Question Zen 6 Speculation Thread

[MS_AT]

You know what I mean. No need to try to look uneducated, or intentionally trying to misunderstand.

Sorry I am lost. When I was intentionally trying to misunderstand something?

For 12C/24T, at 12T basically you won't be using SMT since there'll only be one T per C. When going above 12T, additional threads will have to share core with some other thread. So perf/thread will be lower compared to when only having 1 T per C. So you can call those two threads sharing the same core SMT threads or whatever. Does not matter what you call it. Of course they are all just threads in OS context.

I am afraid you are mixing hardware threads with OS threads. There is a difference there.

But the point is that those two threads will be sharing the resources of one core, so they won't be executing as fast as threads which have a core dedicated to them and do not share it with other threads.

And this is the point you are missing. What you say is true only for contested resources. But what resources will be contested depends on the code you are running. Therefore the performance degradation will depend on the overall environment. But is not guaranteed to be:

much slower than e.g. E cores

But of course there are corner cases (one is quite popular here) where spamming E cores will win.

I mean if SMT was so bad, they wouldn't focus so much in Zen5 for making it easier to get second instruction stream going. Of course, it's server first design, it cares about throughput, but latency hit from SMT is usually negligible and if you empirically find it to be a problem, then you either modify the scheduling algorithm or disable SMT in the BIOS worst case and move on.

Apple of course is another league, it seems they are able to extract enough ILP from single instruction stream that they don't need to care about SMT.

 

[adroc_thurston]

Apple of course is another league, it seems they are able to extract enough ILP from single instruction stream that they don't need to care about SMT.

They really don't have the shipping schedule to have SMT.

 

[Thunder 57]

You know what I mean. No need to intentionally try to misunderstand.

But I'll school you since you asked for it:

For 12C/24T, at up to 12T basically you won't be using SMT since there'll only be one T per C. When going above 12T, additional threads will have to share core with some other thread. So perf/thread will be lower compared to when only having 1 T per C. So you can call those two threads sharing the same core SMT threads or whatever. Does not matter what you call it. Of course they are all just threads in OS context. But the point is that those two threads will be sharing the resources of one core, so they won't be executing as fast as threads which have a core dedicated to them and do not share it with other threads.

And main point: Those SMT threads (or whatever you want to call them) will be executing slower than threads on E cores (and P cores too of course!). Because threads on E cores do not share the E core with other threads, since E cores do not use SMT/HT.

SMT is so terrible Intel is bringing it back. Now stop threadcrapping.

 

[StefanR5R]

I think that everyone needs to start warming up to the idea that Zen 6 desktop is going to get walloped by NVL 52c in MT testing.

Sounds as if you are expecting Intel to ship their 48t CPU with a notably higher default power limit than AMD's 48t CPU.

 

[adroc_thurston]

Sounds as if you are expecting Intel to ship their 48t CPU with a notably higher default power limit than AMD's 48t CPU.

Threads are whatever if you can't clock.
Intel can't.

 

[Fjodor2001]

SMT is so terrible Intel is bringing it back. Now stop threadcrapping.

Depends on the use case / workload. Now stop the threadcrapping.

 

[adroc_thurston]

Depends on the use case / workload

It's a net win in any nT.

 

[Fjodor2001]

Sorry I am lost. When I was intentionally trying to misunderstand something?

If you still do not get it I give up.

I am afraid you are mixing hardware threads with OS threads. There is a difference there.

Point is that a thread that shares a core with another thread will run slower than a thread that has the core to itself.

And this is the point you are missing. What you say is true only for contested resources. But what resources will be contested depends on the code you are running. Therefore the performance degradation will depend on the overall environment. But is not guaranteed to be:

Two threads sharing one SMT core will always run slower, compared to two threads each having a separate core dedicated to themselves (assuming identical core types). The question is just how much slower.

And two Zen6 threads sharing one SMT core will almost always run slower, compared to two threads each executing on their own NVL E core.

But of course there are corner cases (one is quite popular here) where spamming E cores will win.

I mean if SMT was so bad, they wouldn't focus so much in Zen5 for making it easier to get second instruction stream going. Of course, it's server first design, it cares about throughput, but latency hit from SMT is usually negligible and if you empirically find it to be a problem, then you either modify the scheduling algorithm or disable SMT in the BIOS worst case and move on.

Apple of course is another league, it seems they are able to extract enough ILP from single instruction stream that they don't need to care about SMT.

Not a corner case, but the the normal.

Also, there's a difference depending on use case / workload. E.g. gaming vs server differs w.r.t. if SMT/HT is good or bad.

 

[Fjodor2001]

It's a net win in any nT.

But not in gaming. Assuming you want threads with high perf/thread. And then SMT/HT loses, when for xC/yT going above x threads.

 

[poke01]

But not in gaming. Assuming you want threads with high perf/thread. And then SMT/HT loses, when for xC/yT going above x threads.

AMD has SMT , a better L3 cache and higher ST next gen.


Intel will not come close to Zen6 x3D

 

[adroc_thurston]

But not in gaming

Makes no difference there.

Assuming you want threads with high perf/thread

Games want maybe 8 of those.

 

[poke01]

Look I’m sure Nova lake will have wins too like the first to use APX and a bigger NPU but etc but it’s not beating AMD in gaming.

