Soldato
I'm at work so haven't caught up with any of today's events.
What's the TLDR?
What's the TLDR?
-
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
Here is the latest info on Pascal P100.
Taken from here.
https://devblogs.nvidia.com/parallelforall/inside-pascal/
Taken from here.
https://devblogs.nvidia.com/parallelforall/inside-pascal/
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 30,327
CUDA cores per SM are greatly reduced, but seems to follow a trend
Caporegime
Looks like a monster for gaming though, shame it's so far off.
Bear in mind this is not a consumer part this is a Tesla specced part, not even a Quadro part.
Of course having said that the underlying structure of the chip will be the same, but I would expect clock speeds to be very different. for example the GK110 Tesla part was clocked at 745MHz whereas the 780ti came in at 876MHz, the GM200 Tesla part 948MHz whereas the 980ti 1000MHz and we all know that custom AIB raise the base clocks as far as they can, near 1200MHz base clocks for 980ti's are not uncommon.
Maybe this will mean will see a 2 GHz card when these finally make it to market in consumer form.
Soldato
Take with buckets of salt:
https://www.youtube.com/watch?v=A7Lf1Wmt4f8
Supposedly,GTX980TI level performance according to the person who posted it on AT forums.
https://www.youtube.com/watch?v=A7Lf1Wmt4f8
Supposedly,GTX980TI level performance according to the person who posted it on AT forums.
Soldato
- Joined
- 11 Mar 2013
- Posts
- 5,469
The funny thing about this video is DX12 is not on? hmm??
Seems legit! :thumbsup:
On an unrelated note, I've got a bridge in Brooklyn I'm selling if you're interested?
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
It doesn't, it's huge and has 3 times the DP performance of GK110, but only just over twice the SP performance. 580 > Titan had basically 3 times the SP performance. It looks a heavily compute performance increased card and one of the smallest jumps in SP(mostly gaming) performance in recent history.
K20 was 225W while Titan was 250W and came significantly later. This time Nvidia are rating their professional card at 300W.... meaning there is pretty much no reason to expect these to have a higher clock speed at all and no, no way will you see a 2Ghz part. This card is already running hot at the current clocks and there is little to no reason to expect higher clocks. In fact they could well go down to 250W for a consumer card.
Permabanned
- Joined
- 8 Dec 2015
- Posts
- 1,485
Pascal what a load of rubbish then haha. Looks like ill be keeping my Titans alittle longer muahahaha 
Man of Honour
Here is the latest info on Pascal P100.
Taken from here.
https://devblogs.nvidia.com/parallelforall/inside-pascal/
Nice
Roll on 2017
The Titan version is going to be twice as fast as a TitanX judging by the above specs.
Good point I didn't notice that. At 300W this seems a bit disappointing, I'd even go as far to think that something doesn't seem to add up.
Man of Honour
Whack a waterblock on it, job done.
Try overclocking a 5960X, 300W is nothing compared to that.
Almost forgot, the midrange cards we will see this year will be half the size and use about 150W so it is not all bad news.
Soldato
Very surprised that they jumped to 600MM^2 so fast. But it appears that they needed to since they have used dedicated DP units again.
I can see vega being around 300 - 350mm^2 with more SP and DP performance than GP100 if it has 4096 SM's like the leaks mention. The size of the core would not be too much bigger when adding DP support back considering the way AMD perform DP compared to Nvidiaa.
I can also see GP100 not coming to desktop but instead them giving something between GP104 and GpP100. Like GP102 or something which has no DP units in hardware and has SP SM's instead. A part made for SP performance that would surpass GP100 in SP performance.
GP100 would not be a very good gaming part on the desktop for what it is. At the size of die it is and with just under 1/3 of its area dedicated to DP hardware. it would not be as good for gaming as the Titan X was.
IF it does come to desktop it will be a very expensive Titan card, more so than the current and past titans. Unless they have some very good yields already on FF+.
