-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
NVIDIA Pascal GPU to feature 17B transistors and 32GB HBM2, coming Q2 2016 (earliest)
- Thread starter Zombie James
- Start date
Goldenmoney
Member
Sounds powerful, but it also sounds expensive as fuck.
Speculation is fun.
I had to wonder if perhaps Pascal is mostly a shrunk Maxwell whose introduction was a result of the 16nmFF delay, and they wanted to play it safe by using a mature design first ( a bit like Intel's tick-tock*) while making the aforementioned changes/additions (NVLink, HBM, mixed precision) instead of jumping straight to a more significant rejigging on an expensive node. **
I know you mentioned kepler <-> maxwell as not being a major architecture shift, but I consider it to be fairly changed up in a number of ways. IMHO, architecture refresh example would be 750Ti vs 9xx series( or even the GCN series ಠ_ಠ ).
*always hated that the "tick" was the shrunk version of the previous "tock". :|
** i.e. Take tegra x1 GPU, which is Max-based, make it super phat & add/change up a couple things, and shift from 20nm to 16FF+... ¯_(ツ
_/¯Kinda feel like its a long shot that NVLINK will be doable on consumer.
Needs motherboard and CPU to be compliant. Is Nvidia going to go back into the chipset business?I cant see Intel playing nice with them because they are competing companies in the gpgpu space with the Xeon Phi line.
More hoping that they get can get improved multi-gpu scaling as it looks like we are going to need it for Hi-Resolution VR. Seems like that is one of the promises that the silicon interposer offers for multi-die, single card configurations.
I'm just hoping we have a way to do multi-gpu and PCI-E SSD's without compromising bandwidth on the CPU's controller. That might be solvable with Skylake E.
I still feel like my 780TI's run just about everything at 2560x1600. Star Citizen runs flawlessly and that's about the peak of usefully pretty.
Just need to last until Skylake E and the first HBM Titan
Needs motherboard and CPU to be compliant. Is Nvidia going to go back into the chipset business?I cant see Intel playing nice with them because they are competing companies in the gpgpu space with the Xeon Phi line.
More hoping that they get can get improved multi-gpu scaling as it looks like we are going to need it for Hi-Resolution VR. Seems like that is one of the promises that the silicon interposer offers for multi-die, single card configurations.
I'm just hoping we have a way to do multi-gpu and PCI-E SSD's without compromising bandwidth on the CPU's controller. That might be solvable with Skylake E.
I still feel like my 780TI's run just about everything at 2560x1600. Star Citizen runs flawlessly and that's about the peak of usefully pretty.
Just need to last until Skylake E and the first HBM Titan
I'm still maintaining that NVLink will only be used for servers.
Everything they have said about Nvlink, such as the requirement for special physical connectors makes me think that the protocol requires even stronger signal integrity requirements than PCIE. The physical protocol seems to be be using very high clockspeeds otherwise they wouldn't be worried about the Impedance of connectors used in PCIE. Those preliminary connectors they showed look very much like shielded interfaces used in RF.
This essentially means you cannot just shove in something like a SLI connector unless it is very carefully designed. I could definitely see it being used on Multi Gpu cards though.
Everything they have said about Nvlink, such as the requirement for special physical connectors makes me think that the protocol requires even stronger signal integrity requirements than PCIE. The physical protocol seems to be be using very high clockspeeds otherwise they wouldn't be worried about the Impedance of connectors used in PCIE. Those preliminary connectors they showed look very much like shielded interfaces used in RF.
This essentially means you cannot just shove in something like a SLI connector unless it is very carefully designed. I could definitely see it being used on Multi Gpu cards though.
Why would they ever cram 17 billion transistors on a chip even if they could? This article doesn't make a lot of sense.
It's a much, much better business strategy to make incremental improvements (maybe 10%/20% improvements tick over tock) and milk it dry. You can go from 8 bil to 10 to 12 to 14 to 18, and you have 4 generations of products to sell rather than 1. Also prevents the consumer from expecting too much because you gave them too good a deal for Pascal.
