Anand's die size calcs don't seem to add up?

[Hulk]

http://www.anandtech.com/show/6993/intel-iris-pro-5200-graphics-review-core-i74950hq-tested/4

He measures the quad core/GT3e dies at 264mm^2 plus 84mm^2 for Crystalwell
Intel provided the data that a quad core/GT2 is 177mm^2

So Anand surmises that the extra GT3 hardware is 264-177=87mm^2

Then he assumes that the full 40 GT3 EU's would requires twice that amount of die space, or 174mm^2. This is for the full GT3 GPU by itself.

He next calculates the GPU percentage for the quad/GT3 Haswell as 177/(174+90)=67%
Okay that's fine.

Then he compares that to GPU percentage for GT2 Quad Haswell, which he writes is about 33%. I don't see how he comes up with that number. If the GT2 is assumed to be half the size (20 vs 40EU's) then the calc would be 87/177, or about 50%, not 30%.

Of course the initial assumption that GT2 is 87mm^2 and GT3 is 174mm^2 is most likely flawed but I'm still wondering where he got 33%?

But next is where it gets really weird.

Finally, in the last row of his table he shows the dual/GT3 Haswell having a die of 181mm^2. If the entire GT3 area is 174mm^2 then that only leaves 7mm^2 for the dual core Haswell CPU!


?

 

[sm625]

I dont get where the 181mm² comes from. If 20 EU adds 87mm² then the first 20 EU would be 87mm² plus a very significant amount of area that doesnt need to be replicated for the 2nd block of 20EU. It would have to be at least 10mm². Which means the entire GPU on the 264mm² chip would have to consume at least 184mm², leaving just 264-184 or 80mm² for the cpu.

However I think we can throw all of this out the window because it is most likely that the GT3 die contains a great deal of transistors dedicated solely to interfacing with crystalwell. Lets say just for a wild guess that this section adds 25mm². That means that the GT3's second block of 20EU really only adds (264-25)-177 or 62mm². And that means 40 EU = 62mm²x2 = 124mm² rather than the outrageous 174mm² that Anand came up with. Given the stated dual core die size of 181mm², this 124mm² for the EUs sounds much more reasonable. It leaves 57mm² for the cpu cores and the system agent.

 

[Khato]

Yes, as stated in a previous post I made on the subject - http://forums.anandtech.com/showpost.php?p=35088748&postcount=166 - I'm guessing that Anand's lack of time/sleep was responsible for the incorrect line of logic there. I also provide rough estimates for GT2 graphics die space based off the available die photo + the now known die size, and GT3 graphics should be 2x the area of GT2 or less.

 

[mikk]

Yes, as stated in a previous post I made on the subject - http://forums.anandtech.com/showpost.php?p=35088748&postcount=166 - I'm guessing that Anand's lack of time/sleep was responsible for the incorrect line of logic there. I also provide rough estimates for GT2 graphics die space based off the available die photo + the now known die size, and GT3 graphics should be 2x the area of GT2 or less.

GT2 is 53-55 mm² and GT3 ~104 mm². You calculation is a bit off. Also your 2x the die over Trinity is wrong as well. It is 1.5x the die area.

 

[Hulk]

So I did a little very rough math.
We know that 4C + GT2 Haswell is 177mm^2 from Intel
We also know that 2C + GT3 Haswell is 181mm^2, again from Intel
If you sub GT2 as 20EU and GT3 as 40EU and solve the equations simultaneously you get the following very rough averages.

1 Haswell CPU core is 28.83mm^2
1 Haswell EU is 3.082mm^2

Using those number we can calculate the die size of Haswell 4C + GT3 at 239mm^2 which is close to the 264mm^2 that Anand measured with calipers on the actual die.

Haswell 2C +GT2 calculates as 119mm^2 which also seems like a pretty good estimate.

To sum up.
Desktop/Mobile Haswell appears to consist of 4 different parts.

