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

[DisEnchantment]

Speculate at will

 

[Glo.]

Since AMD creates the SOC for both Playstation and Xbox, they could probably do something similar for PC if there were a marketing for it (which I doubt)

From engineering standing point of view - companies can do whatever they want.

The problem is - the market. Its customers and what they want to pay money for, or rather - are used to pay money for.

 

[soresu]

From engineering standing point of view - companies can do whatever they want.

The problem is - the market. Its customers and what they want to pay money for, or rather - are used to pay money for.

I'd hoped that Valve might test those waters for AMD by ordering a semi custom console scale chip to follow Steam Deck's Van Gogh.

Now that Steam OS is proving the viability of Linux gaming through Proton/Wine/DXVK and the various Linux AMD driver efforts it seems like the perfect time to resurrect the Steam Machine hype with a real Valve device to get the ball rolling.

 

[inf64]

I was checking the Zen 4 presentation from HC and noticed this interesting slide that goes against what I posted before (SMT uplift possibly being smaller with a wider and deeper core):


What if Zen 5, being much wider and deeper than Zen 4, gets similarly greater SMT uplift to the tune of ~40%? If the core has an average ST IPC uplift of say ~28% and SMT uplift of ~40%, then it would be roughly 33% more performant in MT workloads / clock versus Zen 4 (what adroc claims):

1.28 x 1.4 / 1.34 = 1.33

So in summary:
~28% ST IPC
~40% SMT uplift
clocks downgrade for desktop version, most likely 5.4-5.5Ghz range for ST and ~4.8Ghz for all core Turbo.
MT throughput around 23% higher, ST around 21% higher

Just speculating here .

 

[HurleyBird]

What if Zen 5, being much wider and deeper than Zen 4, gets similarly greater SMT uplift to the tune of ~40%? If the core has an average ST IPC uplift of say ~28% and SMT uplift of ~40%, then it would be roughly 33% more performant in MT workloads / clock versus Zen 4 (what adroc claims)

Adroc claims Zen 5 SMT uplift is something like one third of Zen 4's.

 

[Saylick]

I'm sure someone pointed this conjecture out before, but I'll state it in case it wasn't: what's the likelihood that MLID's leaked slide of 10-15% IPC uplift was for MT while ST could be much higher? Such a theory would satisfy both the leaked slide and what Adroc is claiming.

 

[inf64]

Adroc claims Zen 5 SMT uplift is something like one third of Zen 4's.

Oh I must have missed that post, do you have a link to it? Interesting indeed. AMD's Zen 4 slides seems to suggest that the wider and deeper the core, it should get the bigger SMT uplift.

Edit:
I just did the napkin math and it doesn't really add up. To get 32% higher MT IPC in nT workloads, with ~11% SMT uplift (versus 34% for Zen 4), you would need some insane ST IPC for the Zen 5 core:

1.6 x 1.11 / 1.34 ~= 1.33

60% is never going to happen.

 

[itsmydamnation]

Oh I must have missed that post, do you have a link to it? Interesting indeed. AMD's Zen 4 slides seems to suggest that the wider and deeper the core, it should get the bigger SMT uplift.

Edit:
I just did the napkin math and it doesn't really add up. To get 32% higher MT IPC in nT workloads, with ~11% SMT uplift (versus 34% for Zen 4), you would need some insane ST IPC for the Zen 5 core:

1.6 x 1.11 / 1.34 ~= 1.33

60% is never going to happen.

But that's not what he said , he said 32%1t not nt

 

[Mopetar]

IPC shouldn't decrease just by using multiple threads unless the cache or the rest of the memory system is creating a bottleneck.

 

[inf64]

But that's not what he said , he said 32%1t not nt

I was under impression he said the IPC uplift refers to the slide he saw, which was a reference to the nT Turin MT performance. Maybe @adroc_thurston can clarify?

 

[HurleyBird]

Oh I must have missed that post, do you have a link to it?

Turin-D SMT on is just a 10% socket-level perf bump over SMT off.

 

[Thunder 57]

Adroc claims Zen 5 SMT uplift is something like one third of Zen 4's.

I'll be glad when he is proven wrong and just spewing verbal diarrhea. Unless of course he is right and I will have egg on my face. Fun times ahead!

IMHO no one but a VIP has the numbers that he is spitting out. And someone that important isn't going to leak those numbers. Plus claiming a price already? Yeah, not likely.

