copy, doesn't impact it much and given no uncore changes of note, also not happening.

Those are actually more useful in a relative sense than GB5 AES.

And it's faster there, just that they didn't put it for some reason.

 

[H433x0n]

If this is a real slide then saying +16% is being super generous. Look at the tests they're using to calculate that average.. They're using subtests of Geekbench (not the actual GB score) and using GB6 AES-XTS to artificially bring up the average.

 

[tamz_msc]

They didn't put y-cruncher or w-prime or anything like that.

Those are actually more useful in a relative sense than GB5 AES.

 

[luro]

Gaming performance will probably be about the same as x3d zen 4. What a disappointment

 

[CakeMonster]

Gaming performance will probably be about the same as x3d zen 4. What a disappointment

I'm happy to avoid the very hacky thread scheduling on X3D. Though I ofc hoped for more, given HandBrake, various AI tasks, and some non-X3D-boosted games I play.

I'm still questioning why they lowered the max wattage on most of the SKU's.

 

[gdansk]

Gaming performance will probably be about the same as x3d zen 4. What a disappointment

Well, it just means everyone will be waiting for Z5 X3D before making a purchasing decision.

 

[Panino Manino]

View attachment 100264

If we remove the Geekbench AES test which I think uses AVX-512 its not good. These have to be fake or old but thts looking unlikely

I did noticed that the TDP decreased.
Maybe it can go faster with more power?

 

[trivik12]

Zen 16% it is.

7/31 is just 8 weeks later. I hope we get 3rd party reviews on Day 1 of the release. Hopefully Anandtech does the deep dive if there is anyone to do it :-( Otherwise we have just Geekrawan.

 

[adroc_thurston]

Maybe it can go faster with more power?

That's only nT-relevant.
Not the talking point here.

 

[poke01]

I did noticed that the TDP decreased.
Maybe it can go faster with more power?

I believe the top sku the IPC is tested against is still 170W

 

[Joe NYC]

View attachment 100264

If we remove the Geekbench AES test which I think uses AVX-512 its not good. These have to be fake or old but thts looking unlikely

If this is the case, then MLID victory lap is going to be something to live for.

 

[adroc_thurston]

I believe the top sku the IPC is tested against is still 170W

Yeah and lower-end parts will look comparatively worse.

 

[Saylick]

The issue is if the slides do leak and I show up to this thread at the start of the Keynote, there is a good chance y’all would be talking about IPC and performance targets and then all the surprise would be gone

Dang, just hopped into this thread and I caught up.

Now my day is ruined and my disappointment is immeasurable.

 

[Tuna-Fish]

How does this work with Turin perf.

I think 96c Genoa is at least somewhat gimped by memory performance, it doesn't quite scale linearly in practice. If they improved that, it could result in Turin improving comparatively more than desktop Zen5.

Does Turin still use the same IOD?

 

[Goop_reformed]

Dang, just hopped into this thread and I caught up.

Now my day is ruined and my disappointment is immeasurable.

Yep lmao. 16% ipc uplift is pathetic

 

[Joe NYC]

I think 96c Genoa is at least somewhat gimped by memory performance, it doesn't quite scale linearly in practice. If they improved that, it could result in Turin improving comparatively more than desktop Zen5.

Does Turin still use the same IOD?

No, on server side, the IOD is new.

 

[Doug S]

If this is a real slide then saying +16% is being super generous. Look at the tests they're using to calculate that average.. They're using subtests of Geekbench (not the actual GB score) and using GB6 AES-XTS to artificially bring up the average.

Wasn't >15% the original claim from AMD? And not long after they said that, weren't people here and elsewhere starting to run wild with speculation about how they were lowballing? So then we'd hear >20%, then 25%, then 40% and some going even higher than that! I laughed every time I saw people pushing the 40% thing, it was wildly unrealistic but people with an ax to grind against Intel or maybe Apple wanted it to be true so badly they somehow fooled themselves into believing it.

Claiming 16% at launch is right in line with the original claims of ">15%", and if it turns out they had a fudge a little to get there that's understandable given their promise of >15% and the knowledge that they had allowed the Zen 5 hype to get so unrealistic that even a very conservative 16% would be disappointing for a lot of AMD fans.

If AMD was smart they would have tamped down the expectations to avoid all the crazy claims of 25, 30, 40% and more. All that does is create disappointment - and probably reduced their sales as no doubt some people who believed Zen 5 was going to be a massive jump (or at least thought there was a chance) held off buying. Some of those buyers might decide to hold off a bit longer to see what Intel can do, and if they can deliver (which remains to be seen, the inflated expectations thing seems to be starting on that side lately) those are sales AMD loses because they were content to let the speculation and hype run wild and uncorrected.