Question Zen 6 Speculation Thread

[adroc_thurston]

Zen 4 to Zen 5 was a massive sub node improvement.

Nope.

Zen 5 being a server focused design is just cope for why it's so mid tbh.

It is.

 

[Thunder 57]

Zen 4 to Zen 5 was a massive sub node improvement.
Zen 5 being a server focused design is just cope for why it's so mid tbh.
ARL's E-core's pipeline is extremely long. Way longer than Zen's.

Have you seen server benchmarks?

 

[Meteor Late]

Zen has always been server focused, this won't change with Zen 6. So with the total midness of Zen 5 in mind, no one should expect miracles from Zen 6, a 10% IPC increase is probably in the cards and a very nice frequency bump because of the node jump and that's it, around 25% 1t jump overall.

 

[Joe NYC]

Zen has always been server focused, this won't change with Zen 6. So with the total midness of Zen 5 in mind, no one should expect miracles from Zen 6, a 10% IPC increase is probably in the cards and a very nice frequency bump because of the node jump and that's it, around 25% 1t jump overall.

In other words, Zen 5% ^ 2?

Are we now seriously growing expodentially, as was foretold in a prophecy?

 

[OneEng2]

OK, I can't find the original post to quote, but where you agreed about the cache fitting in memory, I wanted to add a real world scenario. A primegrid task where you use pinning (can be accomplished by smaller instances that are pinned) time goes from days to 9 hours. smaller examples exist, just giving real world times on a database task.

It was in my discussions about the lack of a "Full Zen 6" 128c DC version vs Zen 5. The argument was that there is no reason for a full Zen 6 with higher core counts.

I know that databases love more performance per core and as much cache as they can get. Licensing like Oracle is "per core" so having max per core performance is very important.

Additionally, in our discussions on latency, (mostly the dismal latency of ARL) the idea that lowering the latency with any means needed would likely have much bigger performance uplifts on NVL than many are suspecting.

I never realized how wildly different the performance of some tasks could be simply by keeping it all in cache though. That is a massive difference!

Thanks!

 

[Markfw]

I never realized how wildly different the performance of some tasks could be simply by keeping it all in cache though. That is a massive difference!

That does reply on pinning the CPUs to the ccd and disabling SMT for full performance, but yes, it can make a massive difference in come apps.

Actually , all that is done to force the CPUs to use the cache !

 

[StefanR5R]

A primegrid task where you use pinning (can be accomplished by smaller instances that are pinned) time goes from days to 9 hours.

The unpinned duration is from BoincTask's estimated time remaining right? I.e. extrapolated with an unknown estimation error, smaller or bigger.

I never realized how wildly different the performance of some tasks could be simply by keeping it all in cache though. That is a massive difference!

Yes, the example given is AVX-512 accelerated multithreaded Fast Fourier Transform, working on a rather large array of FFT coefficients. (Think of Megabytes of FFT coefficients.) It indeed makes a big difference whether the coefficients are pretty much always pulled from a single shared L3$, or are sprinkled across separate caches or spill from cache to main memory. This phenomenon is about as old as CPUs being equipped with vector arithmetic units. But as the latter grow fatter and fatter over CPU generations, their dependence on CPU cache for performance also grows, given that main memory bandwidth lags and main memory latency has almost flatlined for many years now.

 

[poke01]

I guess zen6 mass production starts in in June 2026 and release by November?

 

[Josh128]

In other words, Zen 5% ^ 2?

Are we now seriously growing expodentially, as was foretold in a prophecy?

~25%-29% 1T improvement has always been a very target mark to hit. Thats pretty much what Zen 4 got over Zen 3, what Zen 3 got over Zen 2, and what Zen 2 got over Zen 1. (Zen 2 1t was only about +15% over Zen +, Zen 5 was only about +15% over Zen 4)

 

[Meteor Late]

One could think, well, if Zen 4 already got that 25+% 1t improvement with the help of a node jump, why wouldn't Zen 6 get more than that with a double node jump?
Well, imo, because:
-I don't expect IPC increase to be as high as it was in Zen 4, meaning 3-5% less IPC increase this time around, I have the impression that a bit more priority has been put on incresing core count and CCD core count this time around among other things.
-fmax not as easy to increase as frequency in the lower part of the curve. So while we would see Apple, Qualcomm, etc, getting more frecuency increases from 5/4nm, they are probably at 1.2V and lower, while AMD needs pulses around 1.35V or so to achieve 5.7GHz 1t in desktop parts, I expect less increases in that range. This means no 7GHz for example, but I could see 6.5GHz.

