Nvidia have to provide a big pascal for compute purposes. I wouldn't read too much into TSMCs roadmap tbh.
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
** The Official Nvidia GeForce 'Pascal' Thread - for general gossip and discussions **
- Thread starter Boomstick777
- Start date
Nvidia have to provide a big pascal for compute purposes. I wouldn't read too much into TSMCs roadmap tbh.
Associate
- Joined
- 12 Jul 2015
- Posts
- 1,694
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
He means GP100, generally a big Pascal chip also while TSMC's road map is irrelevant, chip size has little to nothing to do with the process name or capability, if you can make a 75mm^2 chip or a 300mm^2 chip, you can make a 600mm^2 chip. The power/leakage or process name doesn't effect that, only max reticle size does(literally the absolute maximum they can create a mask to get accurate etchings on the chip). Yields are the only difference between realistically being able to make a huge chip and not being able to.
We really don't know what the yields of huge chips are at this point at TSMC or Samsung. Realistically the big chip yields have decreased process on process since 90nm, becoming increasingly difficult at every node.
There is a crossover point where yields are good enough to release a product and before that point while you can make them you might not have a viable product. Even talking about enterprise/professional prices, if you can only get 2 chips per wafer, even if you can charge 10k per chip and make a profit doing it, releasing a product with near zero availability still isn't a realistic idea.
The question with 16/10nm is which comes first, yields on 300mm^2 chips at 10nm or yields of 500+mm^2 16nm chips. Or not quite, if it's predicted today that you can make a 10nm medium sized chip within maybe 3-4 months of a 16nm big chip being doable, it becomes a needless cost to tape out a huge chip on 16nm as a lot of people will just wait for the 10nm chip anyway.
In terms of size, if medium Pascal has the same compute ratio as big Pascal, then there is no need for the bigger one. Presuming it has beefed up compute anyway then a medium chip will beat current compute offerings handily anyway.
- Joined
- 2 Jan 2012
- Posts
- 12,327
- Location
- UK.
Me 'Obviously' I imagine Gregster might enjoy messing with light weight Pascal, until the full fat arrives. Kaapstad will likely buy 4 of each Pascal GPU..
TSMC and Samsung (and Intel) are all trying to be first to ramp 10nm, so their little race may bring us a GP104 followed by a GV104 with no high-end in sight.
These 10nm offerings are meant to be rather lacking though, with very little density/power benefits over 16/14. So not sure NV would spend the hundreds of mil's to tape out on 10nm when they could hang in a little longer and maybe get a bigger chip made on a process variant more suitable for GPUs. Unless TSMC ditch 16 like 20? 10nm does have a new back-end, which means the chip's internal wiring will be smaller as well.
These 10nm offerings are meant to be rather lacking though, with very little density/power benefits over 16/14. So not sure NV would spend the hundreds of mil's to tape out on 10nm when they could hang in a little longer and maybe get a bigger chip made on a process variant more suitable for GPUs. Unless TSMC ditch 16 like 20? 10nm does have a new back-end, which means the chip's internal wiring will be smaller as well.
Last edited:
Intel have said they can make good old Silicon chips down to 7nm.
Seeing as they are the most successful/advanced foundry in the world, it would be wise to believe them
Intel have said they can make good old Silicon chips down to 7nm.
Seeing as they are the most successful/advanced foundry in the world, it would be wise to believe them
Perhaps but Intel just stopped the tick, tock approach as they ran into problems hitting 10nm which is now not going to happen for next gen. 14 and 16nm will stay with us a while longer.
Part of me is tempted to buy another Titan X and when Pascal releases, bury my head in the sand
I am genuinely concerned that performance will not be there on "lightweight" Pascal.
Associate
- Joined
- 3 Dec 2011
- Posts
- 477
- Location
- Shropshire
Apologies for going slightly off topic for a moment but they skipped 20nm (i think) to go to 14/16nm as they had difficulties getting 20nm working properly and now they are already thinking of going to 10nm, as 14/16 isnt very good either? arent they being a bit short sighted as they need to find another viable option than silicone, surley skipping each process shortens this time they have until we get to a point where we cant use silicone effectively enough and we will be stuck until they do find a viable option?? or have they already planned for this and already have another option waiting i know they are testing other things but from what i know (admittedly very limited knowledge about this) they dont have anything concrete yet.
No, they will be working with 14/16nm and expect it for a couple of years minimum.
- Joined
- 2 Jan 2012
- Posts
- 12,327
- Location
- UK.
Part of me is tempted to buy another Titan X and when Pascal releases, bury my head in the sand
Yeah I think we could be in for more of the same. I.e a round of cards to start with that use lower power / run cooler but don't offer a great leap over current stuff in pure performance until we've been thoroughly milked
. Then we'll see the full fat parts at crazy prices, eventually giving way to the real deal 980 Ti / Fury X replacements, in about a 8-12 months from now.
I hope this isn't the case and we see full fat cards come early on..
I think it's Intel who are rushing to 10nm for CPU. GPU wise, we should be on the the same node for 2 gens, or 2 years before transition to 10nm. With current pacing this could turn into 4-5 years lol
Associate
- Joined
- 12 Jul 2015
- Posts
- 1,694
I sincerely hope the GTX 980 (Ti) replacement will be here sooner than later, and we're not left with mid range cards for a long time to come. Is there any hint for one or the other? I really need a better card but going for a GTX 980 Ti now at their current rates is just silly.
Soldato
- Joined
- 11 Mar 2013
- Posts
- 5,511
I'm hoping for a 980 or 980Ti replacement by GTC or by summer at least? Nvidia hasn't said diddly squat on Pascal, but I very much remember feeling the same way before Maxwell came. Just be patient. It will release soon?
I think it's Intel who are rushing to 10nm for CPU. GPU wise, we should be on the the same node for 2 gens, or 2 years before transition to 10nm. With current pacing this could turn into 4-5 years lol
TSMC seems to be abandoning 16nm - its only road mapped through 2016 with medium power level optimisation and they are pouring investment into 10nm and 7nm instead of enhanced 16nm. They are claiming that they will start production for 10nm towards the end of the year in time for products to make it to market early 2017 but I doubt anyone is holding their breath on that one the only thing that might make a difference there is that it will make a big difference as to whether they hold onto Apple as a customer.
Associate
- Joined
- 3 Dec 2011
- Posts
- 477
- Location
- Shropshire
I think it's Intel who are rushing to 10nm for CPU. GPU wise, we should be on the the same node for 2 gens, or 2 years before transition to 10nm. With current pacing this could turn into 4-5 years lol
Ahh right i was getting mixed up yes that makes a lot more sense then
They are cutting corners big time now with 10nm / 7nm, they listed the specs we should expect for both and they were really underwhelming. They previously said they were going to standardize on 16, but it looks like they will just keep saying that each node and hope someone comes up with a new process they can license or something. Like IBM's fancy SOI stuff.
They are now releasing high end 980MX and 970MX laptops, so I would hold off regardless of Pascal.
http://wccftech.com/nvidia-maxwell-geforce-gtx-980mx-970mx/
Last edited: