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

[DisEnchantment]

Speculate at will

 

[StefanR5R]

or they would have said

Correct me if I'm wrong. Somebody, who is not AMD, made some claim on Twitter. And that's it for now.

 

[Fjodor2001]

you're shadowboxing nonexistent stuff again

Sounds like you're talking about yourself. It was you who repeatedly made dead certain claims in this thread about Zen5 DT being released in April 2024. How's that boxing going now?

 

[Fjodor2001]

Correct me if I'm wrong. Somebody, who is not AMD, made some claim on Twitter. And that's it for now.

Regardless of who said it, if it's specified as H2 it usually means Q4, otherwise it's specified as Q3.

 

[Abwx]

So we know that one memory has got 1.385 times the clock compared to the other.
But who can tell us at which ratio the effective bandwidth between CUs and memory was increased?

The numbers were not accurate, the review has been updated, so far they removed the numbers with ocked RAM since they were quite lower than what they have now with stock 5200 RAM, they will probably add them in the coming hours or days.

 

[igor_kavinski]

if it's specified as H2 it usually means Q4

Says who? It could just as easily mean 1st July.

 

[adroc_thurston]

Sounds like you're talking about yourself. It was you who repeatedly made dead certain claims in this thread about Zen5 DT being released in April 2024. How's that boxing going now?

Pretty well.

It could just as easily mean 1st July.

It doesn't mean anything.

 

[Fjodor2001]

Says who? It could just as easily mean 1st July.

Based on past track record of how release dates are usually specified.

 

[StefanR5R]

Regardless of who said it,

Okay, so it's all the same then whether AMD make an announcement or a random passer-by is telling us that he was told something.

[Computerbase Phoenix test]

The numbers were not accurate, the review has been updated, so far they removed the numbers with ocked RAM since they were lower than what they have now with stock 5200 RAM, they will probably add them in the coming hours or days.

They test the GPU only with a set of games. Additional GPGPU tests could be helpful: There is a spectrum of GPGPU workloads from those which manage to utilize the streaming multiprocessors fully almost regardless of VRAM bandwidth, to those which almost entirely depend on VRAM bandwidth and hardly on execution units bandwidth.

 

[Fjodor2001]

Pretty well.

So you're still expecting an April 2024 release for Zen5 DT? Based on what?

 

[adroc_thurston]

So you're still expecting an April 2024 release for Zen5 DT?

hopefully.

Based on what?

Definitely not shadowboxing nonexistent stuff.

 

[Fjodor2001]

Okay, so it's all the same then whether AMD make an announcement or a random passer-by is telling us that he was told something.

No. But the way of specifying the date is usually the same. H2 meaning Q4 in practice, otherwise specified as Q3.

 

[Fjodor2001]

Definitely not shadowboxing nonexistent stuff.

So then what is the April 2024 release date based on? And is it still valid, given recent info surfacing on da Internet about delays etc?

 

[inf64]

I guess it's time for my likely very wrong performance projection for Ryzen 9950X . The streak was always going to end, why not with Zen 5.

Link to the doc ( I reused my previous doc for Zen 4 projections with very minor adjustments)
Computerbase data

Zen 5 core to have:

34% higher ST IPC (geomean for desktop, as AMD always calculates it)
25% SMT yield (same as Zen 3); Zen 4 has 34% SMT yield

Desktop TDP of 170W and 230W
Max ST clock of 5.5Ghz
MT clock of 4.8Ghz for 170W and 5Ghz for 230W variant
ST and MT projections done for the lower TDP model that boosts MT to 4.8Ghz

Versus 7950X (vanilla):
29-30% higher ST performance
15-16% higher MT performance

Notes:

1) I picked 34% ST IPC increase versus Zen 4 because Zen 5 is a monumental change in almost all important aspects that were bottlenecks in previous Zen cores. There will probably be some disappointing IPC jumps in specific workloads and some outrageously high increases in others, but I think that geomean will land in lower 30s percent.
MT IPC should be lower as I think Zen 5 will have lower SMT yield (in line with Zen 3, ~25%) : 1.34 x 1.25 / 1.34 ~= 1.25 or 25% versus Zen 4.