Quad core with GT2 graphics with a die size of 177mm^2 (Intel supplied)
Dual core with GT3 graphics with a die size of 181mm^2 (Intel supplied)
Quad core with GT3 graphics with a die size of 264mm^2 (measured by Anand)
Dual core with GT2 graphics with a die size of 119mm^2 (calculated from Intel supplied data)

Using this data we can estimate GPU die percentage as follow:
Quad core GT2 - 35% GPU
Dual core GT3 - 68% GPU
Quad core GT3 - 57% GPU
Dual core GT2 - 51% GPU

Furthermore given the fact that Quad/GT2 is 1.4B transistors and dual/GT3 is 1.3B we can solve those equations to estimate transistors for the remaining two cores.
Quad/GT3 estimated 1.8B transistors
Dual/GT2 estimated 990M transistors

Yes I know I'm not including many, many variables in the actual die components but this was just for fun given the actual info Intel gave us.

 

[TuxDave]

So I did a little very rough math.
We know that 4C + GT2 Haswell is 177mm^2 from Intel
We also know that 2C + GT3 Haswell is 181mm^2, again from Intel
If you sub GT2 as 20EU and GT3 as 40EU and solve the equations simultaneously you get the following very rough averages.

1 Haswell CPU core is 28.83mm^2
1 Haswell EU is 3.082mm^2

Using those number we can calculate the die size of Haswell 4C + GT3 at 239mm^2 which is close to the 264mm^2 that Anand measured with calipers on the actual die.

Haswell 2C +GT2 calculates as 119mm^2 which also seems like a pretty good estimate.

To sum up.
Desktop/Mobile Haswell appears to consist of 4 different parts.

Quad core with GT2 graphics with a die size of 177mm^2 (Intel supplied)
Dual core with GT3 graphics with a die size of 181mm^2 (Intel supplied)
Quad core with GT3 graphics with a die size of 264mm^2 (measured by Anand)
Dual core with GT2 graphics with a die size of 119mm^2 (calculated from Intel supplied data)

Using this data we can estimate GPU die percentage as follow:
Quad core GT2 - 35% GPU
Dual core GT3 - 68% GPU
Quad core GT3 - 57% GPU
Dual core GT2 - 51% GPU

Furthermore given the fact that Quad/GT2 is 1.4B transistors and dual/GT3 is 1.3B we can solve those equations to estimate transistors for the remaining two cores.
Quad/GT3 estimated 1.8B transistors
Dual/GT2 estimated 990M transistors

Yes I know I'm not including many, many variables in the actual die components but this was just for fun given the actual info Intel gave us.

If you left it up to me, I would've calculated it the same way as you just did. /thumbs up

 

[Idontcare]

If you left it up to me, I would've calculated it the same way as you just did. /thumbs up

That's a "/thread" right there folks QED

 

[mikk]

So I did a little very rough math.
We know that 4C + GT2 Haswell is 177mm^2 from Intel
We also know that 2C + GT3 Haswell is 181mm^2, again from Intel
If you sub GT2 as 20EU and GT3 as 40EU and solve the equations simultaneously you get the following very rough averages.

1 Haswell CPU core is 28.83mm^2
1 Haswell EU is 3.082mm^2

Using those number we can calculate the die size of Haswell 4C + GT3 at 239mm^2 which is close to the 264mm^2 that Anand measured with calipers on the actual die.

GT3 is a bit below a doubling since non-slice parts in GT3 are the same as in GT2. We have a die picture from GT3 here: ftp://download.intel.com/newsroom/k...h_Generation_Intel_Core_Dual_Core_Hero_HR.jpg

This is slighty more than 100 mm².

 

[AtenRa]

Yes according to the Dual Core GT3 pic mikk has linked, *GT3 seams to be close to 100-103mm2 and *GT2 close to 50-55mm2 (Without the Display Controllers) at 22nm. At 32nm GT3 would have been roughly at 163-168mm2 and GT2 close to 81-90mm2.

* The Display Controllers in SB/IV and HW are situated on the other side of the die close to the PCI-e and System Agent .

Trinity/Haswell Richland iGPU (without the display controllers) is roughly 98mm2 at 32nm, at 28nm it could be close to ~70mm2.

