Ok so if theyre exaggerating @ 10x more powerful, can we atleast hope 3-4x more powerful??
Best case scenario for NV is a new architecture + a node shrink. Last time they did this was GTX580 -> 780Ti and we got 2X the performance increase but that was accomplished in a period of 3 years (Nov 2011 GTX580 --> Nov 2013 GTX780Ti), not 2 years. Since Titan X (GM200) launched March 17, 2015, it should take roughly 3 years for NV to double it in performance. If we look at AMD, it's similar. HD7970 was announced Dec 2011 and launched Jan 2012. R9 390X should be 2-2.25x faster around June 2015 or roughly 3.5 years from launch.
The average GPU growth has been on pace at 33-35% per year and has been that way since HD5870 (Sept 2009).
Sept 2009 HD5870 (100%) --> Nov 2013 GTX780Ti (329%):
growth rate of 1.3475X (or 34.75% per annum)
http://www.computerbase.de/2013-12/grafikkarten-2013-vergleich/10/
As you can tell, this pace roughly continues since R9 290X/780Ti launched in November 2013 with the Titan X. Titan X launched roughly 1.5 years since R9 290X/780Ti and its performance is about 45-53% faster. (1.3475X * (half rate to account for half the year or 1.17375) = 1.58X. We are actually
slightly behind pace right now but R9 390X/consumer GM200 haven't launched so we don't have the full picture for 1H of 2015 yet:
http://www.computerbase.de/2015-03/nvidia-geforce-gtx-titan-x-im-test/3/
There is no indication that NV will break the laws of physics. If anything, GPUs used to double in performance every 18-24 months but now it's taking AMD/NV 30-36 months or so.
A rough estimate then is Titan X's performance should increase 2.45X (1.3475^3) around March 2018. I am just using historical data and it's not entirely fair since this time might be a bit earlier due to HBM2 + 14nm. Since Pascal is on the road-map for 2016, I think 3-4x faster is just a pipe dream for Pascal, but Volta is a different story. Also, it's not an exact science. For example, we can't just expect some random GPU to offer 34.75% more performance on March 2016, but it's a rough estimate that by 2H of 2016 we should have at least 1 product like that compared to the Titan X.
I've said several times on this forum since reviews hit that GM200 is off somehow. Either it's 12gb of 7ghz vram is eating too much into it's TDP, or the chip needs another spin, or perhaps 600mm2 is simply too big to nail everything down perfectly. Both GF110 and GK110 were as efficient as their lesser chips (if not slightly more so), but GM200 loses a noticeable amount of efficiency over GM204. GM200 is disappointing in both stock performance and perf/w, from an expectation and precedence standpoint.
I am more inclined to believe that Titan X is not the peak version of GM200 product. As you have alluded, wait another 4-6 months to bin GM200/respin it, drop 12GB for 6GB, allow AIBs to provide open air cooled solutions, and it's possible to have a GM200 with boost clock of 1216-1241mhz out of the box.
EVGA already has 2 such cards:
EVGA Titan X Superclocked = 1216mhz Boost
EVGA Titan X HydroCopper = 1241mhz Boost
http://hexus.net/tech/news/graphics/81787-nvidia-geforce-gtx-titan-x-partner-cards-roundup/
That's about 15% higher clocks than the reference Titan X. With these characteristics, GM200 would definitely live up to the hype/expectations! I have to acknowledge what NV engineers pulled off though. Titan X doubles R9 280X in performance within a similar power usage on the same node. That is remarkable even though it slightly trails GTX980 in performance/watt. Granted, Tahiti uses early gen 28nm node that's far from optimized and it's a DP monster while the Titan X is a pure gaming card with almost non-existent DP capability, while R9 280X has 1.075Tflops of DP.
If NV priced the Titan X at $650 or so and provided after-market open air options for AIBs, I don't see why HD7970/7970Ghz owners would have a reason to wait for R9 390X. Alas, NV didn't do that.....
Facebook
Twitter