The reversal of roles might seem pretty amazing, considering how much credit ATI got for staying conservative during the time of their greatest success, R300. Of course, the other part of the story that it was their superior architecture that gave them the luxury of going with a tried and true process while Nvidia had to push the manufacturing envelope in order to make NV3x at least somewhat competitive.
Obviously, IHVs don't just select a given process for the heck of it. Following R300 and R420, ATI's design parameters have all but required high frequencies to remain competitive. As such, pushing the process became essential.
As I recall, Nvidia beat ATI to 130nm by about 6 months; the 110nm parts came within months of each other with ATI being ahead by a month or so. ATI was 3 months ahead to 90nm (would have been 6 or more had R520 not been delayed) and almost 9 months to 80nm.
Here are GPU frequencies for high-end Nvidia cars, starting with NV25:
300MHz, 500MHz, 450MHz, 475MHz, 400MHz, 430MHz, 550MHz, 650MHz, 575MHz.
For ATI, starting with R200:
275MHz, 325MHz, 380MHz, 412MHz, 520MHz, 540MHz, 625MHz, 650MHz, 750MHz.
Nvidia has been bouncing around 500MHz range for 4 years after making that huge 200MHz jump for NV30 while ATI has been on a steady, consistent frequency ramp.
Let’s look at the extremely simplistic “execution difficulty index†which we can get simply by multiplying number of transistors by frequency and normalizing for R300:
Code:
Freq Tra # ED-Index Freq Tra # ED-Index
R200 275 60 0.47 NV25 300 63 0.54
R300 325 107 1 [B]NV30[/B] 500 125 [B]1.80[/B]
R350 380 107 1.17 NV35 450 130 1.68
R360 412 107 1.27 NV38 475 130 1.78
R420 520 160 2.39 NV40 400 222 2.55
R480 540 160 2.48 G70 430 300 3.71
[B]R520[/B] 625 321 [B]5.77[/B] G70 550 300 4.74
R580 650 384 7.18 G71 650 274 5.12
[B]R600[/B] 750 720 [B]15.53[/B] [B]G80[/B] 575 690 [B]11.41[/B]
I find it interesting that every single part since R300 which had EDi > 2 of the preceding part has suffered delays.