 

[Covfefe]

I don't know why 48 core Nova Lake is getting so much interest in the Zen6 thread. They both have such high core counts that single thread will matter 5x more than multi thread anyways. Who cares if Nova Lake wins in Cinebench. If Zen6 wins in single core benchmarks by 5%, Zen6 will be the better CPU for the majority of buyers.

 

[adroc_thurston]

Look I’m sure Nova lake will have wins too like the first to use APX and a bigger NPU but etc but it’s not beating AMD in gaming.

DT/luggable NVLs are a meme.
Mobile will be generally better (especially given the better cost floor).

 

[Fjodor2001]

I don't know why 48 core Nova Lake is getting so much interest in the Zen6 thread. They both have such high core counts that single thread will matter 5x more than multi thread anyways. Who cares if Nova Lake wins in Cinebench. If Zen6 wins in single core benchmarks by 5%, Zen6 will be the better CPU for the majority of buyers.

Depends on the use case. Not everyone uses their PC only for gaming.

 

[Covfefe]

Depends on the use case. Not everyone uses their PC only for gaming.

I'm not talking about gaming. There are few workloads that scale to more cores endlessly. 48 core Nova Lake will have worse performance than 24 core Nova Lake in many lightly threaded benchmarks due to introducing separate L3 caches.

 

[adroc_thurston]

48 core Nova Lake will have worse performance than 24 core Nova Lake in many lightly threaded benchmarks due to introducing separate L3 caches.

No?

 

[Covfefe]

No?

Yes?

 

[Fjodor2001]

I'm not talking about gaming. There are few workloads that scale to more cores endlessly. 48 core Nova Lake will have worse performance than 24 core Nova Lake in many lightly threaded benchmarks due to introducing separate L3 caches.

If you think all workloads won’t scale to 48T, then 24C/48T Zen6 will have the same problem.

But I don’t agree for all cases, since it depends on the workload. Also, you do not have to run only one App/workload at the same time, you can run several in parallel, each using some of the 48T.

 

[adroc_thurston]

Yes?

No.
NUCA penalties for NINE caches are tiny and you're scheduling the bulk of your stuff to the fast CCD first.

then 24C/48T Zen6 will have the same problem.

Duh.
Again, the part people will buy is the 8c one, with or without the cacheslab.

Also, you do not have to run only one App/workload at the same time, you can run several in parallel, each using some of the 48T

This is the hypothetical usecase that's just not real for desktops.

 

[poke01]

let’s say Intel got to 6GHz for Novalake and AMD Zen6 is above 6GHz

How will Intel beat even AMD in ST, nova lake is no Unified core? And in gaming ST and cache and latency is all that matter. Not if you have 60 cores

 

[CakeMonster]

Duh.
Again, the part people will buy is the 8c one, with or without the cacheslab.

Isn't it unnecessarily wasteful to make 8c versions out of 12c CCD's? That's a bigger disabled area than 6c out of 8c.

 

[adroc_thurston]

How will Intel beat even AMD in ST

That's the neat part: they won't.

And in gaming ST and cache and latency is all that matter

Duh.

Isn't it unnecessarily wasteful to make 8c versions out of 12c CCD's?

That CCD goes into like a whole pile of products so sourcing fast and leaky 8c bins is trivial.

 

[OneEng2]

The other 93% pretty much only cares about Chrome and Office.

AMD's last quarter revenue was nearly 50% data center. Based on operating income, AMD's last quarter statement showed about 84% of its profit was from data center. That is why AMD as stated time and time again that Zen 6 design is "data center first".

I suspect that this is ALSO why Intel is so in the red. It has little to do with desktop.

Sounds as if you are expecting Intel to ship their 48t CPU with a notably higher default power limit than AMD's 48t CPU.

No. I am expecting that both Zen 6 and NVL to be power bound by the mother board socket.... and that Intel's "mont" cores are specifically designed to push through high performance at lower power than AMD's Zen architecture is.

... still, see the cavate below .

Threads are whatever if you can't clock.
Intel can't.

In high MT applications, it's about power efficiency, not clock speed since you are almost always power bound in these applications.

18A is unlikely (IMO) to clock like N2; however BSPDN may well provide good power characteristics if you drop the clock a bit.

Look I’m sure Nova lake will have wins too like the first to use APX and a bigger NPU but etc but it’s not beating AMD in gaming.

Unlikely in the extreme. I agree.

I don't know why 48 core Nova Lake is getting so much interest in the Zen6 thread. They both have such high core counts that single thread will matter 5x more than multi thread anyways. Who cares if Nova Lake wins in Cinebench. If Zen6 wins in single core benchmarks by 5%, Zen6 will be the better CPU for the majority of buyers.

I don't disagree at all with what you said.

I do recall quite vividly Intel's blue man group media blitz and P4 though. I can easily see Intel advertising "More than twice the cores of the competition". That may have an impact ..... or maybe not. I just bought my wife a new laptop (Lenovo Yoga OLED). It's Ryzen 5 AI, but honestly, I have no idea how many cores or what clock speed. It has a gorgeous screen though .

 

[Thunder 57]

I do recall quite vividly Intel's blue man group media blitz and P4 though. I can easily see Intel advertising "More than twice the cores of the competition". That may have an impact ..... or maybe not. I just bought my wife a new laptop (Lenovo Yoga OLED). It's Ryzen 5 AI, but honestly, I have no idea how many cores or what clock speed. It has a gorgeous screen though .

No, just no. When was the last time you saw an ad from Intel? This hasn't been a thing for 20 years or so with perhaps Centrino.