Question Zen 6 Speculation Thread

[Bigos]

Apparently there is also a "stretch goal", which I infer to be higher than 7 GHz.

https://twitter.com/x/status/1943928924704047123

Why are we entertaining a clown?

 

[Bigos]

I will add that the ~40% for Zen5 was true for pure heavy AVX-512 only, but it had some truth to it.

Where did you get 40% PPC for AVX-512 workloads from? Was there mentioned somewhere officially? Or are you deriving that from your own workloads?

 

[Kepler_L2]

This was the target gfxclk for Navi48 about 1 year before launch. They missed by 490 MHz.

I have no doubt their target for Zen6 is 7GHz, the only question is how close they can get to that target.

 

[gdansk]

View attachment 127048

This was the target gfxclk for Navi48 about 1 year before launch. They missed by 490 MHz.

I have no doubt their target for Zen6 is 7GHz, the only question is how close they can get to that target.

And if we extrapolate from N48 with 0.858x of target then Zen 6 would reach 6008 MHz. A most curious proportionality. It's almost like whoever is making up targets is doing the same math...

 

[Bigos]

View attachment 127048

This was the target gfxclk for Navi48 about 1 year before launch. They missed by 490 MHz.

I have no doubt their target for Zen6 is 7GHz, the only question is how close they can get to that target.

If you are deriving that just from how much they missed RDNA4 target then:
1. You do not understand clock rate challenges between 3GHz and 6GHz.
2. The clock target does not mean anything if they can miss it so much and still release product.

I would not compare anything <4GHz like GPUs with things >4GHz like x86 CPUs. What you have to do to attain such high clock rates is not comparable with such low GPU clocks.

AMD might still surprise us with >6GHz CPUs, maybe even close to 7GHz (I highly doubt it). This will have nothing to do with how high they could have clocked their GPUs...

 

[Schmide]

Where did you get 40% PPC for AVX-512 workloads from? Was there mentioned somewhere officially? Or are you deriving that from your own workloads?

The PS3 emulator got a 30% improvement from avx-512.

 

[Joe NYC]

Why are we entertaining a clown?

For very simple reason:

Imagine you are in the middle of a desert, and you have a compass. You can start in any one of 360 degrees to find an oasis.

Then someone (MLID) tells you "Walk straight East". Which is one of the 360 possibilities.

Suppose the likelihood that MLID is correct is 50%. But that's means the other 359 possibilities combined are also 50% and picking any one of them gives you 50%/359 probability of finding the oasis.

 

[Markfw]

Where did you get 40% PPC for AVX-512 workloads from? Was there mentioned somewhere officially? Or are you deriving that from your own workloads?

Primegrid compared to a 7950x vs a 9950x and a 9654 64 core vs a 7742 64 core. My only Zen 3 64 core was a 7763 ES and it only ran at 2.3 ghz/

 

[OneEng2]

You're not getting lower latency.
Lower power? yeah.

While it is POSSIBLE that the new IOD will result in no improvement in latency, with the faster memory and new IOD, why wouldn't you expect a lower latency?

Each nanosecond is 6.5 clocks at 6.5 GHz CPU. So shaving off a few nanoseconds can add up.

Very quickly.

For once, I pray that other forum members can contain themselves from hyping up Zen 6. Let's have a serene launch this time around.

I would encourage those same members to hype up Nova Lake if they want

It's fun to try to make 1+1=3, but I agree, we all tend to do stack ups in a very unreasonable way ..... resulting in launch day depression .

ES samples are out in the wild and interesting part about the two memory controllers in Zen6

I was wondering if this could help latency?

It's not just the overall performance increase, but where it is going to come from.
- from core itself, probably close to 0%
- from bigger L3 slightly lower memory latency - ~5%
- from clock speed 10+%

I think the core will provide a small increase, but in general, I agree with where your thoughts are on this.

 

[DrMrLordX]

I have no doubt their target for Zen6 is 7GHz, the only question is how close they can get to that target.

@adroc_thurston seemed to indicate that total 1T for Zen6 would be +10% * +10-15% = +21-26.5%, I forget which of those was IPC and which of those was clocks but it doesn't seem like they'd hit 7 GHz in any case. MT will be different because of the expansion to 24c on the desktop and unknown all-core boost targets.

 

[poke01]

A

@adroc_thurston seemed to indicate that total 1T for Zen6 would be +10% * +10-15% = +21-26.5%, I forget which of those was IPC and which of those was clocks but it doesn't seem like they'd hit 7 GHz in any case. MT will be different because of the expansion to 24c on the desktop and unknown all-core boost targets.

A 12 core X3D alone in one chiplet is enough to upgrade IMO, the rest is just cherry on top. Biggest problem with 8 core X3D is shader compilation ie its not fast enough. A 12 core N2P X3D is a tough one to beat, Q4 2026/early 2027 will be most interesting since Alder Lake.

 

[511]

We will have to wait until Page 500 for the real answer.

In the meantime, I don't think we will have any trouble filling pages 192-499 with speculation.

Well if we get 502 I doubt we will get to page 500 🤣

Also node doesn't guarantee same max clock speed improvements N2 is 15% iso Performance per watt not peak performance.

 

[Io Magnesso]

No, AMD just likes doing CBT with their physical design engineers. Remember Navi3x had nearly double the frequency target from Navi1x with just a single node shrink.

What is CBT?

 

[jpiniero]

What is CBT?

Cock & Ball Torture

 

[Io Magnesso]

Cock & Ball Torture

It's like being an engineer Penance Isn't it

 

[Io Magnesso]

Here is this the gathering of people with illusions that Zen6 can reach 7GHz or 6GHz or higher?

 

[Io Magnesso]

Unfortunately, MLID is a dangerous person who spreads hoaxes
In order to maintain our good PC life, I have to erase him...

 

[Thunder 57]

Cock & Ball Torture

I hope that is not what @Kepler_L2 was referring to...

 

[msj10]

? Apple got a 10% frequency increase from N3B

yeah while exploding power

 

[Kepler_L2]

I hope that is not what @Kepler_L2 was referring to...

It is

yeah while exploding power

View attachment 127063

That was N4 to N3B, Zen6 is N4P to N2P, plus there is plenty of power headroom for ST if they need it.

 

[Thunder 57]

It is

That was N4 to N3B, Zen6 is N4P to N2P, plus there is plenty of power headroom for ST if they need it.

At least now I know, thanks. I thought I might've been missing something.

 

[poke01]

yeah while exploding power

View attachment 127063

The price you pay for minimal IPC gains. This is why @gdansk is right while increasing clock is the easiest way to increase performance of a core it comes at the cost of power consumption.


N4P 9950X @ 5.7GHz uses ~35 watts. Will AMD maintain this level of power consumption at >6.0GHz?

 

[511]

It is

That was N4 to N3B, Zen6 is N4P to N2P, plus there is plenty of power headroom for ST if they need it.

from TSMC public info we have N5->N3E 18%PPW and N5->N4P 11% PPW that would mean N4P->N3E is 6% better and N3E->N2P is 18%

N5 to N3B was 10-15% PPW(12.5%) factoring in that N3B to N2P is






we have finflex in both N3/N2 so i haven't included that in calculation

 

[Io Magnesso]

There is no dream story in this world
There's no such thing as convenient

 

[poke01]

There is no dream story in this world

but i want a 7GHz core at 35 watts