Nvidia GT300 specs - for real?

Let's look into this with some critical thinking, shall we?

These are the leaked numbers, 2.4 billion transistors, 512 "1D" shaders (which seems to be the G80 based arch that is the GT200) and less than 490mm2 in a 40nm process.

Assuming that Nvidia would perform the, somewhat insane, act of actually building another huge chip, the die size estimates for a chip with 2.4billion transistors holds up.

Let's look at the RV770's 55nm chip vs RV740's 40nm chip:

• RV770: 956 million transistors, 800 shaders, 260mm2
  
• RV740: 826 million transistors, 640 shaders, 137mm2

So, we have around the same transistor budget and almost half the size of the RV770 - let's call this the slight "overhead" on our rough calculations.

Now, the GT200 has 1.4 billion transistors, so 2.4 is not doubling it, and we have a reduction from 65nm to 40nm, based on the table above, and not from 55nm to 40nm. This tells us that the same chip, on a new process, should have around 38% of the size and a die size of 375mm2:

• 40*40/(65*65) = 0.38
• 2.4/1.4*0.38*576 = 375mm2
  

These are areas we're talking, so everything is a square of something, resulting in more reductions than what is expected just by looking at the fabrication process size.
All summed up, we should expect less than double the die-size, since we're now down to 375mm2 plus our "overhead", we creep up to the speculated value of 490mm2 fast. Plausible then.

There's one other factor. These chips are planned around three years in advance, so by the time Nvidia was envisioning the GT300, it had no idea that the GT200 would screw up against AMD's offers. Such a huge die was plausible given past history and high margins.

There's been a lot of speculation about the type of shader cores in use for the next chip, with some people equating that it may not look like this "1D", which I suppose means we don't have vector units here.
AMD uses vector units as it can do a 128-bit vector operation(4 floats) in a clock cycle per shader(that's 160 real, not 800) and a transcendental operation - (4+1)*160 = 800, you get the picture.
Intel has also gone the vector way and Larrabee will have an undisclosed number of cores that are capable of vector operations on 512bit, or 16 floats - in essence, 16 shaders for the marketing people - plus another type of multiple issue that isn't relevant right now.

Given AMD's rather good die size efficiency, especially in games like GRID, Nvidia would probably benefit from a move to vector units for the basis of it's shader processors. The chart above doesn't seem to reflect this in either die-size or shader count, so that part smells fake to me. Then again, by the time this chip was cooked up, Nvidia didn't anticipate the current market.

There's also the memory bus implementation, 512bit again? Given current offerings, it might come true but that's something to reflect upon on another day.

Most of these specs are mostly fake, as the one from RV770 was, some time ago. Shaders and clocks were way off.

Source: VR-Zone