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

[DisEnchantment]

Speculate at will

 

[adroc_thurston]

They put Instinct on the 1 year cadence but they really need to put Zen on a yearly cadence. Or at most 16 months.

6q is what you'll get.

 

[Joe NYC]

RDNA3.5 is curiously not clocking that much higher compared to supposedly broken RDNA3, although I guess there's not much point to clock the 8WGP SKUs that high anyway.

Performance per Watt matters, and RDNA 3.5 seems to be showing some gains there (from the little data we have seen).

 

[adroc_thurston]

Performance per Watt matters, and RDNA 3.5 seems to be showing some gains there (from the little data we have seen).

Yeah it's good.
CPU, turns out, less so!

 

[PJVol]

65W vs 105W. I suspect that if you just open up the power, the chips will match the Zen4 nT clocks.

They lowered TDPs for a reason.

 

[Exist50]

Whoa, that's low.
How does this work with Turin perf.

Because your Turin numbers are just as wrong as your Zen 5 IPC claims.

 

[adroc_thurston]

Because your Turin numbers are just as wrong as your Zen 5 IPC claims.

Hopefully.

 

[burninatortech4]

Should be thankful Microsoft didn't push for NPU's on the desktop.

NPU's on the Zen 5 compute dies would break cross platform compatibility (they wouldn't be used on EPYC). The whole point of the compute die strategy is to use it across multiple product lines.

It makes perfect sense in the laptop dedicated die however - that's why it starts there.

 

[gdansk]

6q is what you'll get.

If that's the plan they keep missing

 

[Geddagod]

Weird....

 

[SpudLobby]

I still think Zen 5 in servers should be fine.

 

[poke01]

I still think Zen 5 in servers should be fine.

I mean yeah. Whose AMD's comp there? Intel? llol

 

[Saylick]

View attachment 100285Weird....

I wouldn't assume core area is accurate based on that artistic rendering... Better to wait for either HotChips or someone gets a real a die shot imo.

 

[burninatortech4]

was Genoa better than that Xeon? answer: yes

Milan gives it a run for it's money in (some) workloads.

 

[gdansk]

I still think Zen 5 in servers should be fine.

More cores solves a lot of problems for servers.
And if Z5c is on N3E, will it even clock much lower?

 

[SpudLobby]

I mean yeah. Whose AMD's comp there? Intel? llol

Exactly.

 

[adroc_thurston]

More cores solves a lot of problems for servers.
And if Z5c is on N3E, will it even clock much lower?

It's 50% moar cores over Bergamo.
The answer is "it depends".

 

[HurleyBird]

How about those new Turin chiplets? Should be a single 16-core CCX now, otherwise not much point to it. Don't think any leaker came close to guessing that. Wonder how it would perform on desktop, if that's even theoretically possible without big sacrifices... if this means there will be more two different V-cache dies, or if AMD just abandons V-cache on desktop now.

 

[Ajay]

PPA-wise it's fine, the CCD is barely bigger.
Rest yeah.

Yeah, kind of wonder what Mike Clark was on about when talking Zen5, idk, two years+ ago. It’s almost like someone made him take a big cut to the Zen4- Zen5 xtor budget increase.

 

[adroc_thurston]

How about those new Turin chiplets

There are dense.
Skinny 16c sausage.

Should be a single 16-core CCX now, otherwise not much point to it. Don't think any leaker came close to guessing that

For Dense, yes.

 

[SpudLobby]

More cores solves a lot of problems for servers.
And if Z5c is on N3E, will it even clock much lower?

Yeah I mean AMD is king of the DC and Xilinx + Pensando open some other avenues. Client I am meh on long term

 

[soresu]

View attachment 100285Weird....

Strix?

 

[HurleyBird]

View attachment 100285Weird....

Don't take AMD die renders too seriously.

 

[igor_kavinski]

The data center part of the presentation was a whole lot more interesting than client. Dang. I wish I had guessed 61% improvement. Then I could claim to have guessed the flipped side of the figure!

So this kinda sets the expectations for AMD client products going forward. Just decent but nothing earth shattering. Guess AMD doesn't want to spend too much resources on making poor people happy 🙁

 

[gdansk]

Yeah, kind of wonder what Mike Clark was on about when talking Zen5, idk, two years+ ago. It’s almost like someone made him take a big cut to the Zen4- Zen5 xtor budget increase.

I'm pretty sure he was excited about the dual decode pipelines. Seems like something an architect would find exciting, anyway.

 

[DisEnchantment]

Because your Turin numbers are just as wrong as your Zen 5 IPC claims.

Just popped in here to celebrate... ehh mate? I totally understand you. 🙂