Question Zen 6 Speculation Thread

[marees]

AMDs GPU game seems to be to make all of the units as small as possible and at some point leverage that and their ciplet tech to makes GPU that is bigger than what Nvidia can provide. Whether that happens or not...they seem to get cancelled

Fingers crossed for the 9990

 

[carancho]

I'd love more than 16 cores but I seriously don't think its very important on desktop in the next ~3-4 years (Z5 and Z6). Though, if there is very little IPC gain (which has made up for more cores in lots of MT apps since Z2) I might change my mind.

I don't understand requests like these for more MT performance. I mean, these posts are like letters to Santa, so why not wish for higher ST instead? AMD is clearly behind the best (Apple). I'd rather they budget whatever extra transistor they can deploy towards extra ST! What's the trade-off that I'm missing?

 

[Fjodor2001]

I don't understand requests like these for more MT performance. I mean, these posts are like letters to Santa, so why not wish for higher ST instead? AMD is clearly behind the best (Apple). I'd rather they budget whatever extra transistor they can deploy towards extra ST! What's the trade-off that I'm missing?

You primarily want more ST perf, others more MT perf. Depends on use cases and preferences. Just the way it is.

 

[gdansk]

I don't understand requests like these for more MT performance. I mean, these posts are like letters to Santa, so why not wish for higher ST instead? AMD is clearly behind the best (Apple). I'd rather they budget whatever extra transistor they can deploy towards extra ST! What's the trade-off that I'm missing?

You're ignoring that it isn't a letter to Santa.
AMD can glue a different CCD that they are making anyway to that IOd. AMD cannot offer better single thread performance so easily.

 

[StefanR5R]

I love more than 16 cores too. And you know what? I already have more than 16 cores. Quite a few more. At home, tugged away in a corner of the living room, currently humming along together with a stand fan (which isn't large enough anymore).

 

[gdansk]

I love more than 16 cores too. And you know what? I already have more than 16 cores. Quite a few more. At home, tugged away in a corner of the living room, currently humming along together with a stand fan (which isn't large enough anymore).

As mentioned before TR is a pricing problem.
I know it has been discussed to death because Fjodor and few others push it heavily but I don't discount the possibility, if Zen 5 sales are poor enough, that AMD may make some weird SKUs again to try to get mindshare back. Basically, if they can't increase IPC much then perhaps core spam is back on the menu.

 

[eek2121]

Apart from the Max die(which I agree AMD cannot do for client or can it? Strix Halo...) the M Pro and base M chips are not big at all.
If we take the M2 and M2 Pro which are on N5P, their total die size is respectively is 153.22mm² and ≈259mm². Note the M2 Pro has a bigger GPU.
If we take the M4 which is N3E, the die size is 166.37mm² and this is the node future Zen6 APUs will be on and AMD will likely have a laptop APU around that area too.

Meanwhile Strix Point has a die size of 232.5mm² on the N4P node. Its got nothing to do with cost, its just a different design philosophy and one that will be useful in laptops/handhelds.

https://hubweb.cn/apple-silicon/chip-m/

The are using a smaller process. AMD is using an older, cheaper process. I did try to make this clear.

 

[StefanR5R]

@gdansk, I don't have a Threadripper myself. There are several more ways to get more cores.

Everybody who actually has real use for more cores, already has more cores. Owns them, or rents them, or both. The others just love hearing themselves talking about more cores, without having an actual grasp of what more cores can do and what they cannot do, and without having ever made a serious plan how to implement the best bang-for-the-buck for an actual use case.

 

[gdansk]

@gdansk, I don't have a Threadripper myself. There are several more ways to get more cores.

Everybody who actually has real use for more cores, already has more cores. Owns them, or rents them, or both. The others just love hearing themselves talking about more cores, without having an actual grasp of what more cores can do and what they cannot do, and without having ever made a serious plan how to implement the best bang-for-the-buck for an actual use case.

And? AMD might still make it anyway. I don't get the point of category excluding it or calling it Santa's wishlist because other people buy used Epyc/Xeon/TR.

They have the capability and Intel may provide them the motivation.

 

[Nothingness]

The are using a smaller process. AMD is using an older, cheaper process. I did try to make this clear.

Is M2 N5P denser than Strix Halo N4P?

 

[StefanR5R]

@gdansk, there are even more ways to have more cores than by buying Epyc/Xeon/TR, new or used. Getting access to MT performance is a solved problem. Has been for very many years now.

Not so long ago, with the two first Threadripper generations, AMD provided high core count for single socket computers for cheap, of course with certain compromises. These were different times, but yes, perhaps AMD makes a move in the future with a similar outcome but for changed reasons, under changed circumstances. But the question whether or not cheap but compromised workstation CPUs/ HEDT CPUs are being on offer or not is merely a peripheral question when looking to satisfy an actual MT performance need.

 

[gdansk]

@gdansk, there are even more ways to have more cores than by buying Epyc/Xeon/TR, new or used. Getting access to MT performance is a solved problem. Has been for very many years now.

Not so long ago, with the two first Threadripper generations, AMD provided high core count for single socket computers for cheap, of course with certain compromises. These were different times, but yes, perhaps AMD makes a move in the future with a similar outcome but for changed reasons, under changed circumstances. But the question whether or not cheap but compromised workstation CPUs/ HEDT CPUs are being on offer or not is merely a peripheral question when looking to satisfy an actual MT performance need.

