By using this site, you agree to our Privacy Policy and our Terms of Use. Close
Pemalite said:
forethought14 said:

Well, if Latte is 40nm, and if its total die size is 156.21mm2 (those measurements are pretty accurate with two decimal places), and we know it uses AMD Radeon technology, and the 40nm cards seem to average around 6 - 6.25 million transistors per squared millimeter, then we could use simple arithmetic, and easily calculate the transistor count:

156.21 millimeter squared x 6,000,000 (million transistors / millimeter squared) = minimum 937,260,000 transistors for all of the Latte die (including the eDRAM). eDRAM is roughly 220 million transistors for 32MB 40nm Renesas eDRAM, so GPU logic for Latte is:

937 million (total die) - 220 million (eDRAM) = 717,260,000 (estimate) minimum transistors for pure GPU logic. "Logic" includes everything like Shaders, Texture mapping units, GPRs, Tessellator, etc.

There is no point calculating it.
You could be off by 20-30% easily because you can pack transisters in more densely on a mature process verses one that isn't, when both are compared at the same fabrication process.

You can also use transisters with better power characteristics if you are targeting a low clockrate too, which can save a ton of power by allowing you to be aggressive with low voltages and minimising leakage.

I did happen to say it was an "estimate" I'm sure it's possible that I'm inaccurate +/- by several tens of millions. Though, the point was to get the idea across that it had a very large amount of transistors for a conventional 40nm 160 shader part (most of them are around 300 million), even more than the Radeon GPUs that released in 2012 with 160 shaders (those likely share the similar mature 40nm processes as Latte, the lower-end 7000 series GPUs that used 40nm and with VLIW core architecture). Though, I'm pretty sure that the million transistors per mm2 aren't going to be significantly more different considering how no 40nm GPU had transistors per mm2 substantially off what I averaged (though I've found mainly 6.01 +/- 0.1 averages, I've heard mention of numbers as high as 6.25, so I included that in). Unless Latte used certain transistors that no other commercial GPU has used before....