To what degree does Intel harvest dies?

[Yuriman]

I'm not familiar with all of the die variants with Haswell, but as an example, I assume desktop i5's are cut from the same wafers as i7's. 2MB of cache is disabled, as well as the transistors used in hyperthreading. How much die space is unused in an i5? Or, from a different perspective, what percent of defectives can be made into an i5? Does Intel bin i7's with less than all of the cache working as i5's, or do they always fuse off the same parts? How does their scavenging/binning differ from other companies, like AMD, nVidia, etc.?

Are the transistors associated with AVX a large part of the die in dual core models?

AFAIK, iGPUs now take up a majority of the die space in 1150 chips. What kind of binning/harvesting does Intel use here? Wouldn't it make sense then to offer GT1 variants of quad core chips if trying to maximize yield?

 

[ShintaiDK]

The only possible thing you can harvest between an i5 and i7 is the cache block. But usually there is so much redundancy that all i5s still got 8MB, just with 2MB disabled.

You cant harvest from AVX, HT etc. You harvest from dead/binned/tdp cores or dead cache blocks. And cores is something Intel only do in the server line.

 

[BSim500]

"AFAIK, iGPUs now take up a majority of the die space in 1150 chips. What kind of binning/harvesting does Intel use here? Wouldn't it make sense then to offer GT1 variants of quad core chips if trying to maximize yield?"

I know the i5-3350P was basically regular i5/i7-3xxx with defective iGPU's that were fused off as a budget i5. Makes you wonder why there's no Haswell "P" equivalent. And a lot of people would agree with your suggestion. Or even a gamer "P" version with no iGPU at all. Non-K i5 Ivy's had HD2500 instead of HD4000, so GT1 is no issue at all if you've got a decent dGPU. There's really no need to launch a "who has the biggest iGPU" p*ssing contest vs AMD on every CPU...

I haven't verified these figures, but from what I remember, it's something like:-

Sandy i3 = 131mm - 6EU / HD2000
Sandy i3 = 149mm - 12EU / HD3000
Sandy i5 = 216mm - 12EU / HD3000

Ivy i3 = 94mm - 6EU / HD2500
Ivy i3 = 118mm - 16EU / HD4000
Ivy i5 = 132mm - 6EU / HD2500
Ivy i5 = 160mm - 16EU / HD4000

Haswell i3 = 102mm - GT2
Haswell i3 = 180mm - GT3
Haswell i5 = 185mm - GT2
Haswell i5 = 264mm - GT3

Haswell CPU die percentage:-
i3 GT2 = 51% GPU / 49% CPU
i3 GT3 = 68% GPU / 32% CPU
i5 GT2 = 35% GPU / 65% CPU
i5 GT3 = 57% GPU / 43% CPU

Looking at Ivy i3-3225 HD4000/16EU (118mm) vs Ivy i5-3470 HD2500/6EU (132mm), technically if they created an i5-P version with no iGPU at all, they could make an i5-3570P for virtually the same cost as an i3-3225, and probably something similar with the Haswells. Food for thought as to just how "free" the iGPU is...

 

[zir_blazer]

I think that to make a proper comparison, you first need to be aware about how many physically different dies are. Haswell had several dies with different amount of Cores, Cache and GPU. Not sure if this list is fully correct (There seems to be several parts overlapping, may be very much the same die on another package), but it gives you an idea.
It doens't looks like Intel does heavy die harvesting in Core count. The most significant thing that Intel seems to disable is 2 MB Cache L2 on all Core i5 Desktop Haswells. The mobile and ultrabook segment is pretty much diving into the unknow due to the variety of dies.

 

[VirtualLarry]

I wouldn't be horribly surprised to find batches of dual-core Celerons / Pentiums / i3 CPUs, that were made from harvested 4-core dies.