Back of the envelope calculations on a 3 module Kaveri

[Khato]

A dual module quad core, 512 GCN core Kaveri at 28nm should be close to 200mm2, adding another module + 2MB L2 should raise the die size close to ~25mm2.

So, a 6 core Kaveri die should be smaller or at least close to Trinity.

Why should a dual module, 512 GCN Kaveri on 28nm be close to 200 mm^2? Even assuming perfect scaling from 40nm to 28nm, GCN adds about about 5% die space. So if you assume perfect scaling from 32nm to 28nm, then the graphics portion of Kaveri would grow slightly (around 7%) compared to Trinity. Non-GPU logic takes up approximately 57% of the Trinity die, so with perfect scaling that would go down to about 107 mm^2. Meanwhile a 7% growth on the remaining 43% would yield about 113 mm^2 for a total of 220 mm^2. Adding another module would, best case assuming that they only had to lay down another core macro and no supporting logic, bring that up to 245 mm^2. (I decided to take the time to do more accurate estimates this time 'round... which means I should also do the following.)

GLF 28nm SRAM cell size is 0.12 um^2 compared to 0.15um^2 for 32nm. This is a scaling factor of 0.8 versus the 0.765625 ideal. That would bring the numbers up to around 230 mm^2 for the dual module and 256 mm^2 for a triple module. (The 40nm to 28nm estimate on the GCN adder is still reasonably close if just going by sram sizes, only 1% larger than the ideal.)

On the consumer end, I think considerations about die size are quite minimal if they exist at all. The only real relevance it has is in how it might impact retail pricing and product choice. As die size considerations may impact whether AMD decides to put out a 3 module + 512 GCN core part at all or stick to a 2 module top end in their APUs.

I mean would any consumer be upset if Intel had to move full 4 core dies into their i3 line? Would ther be "this die is too large" hate mail?

True, for AMD die size doesn't matter much since they aren't a business trying to make money any more... If they could command a premium for their products then the die sizes that we're talking about would make sense. Instead they continue allocating a greater percentage of the die space to graphics despite the fact that no one is paying them extra for that graphics performance. They've had a solid lead in the mainstream integrated graphics market for two years now and what has it done for them?

 

[SlowSpyder]

@slowspyder

AMD have said Q4 2013

Thanks.

 

[AtenRa]

Why should a dual module, 512 GCN Kaveri on 28nm be close to 200 mm^2? Even assuming perfect scaling from 40nm to 28nm, GCN adds about about 5% die space. So if you assume perfect scaling from 32nm to 28nm, then the graphics portion of Kaveri would grow slightly (around 7%) compared to Trinity. Non-GPU logic takes up approximately 57% of the Trinity die, so with perfect scaling that would go down to about 107 mm^2. Meanwhile a 7% growth on the remaining 43% would yield about 113 mm^2 for a total of 220 mm^2. Adding another module would, best case assuming that they only had to lay down another core macro and no supporting logic, bring that up to 245 mm^2. (I decided to take the time to do more accurate estimates this time 'round... which means I should also do the following.)

GLF 28nm SRAM cell size is 0.12 um^2 compared to 0.15um^2 for 32nm. This is a scaling factor of 0.8 versus the 0.765625 ideal. That would bring the numbers up to around 230 mm^2 for the dual module and 256 mm^2 for a triple module. (The 40nm to 28nm estimate on the GCN adder is still reasonably close if just going by sram sizes, only 1% larger than the ideal.)

GloFo 32nm SRAM Cell = 0.149 μm2
GloFo 28nm SRAM Cell = 0.120 µm2

Almost 20% shrinkage

Glofo 32nm Gate Pitch = 126μm2
GloFo 28nm Gate Pitch = 114μm2

Almost 10% shrinkage

TSMC 28nm SRAM Cell = 127μm2
TSMC 28nm Gate Pitch = 120μm2

Glofo 28nm has 5.5% smaller SRAM Cell size and 5% smaller Gate Pitch than 28nm TSMC.

