Ok so within tolerance for intel
Same shit was said about haswell temps but yes way lower tdp to attend to
But there are rumors of soldered ihs

Why do you hate on intel?

why do you think mesh won't work?

This post was edited by yupitsmeh on Jun 28 2017 04:51pm

Mesh is good in theory but it hasn't proven itself to reduce latency significantly yet as far as i've seen.
It also doesn't allow them to do what amd does with die's, its just allows (in theory) better intercommunication between cores.

It's not about thinking it doesn't work, it's about being shown that it does work before acknowledging it.

i.e not being a fanboy

core for core performance doesn't matter, scaling does.

Is it even comparable to infinity fabric? I haven't seen Intel use it to stitch together multiple die's with 1:1 scaling, so I don't see how they can compete with AMD in the server space when their chips run hotter, draw far more power, and cost much more to produce because they need perfect silicon for 20+ adjacent cores to function as needed.

This post was edited by DCSS on Jun 28 2017 04:59pm

Wouldn't it be similar to knights /landing/ corner /mill?

Phi's also operate on single die cores with mesh interconnect, meaning yields are still difficult. Reflected in the price, fuck paying $5000 for 64 Atom cores.
With Ifabric AMD could, in theory, make a chip with 64 zen cores at a similar maybe even lower cost, and there's really no need to mention how much better Zen is core for core vs Atom.

Where are you getting that information on yields?
And there was time while back you could get one pretty cheap

Yields will logically go down as the number of adjacent on the same silicon cores you need operating at a target frequency and voltage range goes up. If even one out of the dozens upon dozens of cores isn't up to snuff you've got to fuse it off and sell it as a lesser product. 16 core chips with one defective core become 14 core, etc. They're cutting into the highest grade silicon they have to bring as many cores to desktop as i9 promises, the silicon usually reserved for server hardware. It's just another example of how desperately they need to find a way to run die's parallel without shit scaling. Right now they'd basically have to use the QPI interface which would result in better yields by being able to put 2 8 core die's on the same chip instead of needing one 16 core die, but the performance would be horrible in comparison.

Don't want your ramble shit
Want your source on yields of knights chip
Remember pascal had pretty Damn good fabs for size
And intel let phi processors go for 200 in 2014 which means they taking huge loss?

Why the hate on intel you haven't explained why

That's proprietary, I don't have evidence beyond the current prices of phi's. Atom cores are so small that they surely did have much better yields than it would be if they were trying to get 72 skylake cores on a die for example, but I'd still assume there's more complexity to them than a single CUDA core.

What even are the die sizes of phi's?

And I can't explain my hate for Intel since I don't have any for them beyond their market segmenting and seeming devotion to give less for more every generation (ever since ivy bridge at least)

This post was edited by DCSS on Jun 28 2017 06:13pm

So they are suppose to dominate and create a monopoly and get fined again?

There is likely to be some degree of corporate sabotage considering how lightly they got off with it last time.