Question Zen 6 Speculation Thread

[adroc_thurston]

it shouldn't need a high power limit

It's still 48 cores.

 

[Fjodor2001]

Regarding those high max power consumption numbers, I'm hoping we'll see behavior similar to this with Zen6:

For example, for 7950X you could limit TDP from 200W to 120W without losing much perf. You would be going from about 2050 to 1850 points. So approximately 40% lower power consumption, but only 10% lower perf.

For Zen6 24C, is there any chance we'll see similar behavior? E.g. instead of 200W or 230W we could run it at 120W or 138W, and only get approximately 10% lower perf compared to max perf?

 

[gdansk]

Probably. Pretty much every v/f curve looks like that.

 

[Hulk]

If you have missing predictions in my chart, I would really appreciate it if you could give me the missing data. It's really tweaking my OCD!

 

[Fjodor2001]

Probably. Pretty much every v/f curve looks like that.

Well, it differs quite a bit per CPU. Check out e.g. 9950X vs 7950X in the graph. At 200W the 9950X has about 10% higher perf than 7950X, but at 120W they are at the same perf.

 

[poke01]

@Hulk , 2700 for ST R23

 

[gdansk]

Well, it differs quite a bit per CPU. Check out e.g. 9950X vs 7950X in the graph. At 200W the 9950X has about 10% higher perf than 7950X, but at 120W they are at the same perf.

Not in the way that matters. Every part shipped since the power limit stupidity (which started with Rocket Lake) can be power limited and give up a fraction of its performance. Zen 5 is actually "bad" in that respect for Cinememe. But like every dual CCD Zen before it, Zen 6 will also have eco mode.

Conversely, why do you think Zen 6 would be different here?

 

[Abwx]

View attachment 133125

There s a mistake in your excel sheet, for ST i stated 2650 pts at 6GHz, guess that you confused with Detox.

 

[Fjodor2001]

Not in the way that matters. Every part shipped since the power limit stupidity (which started with Rocket Lake) can be power limited and give up a fraction of its performance. Zen 5 is actually "bad" in that respect for Cinememe. But like every dual CCD Zen before it, Zen 6 will also have eco mode.

Conversely, why do you think Zen 6 would be different here?

The point is that it differs per CPU how much lower perf you'll get when limiting the power consumption in BIOS settings. And certainly in a way that matters. E.g. for 9950X vs 7950X, the former drops 20% performance while the latter only drops 10%, when going from 200W to 120W. So percentage-wise 9950X drops perf twice as much as 7950X.

So the question is where Zen6 will fit into this picture. How many percent MT performance drop can we expect going from 200/230W to 120/138W? 5%, 10%, 20%, 30%, or more?

 

[gdansk]

How many percent MT performance drop can we expect going from 200/230W to 120/138W? 5%, 10%, 20%, 30%, or more?

That's specific. Something we won't know until it is out. And as usual for Zen it will vary by the instruction mix.
But for the general question, nothing is different. It'll ship with power limit that throws away efficiency for approximately 10% more performance.

 

[Hulk]

@Hulk , 2700 for ST R23

Thanks. Frequency for ST?

 

[Fjodor2001]

That's specific. Something we won't know until it is out. And as usual for Zen it will vary by the instruction mix.
But for the general question, nothing is different. It'll ship with power limit that throws away efficiency for approximately 10% more performance.

Sure. I was thinking there would be someone brave enough to take a guess already now, based on process tech, arch, core count, or whatever. I mean people are brave enough to out their perf estimates in the Hulk table.

In my case I’ll probably evaluate Zen6 vs NVL-S at approx 120W and see which one performs best. For me it’s not worth almost twice the power consumption as needed when maxing it out, just to get 10-20% extra perf or whatever.

 

[Det0x]

Outside of embarassingly parallel benchmark it would be impossible to saturate all the cores also it's PL2

So you agree that Zen6 will most likely win in almost all MT workloads other than cinememe/embarassingly parallel benchmarks since very few programs can actually use all the 48 cores on Nova Lake ?
Should be much easier to saturate the 24 cores on Z6 (granted you wont get full benefit from all the SMT threads)

 

[poke01]

Thanks. Frequency for ST?

6.4GHz

 

[Hulk]

I have a new appreciation for increased ST performance. If it comes from frequency or architecture or both.
I do multitrack audio work from time-to-time with Studio One. A new plugin recently came out from Relab, the 176 compressor. It is a massive compute hog, but sounds great. During playback running many instances of the plugin is not a problem, I've tried as many as 35 (ridiculous) with no playback issues. But here's the thing. During a mixdown (render) the mixdown routine is different from a realtime playback. Basically during realtime playback there is "time" to use various buffers to store data and more effectively use multiple cores. Great, no playback dropout issues. But when you mixdown the routine is more sequential in nature. My mixdown for a typical mix has gone from about 10x realtime to 2.4x realtime due to this new plugin.

