Can you answer my theories about Nvidia Pascal?

[kenrickandbros]

Introduction:
I have two possible theories about Nvidia Pascal. These theories explain why Pascal isn't even more powerful if the GTX 1080 can go to the crazy speed of 2.1 Ghz in an air overclock at 67 degrees Celsius! This clock speed has only been reached previously in a GPU by using liquid nitrogen to cool! The theories cannot be both be true. Here they are:

Theory 1:
Pascal may suffer from what I call "clock speed inflation". Let me explain what I mean by this with an example: AMD's FX-9590 is factory overclocked at 5 Ghz and it has eight cores. Despite this, it can be outperformed by Intel i5's and i7's even in programs that can utilize many threads. This is because the instructions per clock are not as good as Intel's. Regardless, it's clocked at 5 Ghz. Despite being clocked at 5 Ghz, the FX-9590 is not as good as a (higher priced) Intel four core chip clocked at a lower speed. The FX-9590's clock is not worth as much, hence "clock speed inflation". Apply this idea to GPUs.

Theory 2:
This theory applies more to the GTX 1070. Maybe, Nvidia has not given the GTX 1070 and 1080 enough memory bandwidth. Why would they do this? So that they could hold back cards with higher amounts of VRAM, or in the case of the 1070, possibly a GDDR5X version. GDDR5 and GDDRX differ only slightly in terms of the GPU, but 5X is signifcantly better. There will not be an HBM2 version of the 1070 or 1080 because it would be a waste of time for Nvidia to convert these GPU's into an entirely different and more expensive VRAM format if they could easily slap on another few gigabytes.

I'm not an AMD fanboy for saying that Nvidia might consider doing that, given the opportunity. I acknowledge the fact that there are many past examples of both Nvidia and AMD using the same GPU with different amounts of VRAM:
- A lot of AMD's 300 series is just the 200 series with twice the VRAM
- Some models of the GTX 960 have 2GB of VRAM while others have 4GB
- The 980 Ti is the same chip as the Titan X with half the VRAM and is slightly less powerful.

Conclusion:
These "theories" may be more like hypotheses - oops. I know these theories are unlikely, especially because Nvidia has yet to release "big Pascal" or the Titan and 1080 Ti but please tell me what you think.

 

[RobCrezz]

I dont understand? Are you saying that the pascal performance increases are less than you expected? (despite being one of the highest generational increases ever? if you compare 980 to 1080)

 

[Hawkshot]

I believe more in your idea on Theory 2, I don't think that Clock speed inflation is going to effect the new Pascal GPU's. Saying that I think its a distinct possibility that they are holding back on the amount of power that these GPUS are actually capable of, as you said they still need to release the bigger brothers of the 1080 and 1070 which is likely to be the newer version of the GTX 980ti and the Titan X. Even if they are holding back a little, to be honest im not even mad. The proof is in the pudding as they say, I cant wait to get my hands on a GTX1080 and do some real life testing and Benchmarks as the specs they have shown us are outstanding I want to see how that translates to real world testing.

 

[13thmonkey]

clock speed is marketing BS, unless compared to the same architecture. A 3GHZ P4 is no where near the same performance as a 3GHZ i3/i5/i7 even on single threaded. So ignore clockspeed. Instructions per clock per core are what's important with GPUs and CPUs. It's not inflation, it's just that they can run simpler cores faster (assuming that they are simpler) it's nothing sinister.

As for bandwidth for the 900's there was a compression method used that minimised the impact of lower bandwidth, so if that has been used on the 1000's then we should have a lot more effective bandwidth than before.

There will be a 1080ti, just because of binning and to hold in reserve to counter AMD. Also nothing sinister.

The cards will be what they will be.

 

[RobCrezz]

Exactly, plus there is good reason for using smaller 256bit memory buses, power consumption. If you can use other methods for getting the memory bandwidth then its worth it to keep the TDP down.

 

[BadActor]

I guess I don't understand your theories either. The first one you call clock speed inflation is what I believe is commonly referred to as "Instructions Per Clock Cycle" (IPC). Intel has long had the advantage over AMD here, effectively doing more "work" per clock cycle even though not doing the cycles as quickly. If the AMD can only complete half the tasks of an Intel per clock cycle, it would have to run at twice the speed just to do the same amount of "work."

As far as your second theory, I'm sure the engineers knew what they were doing when they assigned the amount of RAM to the new cards. There is more to the story than just the quantity of RAM being used, such as the frequency, architecture and driver support. I agree that intentionally crippling the performance may have occurred, but this isn't uncommon, either. They do this with CPU's by using lower binned parts or disabling certain features of the chip to create different levels of performance. This proportionally spreads the manufacturing costs over the entire product line even though it costs the same to make the different variations.