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

[DisEnchantment]

Speculate at will

 

[Doug S]

Intel got $40 billion from the Biden silicon technology bill. Intel is going to sell silicon to 3rd parties on their fabs. That's something they have never done in the past.

That's not quite true. They tried to become a fab once before in the mid 2010s, but it was a half hearted effort. I think they mostly made FPGAs, nothing complex like SoCs for others, because they were still using their in house workflow instead of standard EDA tools. This time around they are doing it right, and forcing their x86 designers to switch to the industry standard workflow instead of homegrown stuff.

 

[adroc_thurston]

nothing complex like SoCs for others

Oh but they did.
They fabbed the second NUCLUN (the LG phone SoC yes).
And torpedo'd the 3rd with 10nm.
Also Nokia RFSoC (Reefshark) was on 10nm and some other designs.
It was a disaster.

 

[eek2121]

I firmly believe AMD should not have spun off their fabs. They had no choice, I get it, but I think it was a bad idea. You will notice that GloFo is still in business today, despite not having a cutting edge process.

The Conroe moment was when Intel destroyed all of AMD's hopes and dreams for more than a decade. The massive Conroe OC's out of the box that went on over the Intel product cycles for years. I still keep a drawer with my favorite Intel and AMD CPU's.

Yeah I don’t have a habit of collecting CPUs, but I almost never get rid of old hardware. It is a really bad habit.

Let's not go through these discussions again...




So, then, as eek2121 said then Zen 3 was also a Conroe moment. But its supremacy lasted only for a year.
There are so many designs in flight at Intel now that even if it is officially Zen V: the Return of Conroe I don't see anything dominating like Conroe did for so long.

Zen 3 chips are still competitive and are still being made today. If you normalize the clock speeds, a 14900k and 7950X aren’t all that much faster than a 5950X, especially if you also use the same speed of memory. Zen 3 was one of the most power efficient designs ever, just like Conroe.

Zen 3 was the first chip that gave Intel serious trouble.

Dual decoders are the new hotness (for x64 anyway).
But Mr. Thurston shared with us earlier his take (I don't think he claimed to have any special knowledge in how they achieved their gains?) that it likely looks more like A14's "Firestorm". If true it is a bit sad that all good designs will converge on a boring core like that.

“Boring core?” 🤣

Zen 4 is boring. Waiting impatiently on Zen 5 and RDNA4.

 

[Tuna-Fish]

Also Nokia RFSoC (Reefshark) was on 10nm and some other designs.

Did they ever actually ship working product in volume?

 

[adroc_thurston]

Did they ever actually ship working product in volume?

Yeah but not on Intel 10nm.
And years late.

 

[Fjodor2001]

What about perf/watt on Zen5 compared to Zen4, while executing MT loads on all cores? Any guesstimates for that based on e.g. the chip process tech used, and the architectural changes expected for Zen5?

Also, are we expecting the same situation as with Zen4, in that dropping the max TDP setting from 170W to ~120W will not result in much lower performance?

 

[adroc_thurston]

What about perf/watt on Zen5 compared to Zen4, while executing MT loads on all cores?

>20%.

Any guesstimates for that based on e.g. the chip process tech used, and the architectural changes expected for Zen5?

It's not a guesstimate.

 

[Fjodor2001]

>20%.

It's not a guesstimate.

Is the same perf/watt improvement expected when limiting TDP to 120W too (and comparing to same 120W for Zen4)?

 

[Ajay]

Is the same perf/watt improvement expected when limiting TDP to 120W too (and comparing to same 120W for Zen4)?

Only sensible perf/watt measurements are at fixed perf or fixed watts - the relationship in non-linear.

 

[adroc_thurston]

Is the same perf/watt improvement expected when limiting TDP to 120W too (and comparing to same 120W for Zen4)?

No it's the improvement with more TDP iso cc.

 

[Fjodor2001]

So what perf/watt improvement for Zen5 can be expected compared to Zen4 at fixed 120W TDP?

 

[adroc_thurston]

So what perf/watt improvement for Zen5 can be expected compared to Zen4 at fixed 120W TDP?

You can guess.

 

[dr1337]

I firmly believe AMD should not have spun off their fabs. They had no choice, I get it, but I think it was a bad idea. You will notice that GloFo is still in business today, despite not having a cutting edge process.

