Fingers crossed for the 9990AMDs 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
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?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.
You primarily want more ST perf, others more MT perf. Depends on use cases and preferences. Just the way it is.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.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?
As mentioned before TR is a pricing problem.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).
The are using a smaller process. AMD is using an older, cheaper process. I did try to make this clear.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.
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.@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.
Is M2 N5P denser than Strix Halo N4P?The are using a smaller process. AMD is using an older, cheaper process. I did try to make this clear.
"Bro just rent your CPU from Amazon Graviton"@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.
Go on then say it. Is it theft?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.
And with many cores. Gasp it's even better!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.
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.2 x 12600K > than anything 2 x AMD offers at that price.
As I understood we were talking about distributed rendering or container/VM spam? Which do just fine on heterogenous cores.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.
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.The are using a smaller process. AMD is using an older, cheaper process. I did try to make this clear.
N4P is better in every way than N5P. Including density.Is M2 N5P denser than Strix Halo N4P?
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).N4P is better in every way than N5P. Including density.
Exactly! And also there isn't an alternative way to get more ST.@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.
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.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?
How much did the core transistor count increase between Zen4 and 5? Not much, right?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?