I can see vega being around 300 - 350mm^2 with more SP and DP performance than GP100 if it has 4096 SM's like the leaks mention. The size of the core would not be too much bigger when adding DP support back considering the way AMD perform DP compared to Nvidiaa.
I can also see GP100 not coming to desktop but instead them giving something between GP104 and GpP100. Like GP102 or something which has no DP units in hardware and has SP SM's instead. A part made for SP performance that would surpass GP100 in SP performance.
GP100 would not be a very good gaming part on the desktop for what it is. At the size of die it is and with just under 1/3 of its area dedicated to DP hardware. it would not be as good for gaming as the Titan X was.
IF it does come to desktop it will be a very expensive Titan card, more so than the current and past titans. Unless they have some very good yields already on FF+.
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
Honestly could be as simple as needing that much to give it a none awful FP32 performance increase. IE if this was a 225W card with 1Ghz clocks, TF performance would be what under 10TF, not that far ahead of a last gen Fury X... it's still notgreat but I can understand that making a higher power card to extend the gaming/FP32 performance gap between their generations is fairly important to not be seen as stupid cards.
There is also the fact it's a HBM2 card. Simpler layout, smaller PCB, for a similar sized card they can afford to stick a lot more heatsink within the same area. They might simply be more comfortable cooling this card at 300W than a normal card at 225W.
Maybe it's a bit of both, a lot of the numbers are weird from what we're getting. Drive PX2 numbers are just freaking weird, making absolutely no sense at all. GP100, you would certainly expect (based on existing trends) for a smaller TDP card and lower clocks for more efficient power.
Could be down to the process and where it's most efficient. Say that at 1v the card can't break 1Ghz and uses 250W, but with only 1.1v it can work fine up to 1350Mhz and only uses 300W to do so, so 20% more power for 35% more performance. It could just be that those clocks and 300W is where it's best performance/watt happens to be at.
Could be all of them, or none of them.
Soldato
They are using GP106 in the Drive PX2. So each card had around 3-4Tflops Sp with the rest coming from the Tegra cores.
I would think Nvidia would come up with something more imaginative than 1070
No point in 0. Call it 170 and it makes more sense.
X70/X80/X80TI/Titan XX!?
would be nicer,
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
Anandtech have an article up... another very odd number now. The memory is clocked at 1.4Gbps rather than 2Gbps. HBM1 being 1Gbps. When I saw the 720GB/s bandwidth with 4 stacks I initially figured ECC memory might account for a large memory bandwidth reduction though it seemed excessive or that Nvidia were randomly listing effective bandwidth(all gpus/cpus have significantly lower effective than max theoretical bandwidth).
But with HBM2's main thing being 2Ghz instead of 1Ghz clocks... 1.4ghz is odd. More fine grained memory allowing them to simply pick and choose the exact amount of memory bandwidth they need and push more power to the cores? Samsung HBM2 having issues hitting 2Ghz clock speeds, 8 hi stacks proving more difficult than they hoped with temps at 2Ghz or maybe memory controller unstable at 2Ghz on Pascal.
I'm surprised they wouldn't just advertise 1TB/s and have the memory downclock for efficiency reasons because, frankly 1TB/s looks really impressive and iirc Nvidia have been calling Pascal a up to 1TB/s architecture for the past couple of years. It does feel more like a forced decision rather than active choice.
But with HBM2's main thing being 2Ghz instead of 1Ghz clocks... 1.4ghz is odd. More fine grained memory allowing them to simply pick and choose the exact amount of memory bandwidth they need and push more power to the cores? Samsung HBM2 having issues hitting 2Ghz clock speeds, 8 hi stacks proving more difficult than they hoped with temps at 2Ghz or maybe memory controller unstable at 2Ghz on Pascal.
I'm surprised they wouldn't just advertise 1TB/s and have the memory downclock for efficiency reasons because, frankly 1TB/s looks really impressive and iirc Nvidia have been calling Pascal a up to 1TB/s architecture for the past couple of years. It does feel more like a forced decision rather than active choice.