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

[DisEnchantment]

Speculate at will

 

[PJVol]

Maybe in a couple years, if AMD really was expecting bigger jumps from Zen 5, we will get the whole story from a couple of engineers who worked on it.

Some of this may become clear after we see the V/F curve and Cac weights for the cores

 

[HurleyBird]

/s of course (MLID just got lucky here) but still lol

Betting against hype trains is always going to give you a better hit rate than betting for them.

 

[poke01]

https://www.amd.com/en/products/processors/desktops/ryzen/9000-series/amd-ryzen-9-9950x.html

No CCD measurements?

 

[Saylick]

Got this from Olrak on Xitter. Looks like the gaming comparison was made using 6000 MT/s DDR5 for both Zen 5 and the 14900KS. The Intel mobo was set to "Intel Defaults", but I'm not entirely sure what that means these days.

 

[gdansk]

And 7900XTX - so maybe not the best comparison.

 

[poke01]

Got this from Olrak on Xitter. Looks like the gaming comparison was made using 6000 MT/s DDR5 for both Zen 5 and the 14900KS. The Intel mobo was set to "Intel Defaults", but I'm not entirely sure what that means these days.

View attachment 100288

oh its over for the DDR5 5600 people

 

[Joe NYC]

No it's the actual CDNA4.
Different stuff to support way different speeds and feeds.

That one moves the paradigm into a different direction.
Used to be called MI500.

Mi400 should coincide with HBM 4 in 2026.

Since HBM4 will be SoIC stack, I wonder if my prediction will turn true, and at that time, AMD will have HBM stacks on top of base AID dies...

 

[Geddagod]

https://www.amd.com/en/products/processors/desktops/ryzen/9000-series/amd-ryzen-9-9950x.html

No CCD measurements?

I don't think the point of contention here is about how much larger the CCDs are, just how much larger the cores are
Of course, measuring total CCD area differences is going to downplay the increases of area by just the cores, because of how much of the CCD is not cores

 

[adroc_thurston]

Mi400 should coincide with HBM 4 in 2026.

Yea.

Since HBM4 will be SoIC stack

Iirc that's just an option for HBM4.

 

[Joe NYC]

Weird that no pricing info was released for Ryzen 9000 and no release date.

I think the pricing will be roughly the same as Zen 4, and given the IPC uplift is underwhelming, AMD will have to release V-Cache models ASAP, hopefully before Christmas, at the time of Arrow Lake launch.

 

[Hulk]

Arrow Lake: Flop
Zen5: Flop
Apple M4: Flop
Cortex-X5: ? (Flop, obviously, we know how this story ends)

So basically all computing devices: Flop
Back to the the slide rule or abacus I suppose.

 

[gdansk]

Weird that no pricing info was released for Ryzen 9000 and no release date.

I read that as last day of July and same MSRPs.

 

[Joe NYC]

I read that as last day of July and same MSRPs.

I see, says Coming July:

 

[Jayzen]

So basically all computing devices: Flop
Back to the the slide rule or abacus I suppose.

 

[Rigg]

I have a super duper reliable source that says 9950x MSRP is $1000.

 

[Gideon]

I did 3800X - 5800X - 5800X3D

1700X -> 3700X -> 5800X3D for me. Would have skipped zen 2, but at some point it looked to be the end of the road for B350

 

[gdansk]

I really wanted 300MHz more for the nice 6GHz number.
AMD make me a 9950XS, thank you.

 

[Joe NYC]

No it's the actual CDNA4.
Different stuff to support way different speeds and feeds.

That one moves the paradigm into a different direction.
Used to be called MI500.

There were some rumors that the next version (presumably CDNA4) would reshuffle the die sizes, moving to fewer and bigger dies.

If that's part of the plan, would that be part of the Mi350 (using the current naming) or one after that, Mi400?

 

[DisEnchantment]

This slide was spot on, seems Zen 6 with 10% IPC would be what I would bet on.
MLID the winning leaker.

AMD took almost two years for a paltry 16% while Apple and ARM iterating on almost yearly cadence.

 

[Exist50]

There were some rumors that the next version (presumably CDNA4) would reshuffle the die sizes, moving to fewer and bigger dies.

If that's part of the plan, would that be part of the Mi350 (using the current naming) or one after that, Mi400?

Why ask someone just caught red handed lying to everyone for the last year?

 

[branch_suggestion]

So the sitrep is thus:
Granite Ridge: 20% IPC is the goal for a major tock, it failed. V-Cache will save the day in gaming so we are more or less status quo assuming LNC is a similar IPC bump with a small clock regression.
Turin: Still dominant, looks better than desktop does for sure.
Strix: Good, should be enough to justify the big jump in design wins, also looks a bit smaller than expected.
Z5 in general is the smallest bump since pre-Zen, but it does enough to retain its place in the market.

For the other stuff, Versal is great, MI roadmap is excellent and the overall ecosystem is looking very simple and workable now, which is great for everyone.
We will wait for all the final details and deep dives to get the full picture, overall it looks steady and solid, but unfortunately expectations were to kill the comp.
AMD themselves though never lied, this isn't an RDNA3 situation.
And to MLID, well you get to live another day, at least until LNC ends up being 15-20% IPC.

There were some rumors that the next version (presumably CDNA4) would reshuffle the die sizes, moving to fewer and bigger dies.

If that's part of the plan, would that be part of the Mi350 (using the current naming) or one after that, Mi400?

Yeah packaging yield is a concern so less dies balances the net yield. Wafers are plentiful relative to adv packaging.

 

[Saylick]

View attachment 100290

This slide was spot on, seems Zen 6 with 10% IPC would be what I would bet on.
MLID the winning leaker.

AMD took almost two years for a paltry 16% while Apple and ARM iterating on almost yearly cadence.

Based on that trend, whatever comes after Zen 6 is gonna have like 5% IPC increase and then after that 0% IPC uplift.

/s

 

[adroc_thurston]

Yeah packaging yield is a concern so less dies balances the net yield

It's not even that, just that tracking so many L2 instances is hard.
NV is at 2 with B100. MI300 is 8.

 

[phoenix21]

I have a super duper reliable source that says 9950x MSRP is $1000.

No need of sources for that
Simple ceo meth maths
The more you buy, the more you save.

 

[Kepler_L2]

And yet your proclaimed +32% has been proven to be wrong, and the MLID slide's "10-15%+" has been proven to be right.

What are you actually arguing here?

Not an MLID slide. I was the one that convinced Adored not to do a video on it, because the IPC number was lower than expected. My fault for accidentally giving MLID more credibility.