It's a much, much better business strategy to make incremental improvements (maybe 10%/20% improvements tick over tock) and milk it dry. You can go from 8 bil to 10 to 12 to 14 to 18, and you have 4 generations of products to sell rather than 1. Also prevents the consumer from expecting too much because you gave them too good a deal for Pascal.
They will safely do that once they finished curbstomping AMD into bankruptcy.
LiquidMetal14
hide your water-based mammals
Good enough for me. I will have a stash tucked away for Vive and maybe 2 GPU's.
Am I wrong in suggesting there will be a range of cards on market juts like other NVIDIA series? So, you would be able to purchase a 750/760/770 equivalent card that would be sub $300 (770 equivalent more akin to 300-400), and then 780/790 equivalent will be more expensive.
They may start doing the refresheroo for lower-end cards.
The Pascal equivalent of GTX 680 (GK104) and GTX 980 (GM204) certainly won't need 17 Billion transistors. That number is obviously for the flagship GPU, 'Big Pascal'. unofficially referred to as GP100. It's totally reasonable to expect a range of Pascal GPU sizes, serving a full range of cards, as Nvidia always has. That's a given.
Nvidia may have taped out Big Pascal first but I'd expect them to launch cards using a mid-sized Pascal GPU first, as they did with both Kepler (680, march 2012) and second gen Maxwell (980, 970, fall 2014). Big Pascal will probably be reserved for Quadro, the next Titan and the successor to 980 Ti. However, just because Nvidia has done similar things for 2 gens in row, doesn't mean they're locked to do things the same a 3rd time around.
Nvidia may have taped out Big Pascal first but I'd expect them to launch cards using a mid-sized Pascal GPU first, as they did with both Kepler (680, march 2012) and second gen Maxwell (980, 970, fall 2014). Big Pascal will probably be reserved for Quadro, the next Titan and the successor to 980 Ti. However, just because Nvidia has done similar things for 2 gens in row, doesn't mean they're locked to do things the same a 3rd time around.
They almost have to. If they plan on launching big Pascal with HBM2, they're going to have to wait for those to be available, which is Q2 at the earliest.They then have to integrate the HBM dies into the interposer etc. You're basically looking at HBM2 equipped products to appear Q3.
Damerman
Member
This is the best route... Hopefully by then FF XV will be out for PC.
This sounds like its going to be a huge leap. It definitely sounds like they are pushing for a 4k standard. This is good because i settled for a 1080p monitor and a 970, skipping out on 1440p on purpose because i only thought a jump directly to 4k would be worth it.
Damerman
Member
Whose fucking side are you on?
Legato.Reborn-
Member
That man is a capitalist.
dr_rus
Member
Too bad that 16FF+ is actually 20nm physically and from a size savings point it's just the next step from 28nm.
Everyone here seen to also constantly forget about prices of the new nodes - even with a decent size savings it's rather unlikely that a chip of the size of GM200 will cost the same to produce on 16FF+ as on 28HP. I would say that it will cost several times more during the first year or so. Thus a smaller chip at a higher price is the most possible outcome.
There was a reason why we got 680 and 7970 first at their respective price points. This reason is even more pronounced with 16FF+. Everyone should keep their expectations in check.
cpp_is_king
Member
It's literally insane? Sounds scary, are there so many transistors it's going to set my house on fire or something? So the designer was insane to think it was a good idea?
neogaffer1
Banned
Has it been said anywhere what 10nm FF+ might be, physically: 20nm, 16nm, 14nm or a true 10nm ?
I know there's TSMC, Samsung and GF and they might use different method, but...
Well, physically*, it'll be a proper transition. The issue with planar 20nm was that the power & performance characteristics didn't scale as expected.
* node names have been marketing speak at least since 40/45nm anyway
So, when is this expected to come out now? I'm waiting on a 980 TI order from Amazon because it's not in stock currently, but the thought of Pascal on the horizon really dampens the mood.
Should I just use the 980 TI till pascal comes out then sell it?