 

[Markfw]

I'll be glad when he is proven wrong and just spewing verbal diarrhea. Unless of course he is right and I will have egg on my face. Fun times ahead!

IMHO no one but a VIP has the numbers that he is spitting out. And someone that important isn't going to leak those numbers. Plus claiming a price already? Yeah, not likely.

I would not bet anything against his claims. Or for them for that matter. But my gut tells to believe they are closer to the truth than many others.

 

[Tarkin77]

Bergamo successor is called "SONARO".

This Clown doesn't know redacted about Italian cities. It's called "SERRANO".

 

[inf64]

@HurleyBird
Yeah but that is not an IPC claim, but throughput per socket claim (which means that clocks can be different once SMT is enabled). Plus, it's about Turin Dense, not vanilla Zen 5 version.

 

[itsmydamnation]

@HurleyBird
Yeah but that is not an IPC claim, but throughput per socket claim (which means that clocks can be different once SMT is enabled). Plus, it's about Turin Dense, not vanilla Zen 5 version.

I think the other thing that can't be over looked is fabric bandwidth for dense , it's alot of very wide probably very aggressive prefetching cores for 1 ifop link. So fabric could be a limiting factor which then is addressed in zen6

 

[HurleyBird]

@HurleyBird
Yeah but that is not an IPC claim, but throughput per socket claim (which means that clocks can be different once SMT is enabled). Plus, it's about Turin Dense, not vanilla Zen 5 version.

Well this is SIR, about as SMT-friendly as it gets really.
Genoa SMT benefit is ~30%-ish there.

Anyway, I suppose it's possible Adroc only knows of Turin-D and there's some other bottleneck there that makes the SMT yield look bad, but if you look at the full context of the whole SMT discussion back then he's clearly claiming that the yield is way down (and Kepler makes similar claims).

 

[randomhero]

I think the other thing that can't be over looked is fabric bandwidth for dense , it's alot of very wide probably very aggressive prefetching cores for 1 ifop link. So fabric could be a limiting factor which then is addressed in zen6

Power too

 

[exquisitechar]

I'll be glad when he is proven wrong and just spewing verbal diarrhea. Unless of course he is right and I will have egg on my face. Fun times ahead!

IMHO no one but a VIP has the numbers that he is spitting out. And someone that important isn't going to leak those numbers. Plus claiming a price already? Yeah, not likely.

The price is just his speculation based on the performance he thinks Granite Ridge will have.

 

[Timorous]

But that's not what he said , he said 32%1t not nt

Specifically he has said Spec Int 1t is 32%.

 

[Tarkin77]

AMD fights Meteors with paper at Ryzen 8000 launch

It looks like AMD is fighting paper with paper for the launch of the new Ryzen 8000 series.

semiaccurate.com

Ryzen 8000 Paper Launch in December?!

 

[leoneazzurro]

AMD fights Meteors with paper at Ryzen 8000 launch

It looks like AMD is fighting paper with paper for the launch of the new Ryzen 8000 series.

semiaccurate.com

Ryzen 8000 Paper Launch in December?!

If true it will be probably only a teaser, not a launch, with some performance preview.

 

[Bigos]

Ryzen 8020, 8030, 8035, 8040, 8045 incoming?

(same as 7000 but with the first digit changed)

 

[Timorous]

Ryzen 8020, 8030, 8035, 8040, 8045 incoming?

(same as 7000 but with the first digit changed)

I am sure they might exist but anything Zen 5 based would be 805x

 

[Timorous]

Plus claiming a price already? Yeah, not likely.

With 14th gen performance now known and a long wait until 15th gen if Zen 5 offers a 20% + performance increase (so IPC and clocks netted out) over Zen 4 I can see a decent price hike because it will simply be the best thing in town and at best 15th gen might be competitive.

 

[mpumalanga]

AMD fights Meteors with paper at Ryzen 8000 launch

It looks like AMD is fighting paper with paper for the launch of the new Ryzen 8000 series.

semiaccurate.com

Ryzen 8000 Paper Launch in December?!

Charlys articles regarding AMD have turned into a joke.
He is constantly moaning and ranting about the smallest things.

He seems to be more on a personal vendetta against AMD instead of just reporting or commenting on their products.

His sources in Intel also seem to have vanished. He is only rehashing Intel press releases and emphasizing the positive bits.