I'm curious how everyone is so confident that a GM200 Titan X would had terrible DP performance compared to the previous Titan.
Where is the evidence that suggests this? All I've seen is references to the 980's DP performance.
Let's compare:
Model - SP/DP (in GFLOPS) - Die Size - Transistor Count (in millions)
770 - 3213/134 - 294mm - 3540
780 - 3977/166 - 561mm - 7080
780 Ti - 5046/210 - 561mm - 7080
Titan - 4500/1300-1500 - 561mm - 7080
Titan Black - 5121/1707 - 561mm - 7080
980 - 4612/144 - 398mm - 5200
[
All data provided by wikipedia]
After doing some math, what I do see is that Maxwell may have a DP performance about 1/32 of SP for the consumer parts, whereas Kepler had about 1/24.
However, Titan had between 1/4 and 1/5th the performance, and Titan Black had 1/3.
If we extrapolate that with Maxwell, it's likely that a full 985 Ti type of product would likely have around 6912/216 SP to DP. That's likely a low-end guess, based solely on dividing SP/DP into SMMs and extrapolating that to 24 SMMs for the GM200. This may not be a whole picture, as other factors come into play. For instance, making that same comparison of SP/DP to SMM count, the 970 has a lower factor, which demonstrates that the L2 cache and potentially memory controller configuration are likely involved in final compute performance
Now, as DP performance relates to transistor count, the 980 actually has a higher ratio than the 780. It is lower than the 780 Ti. However, this may not be an apt comparison, as the 980 is closer to the 770 in die-configuration. If we compare to that, the GK104 does have a higher ratio, however, it's ratio of DP/SP is essentially the same as the 780 and 780 Ti.
Which means it is harder to actually make a fair guess of what a Titan X class of card would do, at best guess we can possibly get a rough estimate of a consumer-variant of the GM200.
Again, consider these DP/SP values:
780 - 0.04174
780 Ti - 0.04162
Titan - 0.22̅
Titan Black - 0.33̅
980 - 0.03122
So for Maxwell, it does seem obvious that the SP performance has been targeted more strongly than in Maxwell, but not extensively.
I would also say any consumer variant, such as a 985 Ti, will have higher compute performance in both SP and DP compared to the 780 Ti, but it won't be significant on the DP side of the house, essentially any increase owing only to the nearly 1 billion extra transistors over GK110.
Of course, we don't know HOW Nvidia actually removes DP performance, so the DP/SP ratio for a GM200 consumer part might be higher just for the sake of having a little more than past parts. Obviously the actual GPU can handle a much higher DP capability, but I don't know how they achieve this, and if it is a direct scaled factor or simply chosen at will.
And how do you convert those numbers to what a Titan X might achieve?
If you compare the numbers, the Titan had a ~534% relative increase in DP/SP ratio compared to the 780.
The Titan Black? ~800% increase!
Let's try to extrapolate that kind of performance for Titan X from a potential 985 Ti, using the potential values I presented earlier.
Let's also rethink the nomenclature 985 Ti, and settle on 985. This would represent the first consumer release of a GM200, which, as we saw with the GK110, was later refreshed with further performance in the 780 Ti and Titan Black models.
This 985 has 6912 SP, 216 DP, or 0.03125 (DP/SP). Using a modest 500% increase, we would have a 0.15625 DP/SP ratio.
Let's assume that the Titan X has, oh, 7200 GFLOPS for SP performance, a modest increase over this imagined 985. That brings us to 1125 GFLOPS for DP.
That would be above the Titan but not the Titan Black. Perhaps they skip the idea of a refresh and release the most they can in this Titan X so that it's DP performance may be greater than the Titan Black from the start.
For this, we'll again make the SP performance 7200, just in case that is all they can eek out of Maxwell. While the 780 and 780 Ti had nearly the same ratio comparing DP to SP, the 780 Ti did see a significant jump in SP performance (~1000 GFLOPS SP advantage). I won't further muddy this scenario up by attempting to match those same maneuvers for the sake of brevity, not to mention this may as well by all imaginary numbers until we have confirmed specs for the Titan X.
But let us wrap this up. At 7200 GFLOPS of SP performance, an 800% increase would be: 1800 GFLOPS of DP performance.
A small bump over Titan Black: 7200/1800 over 5121/1707. A fairly significant SP increase but modest DP increase.
With that said, it will be rather interesting to see what they managed to extract out of the GM200. And whatever we see, is that the max from the start? Or might they have a refresh down the line?
I'd be shocked if they release a Titan that costs more than the Titan X ever did, yet does not perform the same in Double Precision compute.
For those that want to see the numbers I worked with, and if they want to extrapolate any other fantasies, be my guest.