 

[mikk]

Trinity GPU is 103 mm² big according to AMD. Do you have a source for process scaling from 32nm to 28nm? Also your 16x mm² is too high. IVB scaled only ~1.5 down.

 

[AtenRa]

Trinity GPU is 103 mm² big according to AMD. Do you have a source for process scaling from 32nm to 28nm? Also your 16x mm² is too high. IVB scaled only ~1.5 down.

Trinity iGPU is 103mm2 including the Display Controller. But since we dont count the display controllers on Intel GT3 and GT2 we have to do the same for AMD Trinity, thus we get close to 98mm2 for Trinity's iGPU.

SB at 32nm = 995M Transistors with 216mm2 die size
IV(GT2) at 22nm = 1.2B transistors with 160mm2 die size.
HW(GT2) at 22nm = 1.4B Transistors with 177mm2 die size

SB to HW we have ~1.63

According to the specs known, GloFo 28nm could give up to 30-35% shrinkage over 32nm.

 

[mikk]

m2 die size

SB to HW we have ~1.63

What is your source?

 

[AtenRa]

What is your source?

Just gave you the transistor count and die sizes of SB, Ivy and Haswell above.

 

[Hulk]

Yes according to the Dual Core GT3 pic mikk has linked, *GT3 seams to be close to 100-103mm2 and *GT2 close to 50-55mm2 (Without the Display Controllers) at 22nm. At 32nm GT3 would have been roughly at 163-168mm2 and GT2 close to 81-90mm2.

* The Display Controllers in SB/IV and HW are situated on the other side of the die close to the PCI-e and System Agent .

Trinity/Haswell Richland iGPU (without the display controllers) is roughly 98mm2 at 32nm, at 28nm it could be close to ~70mm2.

If you assume GT3 is 104mm^2 then using
2C + GT3 = 181mm^2 (from Intel according to Anand)
then you get one Core = 38.5mm^2

using
4C + GT2 = 177mm^2 (from Intel according to Anand)

You end up with GT2 being 23mm^2 which doesn't make sense.

 

[AtenRa]

If you assume GT3 is 104mm^2 then using
2C + GT3 = 181mm^2 (from Intel according to Anand)
then you get one Core = 38.5mm^2

using
4C + GT2 = 177mm^2 (from Intel according to Anand)

You end up with GT2 being 23mm^2 which doesn't make sense.

Hell i made a mess, ill recalculate the sizes again

 

[mikk]

Just gave you the transistor count and die sizes of SB, Ivy and Haswell above.

Meaningless due to different GPU and CPU (Haswell). IVB scaled 1,5 down. Apart from the GPU nothing changed to SB basically.

 

[AtenRa]

If Dual Core GT3 die is 181mm2, from Intel's die picture we get roughly the following,

GT3 = 101.5mm2
GT2 = 54mm2
core + L3 Cache at 20,2mm2
core+L2 without L3 is ~13,5mm2

System Agent, PCI-e ports, Display engines etc is roughly at 21,6mm2
Memory Controller is roughly 9,1mm2

Dual core GT3 die = 101,5 + (20,2 x 2) + 21,6 + 9,1 = 172,6mm2
plus the power gating logic around the die and between the cores and we get close to 181mm2

Quad core GT3 would be 101,5 + (20,2 x 4) + 21,6 + 9,1 = 213
plus the power gating logic around the die and between the cores and I believe it would be close to 239mm2.

Quad Core GT2 would be 54 + (20,2 x 4) + 21,6 + 9.1 = 165,5
plus the power gating logic around the die and between the cores + ~2mm2 for the empty space next to memory controllers and we are close to 177mm2.

 

[AtenRa]

Meaningless due to different GPU and CPU (Haswell). IVB scaled 1,5 down. Apart from the GPU nothing changed to SB basically.

From SB to IV = 1.63
From SB to HW = 1.71

If it would be 1.5 then Quad core IvyBridge GT2 die would be ~174mm2 not 160mm2.

 

[mikk]

From SB to IV = 1.63
From SB to HW = 1.71

Once again you don't have a source.

 

[Enigmoid]

Once again you don't have a source.

Kinda easy to calculate.

