Thoughts on Pascal?
#1
Another question from a newbie here.  Please go easy on me.

Despite just building a rig, I'm already saving for new cards.  I've been looking at the literature for the soon to be released Nvidia "Pascal" and realize how much I still have to learn.  The GP100 is touted as being the successor to the GM200 (TitanX), but the technical spec doesn't look to my uneducated eyes like we'll see a massive increase in performance for cracking work.  Yes, there are more transistors and more memory, but a lot of the talk is around the improvement for mixed precision (FP16 support) (cracking is more integer work, no?) and better performance per watt.  If Pascal looks to completely displace Maxwell, I'll go without new toys for  a bit longer and buy Titans once prices drop.  If not, I'm waiting for nothing and will just build another system with cards in the price/performance sweet spot.
(Yes, I know... The obvious answer is to just wait and see rather than bet on theoreticals.  However, I'll be kicking myself if I hold off for months waiting on Pascal and then Titan doesn't drop in price.  I'll also be kicking myself if I let myself get impatient, spend my money on a second GTX970 based system for immediate gratification and then see Titan drop by 50%.  First world problems....)

For the experts in the audience:  Please educate me.  What am I missing, if anything?  What is your take?

tl;dr - I'm saving for new cards.  Wait for Pascal?
#2
No actual technical specs have been released for Pascal, just various sound bites, so everything out there is just rumor and speculation and we really have no idea how much faster Pascal might be. You're correct, hash algorithms are exclusively integer math, so an increase in floating point performance doesn't really mean much if anything. Historically speaking, we should expect a 20-35% increase in performance. I would be surprised if it is more than 40% faster.

The Titan X will not drop in price when Pascal is released, just like the Titan and Titan Z have not dropped in price.

A release date hasn't even been announced for Pascal, so I don't see a point in waiting.
#3
(02-15-2016, 04:44 AM)epixoip Wrote: No actual technical specs have been released for Pascal, just various sound bites, so everything out there is just rumor and speculation and we really have no idea how much faster Pascal might be. You're correct, hash algorithms are exclusively integer math, so an increase in floating point performance doesn't really mean much if anything. Historically speaking, we should expect a 20-35% increase in performance. I would be surprised if it is more than 40% faster.

The Titan X will not drop in price when Pascal is released, just like the Titan and Titan Z have not dropped in price.

A release date hasn't even been announced for Pascal, so I don't see a point in waiting.


Thanks epixoip.  I appreciate the feedback (and the reality check).  You're absolutely right... the rumor mill is in full swing.   If the news is correct, 4 variants just shipped from TSMC to India for testing.  My guess is that Nvidia wants the results in time to support an April announcement at the GPU conf.  If past timelines hold true, we would expect first shipments 5-9 months from the announcement...given this is a significantly new design, my money would be on 9 months.  In other words, a long time to wait.

I hadn't pulled historical price data on cards... it just seemed a good assumption that there would be a significant drop in old tech price on the launch of a new technology; but boy, was I wrong... (yes, I know... do my homework next time!...lesson learned.)... it is shocking how little Nvidia pricing varies (the only slight dip was on the 980 when the 980Ti launched) (source: https://pcpartpicker.com/trends/price/vi...tx-titan-x).  I'm not sure I understand the reasons for the relative price stability, but the stats are the stats. I need to check used pricing in my local market (I'm not in the US). If my current cards hold up well, I may just buy lesser cards to keep me running and then resell them once there's data to support (or not support) moving to Pascal. 

One thing is for sure... I need more speed than I'm getting now and I need to figure out how to get it without breaking the bank!

Thanks again for the advice.
#4
Fab reduction from 28nm to 16nm is a big deal, let's just hope it's done optimally architecture-wise.
#5
its quiet the same as older generation of GPU

i have been following the prices of 770 for quiet some time now and his price didnt drop for the past 2 years  and he worth like an 970.

its not normal...
i was checking the price over amazon new egg and other online site.

pascal is going to be powerfull but my guess is , the cost of a card will be higher as titan x
hoping i am wrong about the price ;d
#6
I remember Jen-Hsun Huang saying, that Titans (Vanilla, Black, Z) were never considered gaming GPUs; rather something for hardware enthusiasts or sciency folk.
Since NV will be betting heavily on fp64 in the high-end models, my guess is that it's (top GeForce model at the time of official announcement) going to be as expensive as vanilla Titan was, $999.

As for the release, I'm still thinking Q2 2016.
Could potentially be rushed by Polaris.
Time will tell, eh?
#7
(02-15-2016, 04:44 AM)epixoip Wrote: I would be surprised if it is more than 40% faster.

With the change from 28nm to 16nm I would be surprised to see less than 40%. TDP/performance ratio will be plus 80%, that's almost a fact. So clocks and shaders will go up and/or TDP will go down. DP also takes away some of the advantage. And they might start with smaller die like 500mm² instead of the actual 600mm² to leave some headroom for later and ramp up yield.

The new features like HBM, higher DP rate and Mixed Precision are nothing useful for hashcat as far as I know. Good for games and HPC, no matter for hashcracking. We need higher clocks and shaders, simple as that. And maybe a nice enhancement in the command set.

From what rumors tell and the past has learned us I'd guess the first to come out might be the GP104 to replace the GTX980/970. Logically price and performance should be lower than the Titan X. Maybe we see a reduced GP104 similar to the GTX970 and then they put a Ti version out later when the Titan X is replaced by the new Titan GP100.

They also might start with the GP100 for the professional market, since Maxwell is useless here.


With AMD bringing its Polaris GPUs in 14nm and with GCN 4.0, NVidia can't play around though it's hard to say how much manpower AMD has left to create something completly new and do it bigtime from the beginning. If they don't mess it up, they might be very competitive again.

But clearly nothing will happen in this half of the year.
#8
Fresh rumours from the Italians.
According to them, the high-end GP100 aka "Pascal" Titan will be released Q4 2016 / Q1 2017, which sounds ridiculously late.
The Tesla variant will debut as promised a while ago, Q2 2016, in April.
#9
The newest rumors suggest a GTX1080 at the end of May with GP104, 8GB GDDR5-RAM and only one 8-pin, so max. power 225W.
#10
in 2 month?! holy moly ^.^

whats ur guess, will it be better than gtx 980 ti or perhaps titan x?