Now, Cape Verde(HD7770) is 123mm2 at TSMC’s 28nm process. It has 640 GCN cores and 128bit Memory controller. A 512 GCN part at GloFo’s 28nm should be close to 110mm2(perhaps smaller) including the 128bit Memory Controller.

So now we have the CPU modules, the NorthBridge and the PCI-e ports. PD module Including the 2MB L2 Cache is 31mm2 at 32nm. SteamRoller should be a little bigger, let’s say 32mm2 at 32nm. I will take a moderate 20% shrinkage for the 28nm over 32nm(not including the Gate Pitch) and that makes each SteamRoller Module including the 2MB L2 Cache at approximately 25,6mm2. So three modules will be close to 77mm2.

Add 20mm2 for the PCI-e lanes and 30-40mm2 for the NB and the rest and we arrive at the die size of 110+77+20+40 = 247mm2. Well as I have said, close to Trinity.

PS: Full nodes bring ~70% shrinkage and half nodes like GloFos 28nm over 32nm can bring 20 to 30%. That could bring the Quad Module PD die size down to ~250-220mm2 at 28nm.

 

[Khato]

And this is where I'll say that your numbers are just as probable as mine since we're both dealing in gross estimates in order to arrive at precise figures

If going to 3 modules instead of 2 allows AMD to increase their ASP on those SKUs by 10% or more then it makes sense to do so. Well, it could also make sense if it allowed them to make significant gains in marketshare... but good luck with that. Such is what has bothered me about AMD's mainstream offerings since Llano - they have the die size of a premium product but they're competing in the i3 price range.

 

[SiliconWars]

AMD has a WSA to deal with, so the more wafers they use the better. It makes sense for them to have competitive parts no matter what the area penalty is.

 

[Montosaurous]

Some posters seem to be on a crusade against the i3 basically. It is still adequate as a gaming processor. That said, it would not be my choice. Considering the entire cost of a system, buying games, internet, etc, etc. except for very unusual circumstances, it seems foolish not to get at least a low end i5 cpu for a gaming build.

I also think a non-hyperthreaded intel quad will be more than adequate for several years despite the claims of those who think everything is suddenly going to need 8 cores. I would be concerned about a dual core, even with HT though as far as longevity is concerned.

In the long run the more cores the better, though by the time that happens most chips today won't be up to speed. The i3 is no longer viable for a new gaming build, that is a fact. You can still have one for another year or two I guess. I gotta upgrade my chip soon so I can future proof it some (not too sure how good my motherboard will be on overclocking an FX 8320 however). If a new build, opt for a Phenom II 965 BE sub $100, an FX 6300 sub $150 and a low end i5 or FX 8350 sub $200.

 

[SPBHM]

Any kind of serious gaming is going to require a dedicated gpu, even with kaveri. At best it will have only hd7750 level of performance, not really adequate for higher image quality at 1080p and above in modern games.

the 7750 runs at 800MHz and uses GDDR5 (at 4500MHz),
if the APU only uses DDR3, even at 2133 it's going to see a significant decrease in performance,

here is what it looks like with DDR3 1600, it's slower than a 5750/GTS450, a little bit faster than the 6670 DDR5 probably.

hardly a good gaming card for 2014.


even if they run at 900MHz (like some models of the 7750 ), GDDR5 will be the main factor, how much cost will it add?

if they use DDR3, Haswell GT3 could be a problem, GT2 not much.

Some posters seem to be on a crusade against the i3 basically. It is still adequate as a gaming processor. That said, it would not be my choice. Considering the entire cost of a system, buying games, internet, etc, etc. except for very unusual circumstances, it seems foolish not to get at least a low end i5 cpu for a gaming build.

I also think a non-hyperthreaded intel quad will be more than adequate for several years despite the claims of those who think everything is suddenly going to need 8 cores. I would be concerned about a dual core, even with HT though as far as longevity is concerned.

I totally agree,