In this world where AMD tried to plough through the market with a subpar architecture only to pull themselves up by their bootstraps with Zen I firmly believe they would have had no chance splitting their R&D brainpower let alone funding.

If Globalfoundries can't break through 12nm on their own, but AMD can develop a good arch on their own, logic would dictate one group is more fruitful than the other. If it were easy enough to throw money at litho wouldn't all the other fabs also be on 7-5-3nm right now? Occam's razor says it all when Intel was the only mass market cutting edge 'foundry' for over a decade until 2019.

 

[DrMrLordX]

Yay, not the worst guess even!

I'll toss my hat into the ring.

Zen5 is actually a bunch of tiny "volunteer" workers from a politically-sensitive region of Asia that I won't be mentioning by name that have been shrunk by the shrink ray from Honey, I Shrunk the Kids, each equipped with an abacus. Keep in mind that they're REALLY small.

 

[Fjodor2001]

You can guess.

So for once, you don't know either?

 

[adroc_thurston]

So for once, you don't know either?

Samples aren't fused at 120W PPT, what did you expect.

 

[Fjodor2001]

Samples aren't fused at 120W PPT, what did you expect.

For anyone claiming to have access to samples, they can change max TDP to 120W for the CPU in the motherboard settings. So do you have access to such samples, and can try changing to 120W TDP?

Also, I assumed the info you have claimed in this thread previously wasn't based on CPU samples, but targeted final CPU performance. So it would be good if you can clarify whether your Zen5 info based on measurements from engineering CPU samples, or release CPU expectations?

 

[adroc_thurston]

they can change max TDP to 120W for the CPU in the motherboard settings

No you can't.
That's not how ref platforms work.

 

[StefanR5R]

Also, are we expecting the same situation as with Zen4, in that dropping the max TDP setting from 170W to ~120W will not result in much lower performance?

Not the same, but probably similar.

First of all, consider that 120/170 W TDP of desktop parts translates to 162(?)/230 W PPT limit. And that's what counts; let's forget about the rather inconsequential TDP figure.

Consider that Genoa 64…96 core parts are sold with 320/360/400 W cTDP_low/TDP/cTDP_high = min/default/max PPT limit. Divide these figures by 4…6 and you will arrive at power limits at which a 16-core Zen 4 part (with dual-channel-IMC etc.) would perform very power-efficiently and still with decent throughput.

Now consider that Turin will be released with …/…/500 W …/…/cTDP_high. That is, with +25% socket power budget compared to Genoa.

From that you can roughly estimate at which socket power range a 16-core Zen 5 part should perform with its best power efficiency, and still good throughput.

 

[Fjodor2001]

No you can't.
That's not how ref platforms work.

I thought you had insider knowledge, so you should know what AMD is targeting anyway. Not having to rely on your own measurements from engineering samples.

And you did not comment on this:

Also, I assumed the info you have claimed in this thread previously wasn't based on CPU samples, but targeted final CPU performance. So it would be good if you can clarify whether your Zen5 info based on measurements from engineering CPU samples, or release CPU expectations?

 

[adroc_thurston]

I thought you had insider knowledge, so you should know what AMD is targeting anyway

They're not targeting 120W SKUs.

 

[Thunder 57]

I thought you had insider knowledge, so you should know what AMD is targeting anyway. Not having to rely on your own measurements from engineering samples.

And you did not comment on this:

Because he is full of it. I will be glad to be wrong be he seems to be a prima donna.

 

[Fjodor2001]

They're not targeting 120W SKUs.

Did not say so either. But a lot of people are running Zen4 at 120W, since the performance difference is not big compared to running it on 170W.

Such people are interested in knowing the perf/watt difference between Zen5 vs Zen4, at 120W TDP.

So do you have any numbers for that or not?

 

[adroc_thurston]

Because he is full of it.

Wish I was.

But a lot of people are running Zen4 at 120W, since the performance difference is not big compared to running it on 170W.

It doesn't matter you do.
All that matters is what ships fused out of the fab.
And that is 170W GNR parts.

Such people are interested in knowing the perf/watt difference between Zen5 vs Zen4, at 120W TDP.

120W GNR doesn't exist.

 

[Fjodor2001]

Because he is full of it. I will be glad to be wrong be he seems to be a prima donna.

If he is, I assume his account will become silent after April 2024, and then a new recently created prophet account will show up in the Zen6 sister thread posting dead certain claims about Zen6. 🤣