NVidia Pascal P100 announced
#11
(04-10-2016, 07:26 AM)epixoip Wrote: It certainly does matter what Titan X shows under hashcat because this is the hashcat forums and how the card performs under hashcat is all we care about...

Sure. But you cannot compare spec facts of one product against real world experience of another. Spec vs. spec and later real vs. real. So far there is an advantage of 50% on the paper. How much it will be later under hashcat - well, we'll see.


(04-10-2016, 07:26 AM)epixoip Wrote: Anyway I think you're crazy, 50% is way off base from what will happen in reality. I'm still expecting max 20% and that's being optimistic.


I'm not crazy at all, just going with the facts. I might agree with you, if there would be just a new generation and no die shrink involved. But the fabric prcoess jumps from 28nm to 16nm. They skip the 20nm which would be one step down. But 16nm is a reduction by more than 3 times. They can use this for
- a die size shrink
- higher density aka more transistors
- higher clock rates
- lower power

Each step (28nm -> 20nm or 22nm -> 14nm) roughly doubles the above possibilities. Compare the new GP100 against the GM100:
same die size -> 1,0x
more transistors -> 1.9x
higher clockrate -> 1.33x
power efficiency -> 0.8x
overall advantage: 1 * 1,9 * 1.33 * 0.8 =~2x

At all the advantage of the new chip is roughly two times. That's fine, since they come from a highly optimized 28nm and will be more conservative on the new 16nm process. Most of that advantage goes into more transistors and they use it for their FP64 cores.

If the GP104 would see slightly smaller die size compared to the GM104 and would have the same power level, it will still have more transistors and higher clock rates. For example:

die size shrinked -> 1,15x
more transistors -> 1,33x
higher clock rate -> 1,33x
same power level -> 1,0x
overall advantage: 1,15 * 1,33 * 1,33 * 1 = ~2x

Following the GP100 it could have 10 SMX units with 2560 shaders and the same clock range (around 1500MHz boost). Surely not all of that performance necessarily ends up in hashcat. But most of it will.

This is not blind speculation since a shrink always has positive advantages. In one way or the other. It's a fact that I've studied in countless shrinks over the last 25 years. Trust me here, this is a huge step. And AMD does an even bigger one with Polaris (28nm to 14nm). NVidia is aware of that and they cannot afford to play around.

We will see who got closer to reality, your optimistic 20% or my pessimistic 40% Wink


(04-10-2016, 07:26 AM)epixoip Wrote: Also where are you seeing the GTX 1080 will be shipping at the end of July?

I've changed the "will be shipped" into "might be shipped" since I'm referencing to the newest rumours from last week that the announcement will be on Computex and shipment in July. The GTX980, GTX970 and the GTX980Ti shall be replaced. The latter because the new card is too close or even outperforms the 980Ti.