I'm rendering 27 tracks right now and it's taking forever. You'll notice only one core is really in use. Total core usage is 5.2%! From what I have learned this is not a Studio One issue but a DAW issue in general. I wrote a message to Presonus tech support asking them if these mixdowns can be done in parallel. Meaning if only a couple cores are used per track then allocate say 2 cores per track and mixdown 8 tracks at once. I doubt they'll get back to me but we'll see.

Anyway, 6.5GHz. Yes! I'll take it!

 

[Fjodor2001]

I have a new appreciation for increased ST performance. If it comes from frequency or architecture or both.

Based on what you wrote in the rest of that post I have more of a despise for poorly written SW.

 

[Hulk]

Based on what you wrote in the rest of that post I have more of a despise for poorly written SW.

I hear you but apparently the nature of the mixdown makes MT problematic. But yeah. It's annoying. Luckily no slowdowns during the actual mixing. But mixdown is definately a time to do other things.

 

[gdansk]

Based on what you wrote in the rest of that post I have more of a despise for poorly written SW

Poorly written or is it just that some problems are not parallel

NC (complexity) - Wikipedia

en.wikipedia.org

If the steps must be applied in order and there are 35 independent steps, each step takes 65ms, how do you parallelize it? Spawn n worker threads and at best it can only speed up by whatever factor each step is parallelizable.

 

[Fjodor2001]

Poorly written or is it just that some problems are not parallel

NC (complexity) - Wikipedia

en.wikipedia.org

If the steps must be applied in order and there are 35 independent steps, how do you parallelize it?

And why can’t each of those 35 steps be parallelized in this case?

 

[gdansk]

And why can’t each of those 35 steps be parallelized?

Most probably each step itself can be. But it ends up waiting for the slowest "substep" of each step, then add synchronization time on top of that. There are truly problems which scale poorly. I'm not sure if this is one but it easily could be the outcome of having a modular plug-in architecture where each plug-in expects memory not to be modified beneath it between its own actions.

 

[Josh128]

Probably. Pretty much every v/f curve looks like that.

The curve might look the same, but the number axis can shift gigantically.

7960X, 24C, 5nm, 337W




4.650 GHz all core freq.



Even with 2nm magic, the above is why Im conservative on Zen 6 24C all core clocks, especially at 200W. Zen 6 24C would have to be 68% more power efficient to match this 7960X nT loaded freq with 200W. Im sure AMD is going to make some good strides on fixing idle /uncore power, but asking for more than 68% better power efficiency is a fairly tall order, even for a 2 node advance.

 

[gdansk]

Zen 6 24C would have to be 68% more power efficient to match this 7960X nT loaded freq with 200W

Would it? What's the IOd power for TR? and AM5 Zen 6 is TBD but most probably much less.

But regardless, that Zen 6 would have to much more efficient than Zen 4 isn't surprising nor the point of what I said.
Zen 6 dual CCD part is almost guaranteed to not to ship at the best point on that curve and that's not new. That's just what they do for consumers parts that will have a high MSRP since neither competitor will step back.

 

[poke01]

ST should always be appreciated, it’s the hardest to achieve and win at.

Just ask Intel.

 

[Kepler_L2]

The curve might look the same, but the number axis can shift gigantically.

7960X, 24C, 5nm, 337W


View attachment 133137

4.650 GHz all core freq.

View attachment 133138

Even with 2nm magic, the above is why Im conservative on Zen 6 24C all core clocks, especially at 200W. Zen 6 24C would have to be 68% more power efficient to match this 7960X nT loaded freq with 200W. Im sure AMD is going to make some good strides on fixing idle /uncore power, but asking for more than 68% better power efficiency is a fairly tall order, even for a 2 node advance.

IOD alone is 100W on EPYC/TR

 

[Josh128]

Would it? What's the IOd power for TR? and AM5 Zen 6 is TBD but most probably much less.

But regardless, that Zen 6 would have to much more efficient than Zen 4 isn't surprising nor the point of what I said.
Zen 6 dual CCD part is almost guaranteed to not to ship at the best point on that curve and that's not new. That's just what they do for consumers parts that will have a high MSRP since neither competitor will step back.

Fixing idle power is likely one of AMD's major goals with Zen 6. Even my 12 core 9900X consumes 40-50W PPT just idling on Windows desktop with HWinfo monitor running, which is not great. If they can knock that down to the sub 15W range that would be a tremendous accomplishment, but that remains to be seen. Even if they shave off 30-40W of that idle power draw, 200W PPT is still not going to be enough fuel to allow a high frequency capable 24 core to show its true power. 200W might actually be well within the sweet spot of the efficiency curve, but thats just all the more reason for them to push even more power, as you noted.