The die shots of the GM200 GPU were spotted by Videocardz.com a few hours ago, and seeing that these are the first authentic images of the GM200 core, I seized the chance to do a die analysis. The pictures weren't as high quality as I had hoped, but they will have to do. Before we begin, in all fairness, I should point out some things that could make this experiment inaccurate: lens distortion, inaccurate perspective correction and warping due to rolling shutter just to name a few.
Nvidia GM200-400-A1 Core and PCB Analysis - Die Size and Memory Specs Revealed
The die size of a GPU is a very interesting specification. It indicates just how much more power can be squeezed out of the current combination of architecture and process size. Since TSMC's limit is 650mm^2 if our results are anywhere close to this than Nvidia's next die will either have to have 1) A brand new architecture or 2) process shrink with the same architecture. Lets begin with some preliminaries first. Our reference die shot is an upright shot, high resolution picture of the GM204 die - in other words, a perfect candidate for comparison.
Our GM200 die shot was taken from the leak that occurred just a few hours ago and is not so helpful a case. Not only is the perspective distorted but it isn't that high res either. Scaling it and correcting the perspective distortion will prove to be slightly tricky. Given above is a side by side comparison of our two starting images. For those that are curious, here is how my correcting methodology: Since the base plate surrounding the GM200 is obviously a square, with a few tries I got it more or less upright, perspective wise. For aligning and scaling purposes I used the holes (where the heat sink usually screws in) to match the two images and correct the relative size. Since we know for a fact that both holes will be of the same size, we can approach almost 100% accuracy. We finally get the following result:
The fact that the half cropped screws of each PCB align perfectly means I was successful in maintaining a very high level of accuracy while doing perspective/scaling correction.
Now we come to the slightly tricky part: Pixel analysis and comparison. However during pixel selection I hit a slight snag. While I was able to get a very very tight fit on the GM204 die, the GM200 die had a slight offset to its rotation (notice the non-die area selected on the right and left side of the core), which will need to be manually removed after the calculation. This won't ruin the integrity of the test but will introduce a certain expected deviation in our result which I have given in the para below. Anyways here are the results of the analysis:
|Die||Scale Units||Scale Factor||Count||Area||Perimeter||Circularity||Height||Width||Grey Value (Mean)||Integrated Density|
Since the GM200 shot is slightly low resolution and it is difficult to decipher the edge correctly due to the offset, so I can expect a maximum error of ± 2.5%. That is a pretty decent margin so I can live with that. As you can see from the circularity data that my selections were more or less consistent. While I have also included the other data I calculated for any enthusiast out there, we will only concern ourselves with the Area for this article. The Area is in Pixel^2 which is great because that means we can directly calculate the die seize of the GM200 using ratio method.
The GM204 has an area of 10803 and a die size of 398mm^2, while the GM200 has an area of 17928 and an unknown die size: X. Using simple ratio we arrive at the following answer = 660.49mm^2 Well well well, but just before we wrap this up we need to account for those empty pixels I mentioned back and we arrive at the following answer: 632mm^2± 2.5%.
Wow, that is a pretty big die. That is pretty close to TSMC's max size so this is really as 'Big' as Maxwell can get. After this Nvidia will either have to move to a smaller node if it wants to iterate Maxwell or introduce a brand new architecture if god forbid it stays on the same node. The PCB also tell us exactly which kind of memory the GM200 board will be using. Since there are exactly 24 chips with the code name H5GC(Q)4H24MFR, a simple lookup at SK Hynix's database confirms the specification we already know. GM200's board is rocking 24 chips of 4 Gigabit each, dividing by 8 we get 512 MBs, a total of 12 GB of vRAM. As far as the nomenclature is concerned, a better variant of this die (in the future) will be called the GM200-400-A2 while a process port will be called GM200-400-B1 if Nvidia decides to stick with the GM200 branding. That's all folks.
— Usman Pirzada (@usmanpirzada) January 16, 2015