Unfortunately TPU doesn't give average performance metrics further back than Tesla 2.0 launched in Dec 08 in its fastest GTX 285 iteration, but that allows us to compare all the way back to the good old 8800GTX launched in Nov 2006, 14 years ago, and all the way back to the start of the PS3 console generation. That covers 9 architecture iterations so should be a reasonable sample for predicting future performance scaling.
Looking at the fastest iterations of each architecture (using Ti's rather than Titans where they are available) which seem to have launched roughly every 2 years gives us the following performance uplifts architecture to architecture starting with the GTX285 uplift over the 8800GTX and ending with the 3090 uplift over the 2080Ti. All comparisons are taken at the time of the newer GPU's launch thus removing any driver optimisation questions for the older GPU and if anything favouring the older GPU (especially in the Turing/Pascal comparison) given the new architectures inability at that point to stretch it's legs. I used resolutions appropriate to both GPU's at the time of comparison.
Code:
Resolution Performance Uplift New Architecture Launch Date
3090 2080Ti 4k 145% Ampere Sep-20
2080Ti 1080Ti 4k 139% Turing Sep-18
1080Ti 980Ti 1440p 175% Pascal Mar-17
980Ti 780Ti 1440p 141% Maxwell Jun-15
780Ti 580 1080p 169% Kepler Nov-13
580 285 1080p 164% Fermi 2.0 Nov-10
285 8800GTX 1680*1050 149% Tesla 2.0 Dec-08
Since the table's not so easy to read, the summarized performance uplifts roughly every 2 years, oldest to newest are 49%, 64%, 69%, 41%, 75%, 39%, 45%. While there may be a very slight downwards trend there, we have to remember that Turing (the 39%) introduced Ray Tracing which was a major change in approach to real time graphics. And where RT is used, that 39% increase would obviously be far, far larger.
EDIT: I forgot to mention, the cumulative increase from the 8800GTX to the RTX 3090 over 14 years is about 2000%.
So the 3090 is roughly 20x faster than the 8800GTX.
And that;'s literally a 100% increase in bandwidth.
There is a difference between what the chips did/are doing vs what they are capable of. 980ti was one of the most conservatively clocked card there has ever been and the 41% it has on your table does not fully represent what that chip could do. There were customer cards that out of the box offered 20% more performance and still could OC 10% on top of that. The chips after that were clocked closer to the max in the reference model. If you take that into account the downward trend is slightly steeper especially when you combine it with the upward pointing trends in die sizes and power consumption.