SB 4 core = 1.16 billion transistors @ 216 mm^2.
Density = 5.37 million transistors per mm^2

IV 4 core = 1.4 billion transistors @ 160 mm^2.
Density = 8.75 million transistors per mm^2

Ivy bridge then increased the density by 63% (8.75/5.37).

Do the same for haswell.

 

[mikk]

Kinda easy to calculate.

SB 4 core = 1.16 billion transistors @ 216 mm^2.
Density = 5.37 million transistors per mm^2

IV 4 core = 1.4 billion transistors @ 160 mm^2.
Density = 8.75 million transistors per mm^2

Ivy bridge then increased the density by 63% (8.75/5.37).

Do the same for haswell.

That doesn't work. GPU logic has a different package density. iGPU percentage increased heavily in IVB and Haswell compared to Sandy Bridge. Your calculation is flawed.

 

[Enigmoid]

That doesn't work. GPU logic has a different package density. iGPU percentage increased heavily in IVB and Haswell compared to Sandy Bridge. Your calculation is flawed.

Margin of error.

Okay lets try this.

SB->IB->HW igp percentage increases.

Looking at whole die scaling is 63%. Question now is if igp transistor density is different than cpu core + cache density. Also if igp transistor density relative to the cpu cores changed between generations (i'm guessing not within +/- 15%).

I don't have the numbers for this and if you do feel free to correct me but in the absence of some more numbers scaling looks to be overall 1.63 from SB to IB (probably margin or error of 15% at most).

 

[mikk]

Margin of error.

Okay lets try this.

SB->IB->HW igp percentage increases.

Looking at whole die scaling is 63%. Question now is if igp transistor density is different than cpu core + cache density. Also if igp transistor density relative to the cpu cores changed between generations (i'm guessing not within +/- 15%).

I don't have the numbers for this and if you do feel free to correct me but in the absence of some more numbers scaling looks to be overall 1.63 from SB to IB (probably margin or error of 15% at most).

Sandy Bridge GT2 216 mm² - 17% for the graphics = 37 mm²
Ivy Bridge GT2 160 mm² - 27% for the graphics = 43 mm²

216-37= 179 mm² Sandy Bridge GT2 without graphics
160-43= 117 mm² Ivy Bridge GT2 without graphics

179:1,5= 119 mm²

Because Ivy Bridge basically didn't change excluding graphics it should be a pretty accurate measurement and it is almost a 1,5 scaling down.


Or we can do this the other way around.

SB GT2 graphics= 37 mm²
IVB GT2 graphics= 42 mm²

42*1,5= 63 mm²

179 mm²+63 mm²= 242 mm²

242:1,5= 161 mm².......almost matches the size of IVB GT2.

 

[Enigmoid]

Sandy Bridge GT2 216 mm² - 17% for the graphics = 37 mm²
Ivy Bridge GT2 160 mm² - 27% for the graphics = 43 mm²

216-37= 179 mm² Sandy Bridge GT2 without graphics
160-43= 117 mm² Ivy Bridge GT2 without graphics

179:1,5= 119 mm²

Because Ivy Bridge basically didn't change excluding graphics it should be a pretty accurate measurement and it is almost a 1,5 scaling down.


Or we can do this the other way around.

SB GT2 graphics= 37 mm²
IVB GT2 graphics= 42 mm²

42*1,5= 63 mm²

179 mm²+63 mm²= 242 mm²

242:1,5= 161 mm².......almost matches the size of IVB GT2.

I think your igp percentages are off.

By rough calculation SB is 22-25% igp.

http://www.anandtech.com/show/5174/why-ivy-bridge-is-still-quad-core

The SNB graphics occupy roughly 25% of the total die area, or the space of three cores if you prefer to look at it that way,

Ivy bridge is even more igp (looks like 32-35% igp).

Thanks for the calculations though.

 

[mikk]

You are definitely wrong here. 32-35% for IVB would be in the Haswell range. Even AMD confirms: http://cdn.wccftech.com/wp-content/uploads/2013/04/AMD-Richland-vs-Haswell-GPU.jpg