"Bro just rent your CPU from Amazon Graviton"
- AMD (a CPU salesman)

They're gonna add more cores at some point. Even if it isn't very helpful. Why continue this charade that it is a solved problem with no market? People are buying 24 core 14900K/13900K despite all its faults.

 

[StefanR5R]

I never said there is no market for more cores. I actually said the contrary. And no, by considering the possibility to rent instead of to own, you still haven't exhausted all of the actual ways to get more cores. For cheap even.

 

[gdansk]

I never said there is no market for more cores. I actually said the contrary. And no, by considering the possibility to rent instead of to own, you still haven't exhausted all of the actual ways to get more cores. For cheap even.

Go on then say it. Is it theft?

The options don't matter. AMD has the means to make a stupid product and there is no reason they shouldn't. AMD has no qualms throwing more cores on the same speed memory as Strix Point shows.

 

[StefanR5R]

Ethernet clusters can be dirt cheap. If lower latency is needed, small Infiniband clusters are dirt cheap as well. (You don't need a switch if you can get by with up to three nodes.) Strange that I have to spell it out: In many jurisdictions, individuals are permitted to own two computers, if not more.

Yes, many MT problems call for a single system image solution (edit: or single node solution even) instead of a cluster. But the usual suspects who moan for more cores have not the slightest idea about when to choose one or the other anyway.

 

[gdansk]

Ethernet clusters can be dirt cheap. If lower latency is needed, small Infiniband clusters are dirt cheap as well. (You don't need a switch if you can get by with up to three nodes.) Strange that I have to spell it out: In many jurisdictions, individuals are permitted to own two computers, if not more.

And with many cores. Gasp it's even better!
In almost no way is it an alternative, but complementary to core spam parts.

And for that the Lakes are king at almost every price point. 2 x 12600K > than anything 2 x AMD offers at that price. And why? Because they spam cores even at low price points. And this trickles down to any configuration using new parts no matter how many nodes you add.

 

[StefanR5R]

2 x 12600K > than anything 2 x AMD offers at that price.

No, they are <, not >, if we talk about parallel computing beyond Cinebench. The uneven performance of the two core types makes these parts less attractive for parallel computing, if it doesn't exclude them from consideration outright.

 

[gdansk]

No, they are <, not >, if we talk about parallel computing beyond Cinebench. The uneven performance of the two core types makes these parts less attractive for parallel computing, if it doesn't exclude them from consideration outright.

As I understood we were talking about distributed rendering or container/VM spam? Which do just fine on heterogenous cores.

You know your "just add more nodes" is specious. Adding more nodes doesn't change anything. It just doubles or triples any throughput/$ disadvantage.

 

[poke01]

The are using a smaller process. AMD is using an older, cheaper process. I did try to make this clear.

What? I showed the die sizes for the M2 and M2 Pro which are on N5P which is an older node than N4P which Strix is using.

There is clear reason why I didn’t use the M3 series because it’s on N3B and it’s irrelevant to AMD.

 

[poke01]

Is M2 N5P denser than Strix Halo N4P?

N4P is better in every way than N5P. Including density.

 

[Nothingness]

N4P is better in every way than N5P. Including density.

Yes, my question was rhetorical It's also likely that N5P is less expensive (though back when M2 was released it was probably more expensive).

 

[LightningZ71]

The ACHIEVABLE triangle (performance, efficiency, density) of N4P is better than N5P in every way. This does NOT preclude a customer from using a dense N5P configuration that has greater density than a performance, relaxed density product on N4P.

While I don't have hard figures in front of me, it is entirely possible that the density and efficiency focused M2 product could be denser than the performance and efficiency focused Strix Point product, even if N4P is a better node than N5P.

 

[carancho]

@gdansk, there are even more ways to have more cores than by buying Epyc/Xeon/TR, new or used. Getting access to MT performance is a solved problem. Has been for very many years now.

Exactly! And also there isn't an alternative way to get more ST.

What kind of client workload benefits more from say 2x MT than from 1.5x ST+MT? What kind of client workload needs so much MT that isn't accelerated by GPUs or NPUs already?

And if any of you identify such workload, what's the share of time that a use will spend under that workload such that they'll prefer 2x MT to 1.5x ST+MT? If you're **exclusively** doing MT workload flat out on a client, what the heck are you doing using a client CPU?

 

[gdansk]

Exactly! And also there isn't an alternative way to get more ST.

What kind of client workload benefits more from say 2x MT than from 1.5x ST+MT? What kind of client workload needs so much MT that isn't accelerated by GPUs or NPUs already?

And if any of you identify such workload, what's the share of time that a use will spend under that workload such that they'll prefer 2x MT to 1.5x ST+MT? If you're **exclusively** doing MT workload flat out on a client, what the heck are you doing using a client CPU?

You're not getting that much more ST. We already saw the slide. Zen 6 targets "10%+" IPC which means 10-11%. It's all tapped out. There is basically no more ST performance in the x64 ghetto going forward.

You really, really don't think that is the time to add more MT? How are they going to get people to buy if it's merely 10% faster?

No surprise that same slide says larger complexes too.

 

[carancho]

You're not getting that much more ST. We already saw the slide. Zen 6 targets "10%+" IPC which means 10-11%. It's all tapped out. There is basically no more ST performance in the x64 ghetto going forward.

You really, really don't think it is time to add more MT? How are they going to get people to buy if it's merely 10% faster?

How much did the core transistor count increase between Zen4 and 5? Not much, right?

And why wouldn't it be possible to achieve the ST performance levels of Apple? Without resorting to discredited ISA superiority arguments....