 

[Josh128]

One could think, well, if Zen 4 already got that 25+% 1t improvement with the help of a node jump, why wouldn't Zen 6 get more than that with a double node jump?
Well, imo, because:
-I don't expect IPC increase to be as high as it was in Zen 4, meaning 3-5% less IPC increase this time around, I have the impression that a bit more priority has been put on incresing core count and CCD core count this time around among other things.
-fmax not as easy to increase as frequency in the lower part of the curve. So while we would see Apple, Qualcomm, etc, getting more frecuency increases from 5/4nm, they are probably at 1.2V and lower, while AMD needs pulses around 1.35V or so to achieve 5.7GHz 1t in desktop parts, I expect less increases in that range. This means no 7GHz for example, but I could see 6.5GHz.

Because node jumps dont guarantee fmax increase.

 

[adroc_thurston]

Because node jumps dont guarantee fmax increase.

yeah they do

 

[exquisitechar]

-I don't expect IPC increase to be as high as it was in Zen 4, meaning 3-5% less IPC increase this time around, I have the impression that a bit more priority has been put on incresing core count and CCD core count this time around among other things.

Those other things being worked on doesn't mean the core has been neglected at all. AMD can do both.

 

[Josh128]

yeah they do

No they dont.

 

[adroc_thurston]

I have the impression that a bit more priority has been put on incresing core count and CCD core count this time around among other things.

?
This doesn't impact IPC targets at all.

No they dont.

Yes they do.
Incompetents like Intel are not part of that equation.

 

[OneEng2]

View attachment 132124
I guess zen6 mass production starts in in June 2026 and release by November?

Which Zen 6?

One could think, well, if Zen 4 already got that 25+% 1t improvement with the help of a node jump, why wouldn't Zen 6 get more than that with a double node jump?
Well, imo, because:
-I don't expect IPC increase to be as high as it was in Zen 4, meaning 3-5% less IPC increase this time around, I have the impression that a bit more priority has been put on incresing core count and CCD core count this time around among other things.
-fmax not as easy to increase as frequency in the lower part of the curve. So while we would see Apple, Qualcomm, etc, getting more frecuency increases from 5/4nm, they are probably at 1.2V and lower, while AMD needs pulses around 1.35V or so to achieve 5.7GHz 1t in desktop parts, I expect less increases in that range. This means no 7GHz for example, but I could see 6.5GHz.

Agree!

Furthermore, as process shrinks have gone FAR beyond the point of diminishing returns, I think 5% IPC is the new norm. I think it might be a little higher for Zen 6, but without the huge transistor budget increases of the past, we can't expect big IPC increases either.

Your point of where phone processors are in frequency relative to PC/Server is also quite valid. As you move up in frequency, you QUICKLY get beyond linear heat production, and at the top of the process spectrum, you are at a square of the frequency for heat production.

This is nothing new, so it confuses me why so many here are ignoring it. Dreams die hard I guess .

 

[511]

Which Zen 6?

There is only one Zen6 lol

 

[Joe NYC]

yeah they do

18% for N2P over N3E:

https://twitter.com/x/status/1978737339171017215

 

[inquiss]

Zen has always been server focused, this won't change with Zen 6. So with the total midness of Zen 5 in mind, no one should expect miracles from Zen 6, a 10% IPC increase is probably in the cards and a very nice frequency bump because of the node jump and that's it, around 25% 1t jump overall.

Yep, that's it. Just a teeny tiny 25%

 

[inquiss]

There is only one Zen6 lol

Eh, sorta. Volume on n3 class and fun stuff on n2 class. Gonna have different frequencies...