Will the 980 TI hold it's value till then?
Should I just use the 980 TI till pascal comes out then sell it?
Will the 980 TI hold it's value till then?
dr_rus
Member
No idea. With how often things are changing around TSMCs nodes I doubt that anyone has. Theoretically 10nm should be a "proper" 10nm as it retains the same FINFET tech used to create "16nm" out of their 20nm.
It will of course go down a bit, but if you need to performance now you should get it. It's still ways off until we get these new cards.
AMD readies three new GPUs: Greenland, Baffin and Ellesmere
http://www.kitguru.net/components/g...gpus-for-2016-greenland-baffin-and-ellesmere/
http://www.kitguru.net/components/g...gpus-for-2016-greenland-baffin-and-ellesmere/
I'm buying new GPU in 2016, and most likely AMD will give me best price/performance.
Bring it read team. Make me happy.
GatticusFinch
Member
Don't hold your breath. Prepare to be disappointed as usual with AMD. I waited for Fury X and now I'm a 980 Ti owner.
I was all set to switch back to team green (I was due for a flip-flop anyway) but I guess I'll wait and see what AMD is doing. It's for VR so I have a little time and my 290x is still kicking.
Chucke1992
Banned
This basically proves that Pascal is Maxwell on 16nm + HBM2 + some compute stuff that was cut from Maxwell due to 28nm.
Wow. Maxwell was great and transforming into Pascal will make it even greater!
DX12 context switch great?
This basically proves that Pascal is Maxwell on 16nm + HBM2 + some compute stuff that was cut from Maxwell due to 28nm.
So NVLINK is coming with Pascal as well. I assume if it's up there, it will appear in some high end motherboards.
That's when I'm building my 4k/60 PC.
EDIT: Actually, do they use the same chip names for their pro cards?
icecold1983
Member
if maxwell does turn out to be a poor performer in dx12, pascal unfortunately wont change it
This basically proves that Pascal is Maxwell on 16nm + HBM2 + some compute stuff that was cut from Maxwell due to 28nm.
Thus Volta is the real deal.
This basically proves that Pascal is Maxwell on 16nm + HBM2 + some compute stuff that was cut from Maxwell due to 28nm.
Will it be enough to be competitive with AMD's 400 series ?
No idea if they chose Samsung 14nm or TSMC FF+. Also whether maintaining GCN is holding them back or not.
Chucke1992
Banned
Why should I care? End user does not know and care regarding how many calls is done inside GPU or whatever.
Whatever provides me with best experience - that will be my choice.
TSMC will manufacture their chips, not Samsung. Obviously we don't know much at all about the new rounds of GPUs but if async compute is going to be as big as some people claim then it's a bit short-sighted on Nvidia's part not to include something akin to AMD's ACEs. Or perhaps async compute can still be fully supported with their Pascal setup, no idea for now.
This is what happens when a node is skipped. Not that surprising really to be honest but I guess the date assumes no respins or other issues with the die, which on a new node using new tech feels a bit optimistic but still doable.
Being stuck with 28nm has really stagnated GPU performance so I am glad we they are finally moving on.
Being stuck with 28nm has really stagnated GPU performance so I am glad we they are finally moving on.
Also optimistic now that Samsung has a competing foundry and smartphone uptake slowed. Two years ago it felt like TSMC was the only viable foundry yet preferred mobile business.
Finished? You mean that's not the current state of play?
Maxwell was a pretty significant performance enhancement considering it was still on 28nm. By all reports there's nothing to suggest Pascal is going to be revolutionary; just a shrunken down Maxwell.
Yeah, but presumably, Volta won't be out until like...2018.
You are obviously not following the most recent developments.
As good as Maxwell is the improvements are nothing compared to what a die shrink would have provided. Instead of a 7Tflop Titan X / 980Ti a 20nm Maxwell would likely have been a 10-11 Tflop Titan X / 980Ti which is a significant improvement. Along with the fact the increased transistor density would have enabled more ROPS/TMUs and other parts to have had their numbers boosted as well.