-
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.
Nvidia GTC16 Webcast Live Stream@5pm/9am PDT *Jen looks great in leather.*
- Thread starter tommybhoy
- Start date
Soldato
Everything is falling nicely into AMDs favour here.
Leading DirectX12, leading VR and now looking at dropping GPUs before the competition.
This could all come back to bite nvidia on the back side.
Leading DirectX12, leading VR and now looking at dropping GPUs before the competition.
This could all come back to bite nvidia on the back side.
So, do we have any timeframes for Polaris yet?
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
Pascal is a HPC(FP64) and deep learning(FP16) architecture. FP32(gaming) is the unfortunate casualty in improvements. They've moved to a architecture of shaders that can work together as FP64 or push 2 FP16 instructions through one shader. This has meant a very small overall increase in core count and FP32 throughput. But a really very very large increase in FP16/FP64 throughput.
It is perfectly possible that Nvidia manage to push through a bunch of optimisations for gaming using FP16 rather than 32. In the past there was a noticeable quality increase on many effects moving from 16 to 32bit, but some things can probably be done a little faster in FP16. Though how much and how the gpu delegates these tasks is yet to be seen.
It's still looking to me to be Q1 2017(wouldn't be shocked to see this slip a little) for professional GP100, 3-4 months later for Titan, 3-4 months later for a cheaper Ti version of it.
What that means for other cards... I wouldn't read too much into any link. Where as GK104 was what, around the time the professional GK110 was available, they were similar cards, gddr5, same process, same packaging. This time GP100 packaging/memory is so different I wouldn't link their time frames.
It's also possible that Nvidia could make a fairly different GP104, I don't automatically think they will but they could design a different consumer card with more shaders and less shader flexibility, IE optimised for FP32, more shaders in the same die size and more gaming performance.
I did try to tell String and Simon that It would be June. They both seemed to think that it was a dead cert that Nvidia were going to announce Pascal today and be able to order it.
I don't think too many of the others in thread were expecting a launch or much information, but, I do think everyone, including me, was hoping for a little teaser today about the consumer cards.
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
Jesus, the professional cards with GK110 came months before the consumer cards using GK110.
http://www.anandtech.com/show/6446/nvidia-launches-tesla-k20-k20x-gk110-arrives-at-last
They also weren't late, K20 only launched late in 2012. Nvidia went for a mid sized core first then a big core, it was intentional for the big core to come later. Not likely that much later but it really wasn't late. We only got information about GP100, I was quite clearly talking about when you would sensibly see a Titan based off it, not when the first consumer Pascal would appear.
Last edited:
I believe AMD were claiming (very) late summer but I've not kept upto speed.
Good show, disappointing not to see any consumer level info, but such is life, we can now wait for Computex to see what is said there.
On the subject of volume production now, please correct me if I'm wrong, but don't these chips actually take many months (six?) to bake at TSMC and other fabs. If so then they probably are in production now, by the time they are ready the HBM2 should be ready and then they can start putting it all together.
Also did he say when he first mentioned the Tesla P100 that it was several Pascal P100 dies stacked on top of each other? as I could have sworn that is what he said.
On the subject of volume production now, please correct me if I'm wrong, but don't these chips actually take many months (six?) to bake at TSMC and other fabs. If so then they probably are in production now, by the time they are ready the HBM2 should be ready and then they can start putting it all together.
Also did he say when he first mentioned the Tesla P100 that it was several Pascal P100 dies stacked on top of each other? as I could have sworn that is what he said.
AMD have working cards which have been demoed yet no info at all about pascal performance or architecture. I'm sure most companies release some bits of info well before any launch.
It's either because Nvidia don't have anything to reveal just yet or they are waiting for AMD to show their hand so that they can try and make sure they can beat them, a bit like the FuryX launch.
It's either because Nvidia don't have anything to reveal just yet or they are waiting for AMD to show their hand so that they can try and make sure they can beat them, a bit like the FuryX launch.
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
4-8 weeks in production usually, depends on the complexity and number of layers. 6 weeks is about average for a complex chip. I expect they might be in volume production... but volume can be anything. I'm pretty much expecting they will have utterly dire yields, ludicrously low availability and trickle out a couple of boxes. That they will only give their biggest customers availability in Q1/2017 really highlights when Nvidia believe yields/production will be viable for decent profits and volume.
No they weren't. Stop exaggerating
I don't think anyone expected an April launch! I also don't think they will release much/if any info much before launch either so as not to kill sales.
I they don't mention Pascal at Computex though, then yes a Q2 release looks very unlikely.
Soldato
- Joined
- 30 Nov 2011
- Posts
- 11,358
No, its 6-8 weeks generally to go from a blank wafer to packaged chips - a complete product launch to have enough completed cards to start selling takes at least 3 months and then you need to ship them so if using containers another 4-6 weeks
Thanks. 980ti it is!
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
There is a decent amount of information on the Pascal architecture from the link someone put up earlier. They've given shaders per SM, number of SMs and some fairly detailed numbers, number of registers, number of threads, clock speeds/performance for GP100. Assuming that Pascal is the same up and down the range in base design then that is a lot of Pascal architecture info. It's heavily FP64/FP16 performance increase focused and FP32 performance increase is pretty painfully low.
Nvidia will not give us any hint at perf/price much if at all before launch as that just kills current sales.
Even AMD's Polaris info was pretty sparse and hasn't given us much to go on.
I don't think Nvidia officially revealed anything about Maxwell 970/980 before launch. I think even the release date was up in the air and only rumoured until it happened
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
No they weren't. Stop exaggerating
I don't think anyone expected an April launch! I also don't think they will release much/if any info much before launch either so as not to kill sales.
I they don't mention Pascal at Computex though, then yes a Q2 release looks very unlikely.
A very large number of people on the Pascal thread were arguing the toss over stupid rumours/info from November about an April launch and they have been for months.
Soldato
I think it indicates that Nvidia are quite worried about the Intel compute cards,whch are slowly creeping into the supercomputer market.