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

[DisEnchantment]

Speculate at will

 

[Hail The Brain Slug]

Presenting with permission from our wonderful ES user:

View attachment 102528

Did they reset the HWINFO data so we can't see what max power, thermals, clocks, or effective clocks were?

 

[Philste]

You can also see that a limit of 142W costs almost no performance. This makes it clear that they could have released the 7900X as a 120W TDP model with basically no difference in performance. AMD adjusting the TDP of the 9900X down to 120W is just acknowledging that.

And that's what I wanted to say. We shouldn't hype ZEN5s Efficiency, when it's basically just ZEN4s Stock profiles being bad.

 

[Mopetar]

And that's what I wanted to say. We shouldn't hype ZEN5s Efficiency, when it's basically just ZEN4s Stock profiles being bad.

Telling people not to hype in an AMD thread? Do you have any idea where you are right now or what's about to happen to you?

I call dibs on his stuff.

 

[KompuKare]

Oh, wait... they stole my "bad math" ?

Bad maths is almost a form of entrapment when it comes to wccftech articles at this point!

 

[Rheingold]

And that's what I wanted to say. We shouldn't hype ZEN5s Efficiency, when it's basically just ZEN4s Stock profiles being bad.

That, and at the least it tells us that power consumption did in fact not increase very much due to the wider core design. Perhaps some small improvement due to switching to some kind of N4 node and the wider core pretty much eliminate each other.

 

[yuri69]

Rapydmark score (probably unoptimized since it depends on a VC2017 runtime and possibly not using any SIMD) seems to suggest that software will need to be optimized for Zen 5 to unlock its potential. That's my thinking at the moment. The score? Not earth shattering so no point in revealing it or we'll get a WTFtech article saying Zen 5 disappoints!

Then it simply disappoints. Relying on SW optimization is rather bad for a product targeting general purpose codebases.

Moreover when the product is pushed by a party not owning a majority of the target market.

Moremoreover when the player has been utterly bad at both software and 3rd party relationships at the same time.

Anyways, the Zen core was meant to be balanced, to be mainstream - to be the opposite to Bulldozer - to be great at existing codebases - great at Intel-tailored binaries.

Zen 5 is not a Zen

 

[igor_kavinski]

Zen 5 is not a Zen

Bit premature to say that. Rapydmark is synthetic. We need something like TPU's test suite to form a proper opinion.

 

[Goop_reformed]

Presenting with permission from our wonderful ES user:

View attachment 102528

How does this compare w/ zen 4?

 

[jpiniero]

But I thought 32%, at least, was inevitable! Tier 1 OEMs don't lie and neither do Anandtech posters!

Could still be true. After all, they were talking about Epyc and not Ryzen.

 

[Saylick]

Looks like WTFtech caught wind of your latest info, @igor_kavinski

 

[StefanR5R]

We shouldn't hype ZEN5s Efficiency, when [...]

[...] EPYC 9004 vs. 9005 comparison benchmarks with task energy measurements aren't available. Or as a stopgap, look for Ryzen 7000 @ eco mode vs. 9000 @ eco mode comparisons after Ryzen 9000's launch.

[...] it's basically just ZEN4s Stock profiles being bad.

As far as I am aware, Ryzen's stock profiles of any generation are outside the sweet zone far enough to make them generally unrepresentative for Zen power efficiency assessments.

 

[SarahKerrigan]

Could still be true. After all, they were talking about Epyc and not Ryzen.

Are you implying that Epyc will use a different microarchitecture with drastically higher iso-clock single-thread performance? Because if so, wow.

 

[jpiniero]

Are you implying that Epyc will use a different microarchitecture with drastically higher iso-clock single-thread performance? Because if so, wow.

There's been rumors that AMD is doing a different (presumably bigger) core in Zen 6, not Zen 5... but I wonder...

 

[SarahKerrigan]

There's been rumors that AMD is doing a different (presumably bigger) core in Zen 6, not Zen 5... but I wonder...

Of course! +30-40% in Zen5 is happening... but on the other, secret, Zen5.

Y'all are hilarious.

 

[KompuKare]

Are you implying that Epyc will use a different microarchitecture with drastically higher iso-clock single-thread performance? Because if so, wow.