2) My numbers for 7950X on ST chart are from the actual Computerbase results (latest update from their webpage), and not from my projection sheet (document) for 7950X. Because Computerbase uses a specific non-favorable to AMD ST benchmark selection (mostly Cinebench aka Cinememe), actual ST index they got is ~7% lower for 7950X versus 5950X than what IPC x clock difference should show ( 1.1 x 5.7 / 4.9 ~= 1.28 or 28% )

 

[Abwx]

Okay, so it's all the same then whether AMD make an announcement or a random passer-by is telling us that he was told something.

[Computerbase Phoenix test]

They test the GPU only with a set of games. Additional GPGPU tests could be helpful: There is a spectrum of GPGPU workloads from those which manage to utilize the streaming multiprocessors fully almost regardless of VRAM bandwidth, to those which almost entirely depend on VRAM bandwidth and hardly on execution units bandwidth.

The issue was that the firmware was mobile dedicated and that it throttled the APU power uselessly, since the CPU part doesnt seems to be affected all GPU tasks will be improved by the 13% better average they got, that s at stock.

The 7-8% improvement with 38% faster RAM was indeed wrong since the GPU power was limited anyway, they could have increased RAM speed by 50% that the improvement would still be lacking, we ll see what is the RAM limitation magnitude once they publish the numbers with 7200 RAM.

 

[Glo.]

So mass production of Zen5 DT started in Oct-Nov 2023. Then why is it now not expected to be released until 2024H2 (likely meaning Nov-Dec 2024, or they would have said 2024Q3).

Are we expecting ~12 months from start of mass production until release for Zen5?

They said it this way so they can still sell their current products, whilethey are manufacturing them.

It usually happens when you DO have something good inside your engineering labs.

 

[igor_kavinski]

Versus 7950X (vanilla):
29-30% higher ST performance
15-16% higher MT performance

Seems conservative based on Adroc's prediction.

I believe:

45-55% higher ST
20-35% higher MT

I think the cores will initially be bandwidth starved in MT workloads due to lower maximum DDR5 speeds (7200 max). A year later, it may go up to 8400 MT/s or even higher.

 

[Fjodor2001]

They said it this way so they can still sell their current products, whilethey are manufacturing them.

It usually happens when you DO have something good inside your engineering labs.

Who is 'they'? Sounds like you mean AMD. But the info did not come officially from them.

And why would 'they' imply a 2023 Oct-Nov start of mass production , if they want to empty Zen4 stock? Why not just keep silent, so people would expect an even later start of mass production, and thus also later release of Zen5?

 

[Glo.]

Who is 'they'? Sounds like you mean AMD. But the info did not come officially from them.

Haven't they(AMD) said that Zen 5 products are coming this year?

 

[Fjodor2001]

Haven't they(AMD) said that Zen 5 products are coming this year?

Yes, meaning it could be any time during 2024. But that was not what was being discussed in the posts I commented on.

What was being discussed was the 2023 Oct-Nov start of mass production claim, and the April 2024 release claim. And that info does not come from AMD.

 

[Glo.]

Yes, meaning it could be any time during 2024. But that was not what was being discussed in the posts I commented on.

What was being discussed was the 2023 Oct-Nov start of mass production claim, and the April 2024 release claim. And that info does not come from AMD.

If it does not come from AMD, why do you care?

 

[igor_kavinski]

I think adroc's saying "hopefully" with regards to the April 2024 timeframe is coz Microsoft may try to force or convince OEMs to hold the release of their products until Win12 goes gold. So while AMD may be ready to release in April, OEMs may not want to. Microsoft's fault.

 

[StefanR5R]

[Computerbase Phoenix review]

we ll see what is the RAM limitation magnitude once they publish the numbers with 7200 RAM.

No. We will only get to see how faster clocked RAM will influence Phoenix' game performance. We will not get to see how it influences bandwidth between CUs and RAM.

 

[Fjodor2001]

If it does not come from AMD, why do you care?

Because some people in this thread seem to care about info not officially coming from AMD.

E.g. the April 2024 Zen5 DT release date was being generally accepted as true by a lot of people.

 

[Abwx]

No. We will only get to see how faster clocked RAM will influence Phoenix' game performance. We will not get to see how it influences bandwidth between CUs and RAM.

If the limitation is due to CUs to RAM bandwith then with 38% more bandwith the FPS should increase tremedously, if not this means that limitation is elsewhere, currently
the 12CUs RX6400 is only 22% faster than the 8700G/5200 RAM.

 

[igor_kavinski]

Intel should have released the 14900KS by now. I think they are waiting for the Zen 5 announcement.