I would think that if some Epyc workloads do have higher gains than desktop, that those will either be due to AVX-512 or ,far more likely IMO, due to the power savings allowing better sustained loads for both ST and MT.

Server chips with tons of cores tend to have far lower clocks so any gain in efficiency will show there.

I know that's not IPC but server buyers and vendors probably care about throughout per server per watt far more than other users.

If that's the case, Zen will be great for servers but probably disappointing for desktop. Mind you, I'm also not buying the hype for team blue on the other thread!

 

[Makaveli]

Looks like WTFtech caught wind of your latest info, @igor_kavinski

lol Igor needs to start getting compensation from hassan for all the traffic its driving to their site.

 

[SarahKerrigan]

I would think that if some Epyc workloads do have higher gains than desktop, that those will either be due to AVX-512 or ,far more likely IMO, due to the power savings allowing better sustained loads for both ST and MT.

Server chips with tons of cores tend to have far lower clocks so any gain in efficiency will show there.

I know that's not IPC but server buyers and vendors probably care about throughout per server per watt far more than other users.

If that's the case, Zen will be great for servers but probably disappointing for desktop.

So... not 30-40% generalized iso-clock perf. How unexpected.

Mind you, I'm also not buying the hype for team blue on the other thread!

I'm sure I̶C̶L̶-̶S̶P̶ ̶S̶P̶R̶-̶S̶P̶ ̶E̶M̶R̶-̶S̶P̶ ̶S̶R̶F̶ GNR-SP will be the one to change everything.

 

[yuri69]

So... not 30-40% generalized iso-clock perf. How unexpected.

To use historical examples of cope, wait for:
* non-ES silicon
* final BIOS/firmware
* a very important OS update
* new chipset drivers
* all binaries recompiled using a new magic compiler
* upcoming BIOS/firmware update
* AMD-announced custom power profile
* new silicon revision
* Zen 5+ fixing all the Zen 5 problems
* Zen 6 is the original 40% Zen 5!!!!

 

[gaav87]

This ?

 

[KompuKare]

So... not 30-40% generalized iso-clock perf. How unexpected.

Those always seemed unrealistic.
Zen5 does seem to be server 1st, server 2d, and server 3rd. Far more than previous Zens.

The danger for AMD is that like last time with Conroe, it is not the monster P cores they need to watch - as they have almost got P4 crazy now - but rather the smaller team working on the efficiency stuff. That is the E team.

If those tiny cores can work well for servers then in hindsight AMD neglecting desktop will have been in vain.

 

[Saylick]

Those always seemed unrealistic.
Zen5 does seem to be server 1st, server 2d, and server 3rd. Far more than previous Zens.

The danger for AMD is that like last time with Conroe, it is not the monster P cores they need to watch - as they have almost got P4 crazy now - but rather the smaller team working on the efficiency stuff. That is the E team.

If those tiny cores can work well for servers then in hindsight AMD neglecting desktop will have been in vain.

Those E cores will only be a true threat to AMD’s full featured Zen cores if they can compete in HPC workloads, imo. Part of the reason why those E cores are so area and energy efficient is because they aren’t designed to clock stupidly high and they have limited AVX capability. Zen dense specifically aims to deliver a full featured Zen core in a denser, more efficient package. If AMD were to strip out even more AVX capability, it will be even denser and the gap narrows even more.

 

[jpiniero]

Course the 32% server number could have included clocks.

 

[SarahKerrigan]

Course the 32% server number could have included clocks.

If you look at what was originally claimed, it didn't. Adroc explicitly claimed 32% iso-clock ST int.

 

[jpiniero]

If you look at what was originally claimed, it didn't. Adroc explicitly claimed 32% iso-clock ST int.

That's right.

 

[H433x0n]

Course the 32% server number could have included clocks.

Tin foil hat time - I don’t think the 32% number was ever real. I think it was an honest mistake from extrapolating results from what appeared to be legit benchmark scores.

We’ve seen multiple instances where benchmark software incorrectly reports Zen 5 clocks much lower than it’s actually running. We’ve made this same mistake in this thread with Strix GB6 scores. I think this same thing happened with SIR2017 last year and it accidentally spawned a year long hype cycle.