Please excuse the ignorance, but there are 3584 Stream Processors of the FP32 variety correct? So does that mean that the FP64 CUDA Cores per SM are separate hardware-wise and therefore there are 1792 of those on chip? Also could FP64 cores, theoretically, be used for games either by having them crunch 32bit math or however else?
LOL, let me just say sorry Ryan. That is my younger brother asking that question since he finds it easier to use my login than to take the minuscule amount of time to sign-up for his own. It might explain some previous comments/questions too not sure how far back tho, not that I think its a big deal. Sorry, n its just funny on a personal level because he definitely does not talk like that or type like that from what I have seen haha. He's a pretty bright kid n Im glad he enjoys the workings of tech. But his last question does seem intriguing. I'm not sure if it would take some architecture changes (maybe?) but could FP64 cores be used in conjunction with FP32 cores to crunch the same type of dataset ("gaming" data). Is the gpu designed so that you are either using the 3584 FP32 cores OR the FP64 cores but not both at the same time? Could this just be as simple as being a power or thermal limit (running ALL the cores at the same time being too power hungry)?
Not sure what you are asking, but yes this is "Big Pascal" although they are currently only able to sell partially disabled versions because of the defect rate in the 16nm process.
Considering how functional the die is (over 90%) yields appear surprisingly good. They might have a fully enabled die to market in a year, if there is even any competition to push them to do so. Here's hoping AMD pulls off something good.
not much more because they are already pushing the boundaries on die size.. unless they are counting the HBM in those figures, or they are counting multiple layers.
Also, to add, GTX 1080 will probably be GP104. GP100 may end up in a GTX 1080 Ti, in a partially disabled form, and then probably also in a future Titan successor (possibly called Summit).
None of the Titans are crippled. It's just that GM200 has very poor FP64 performance by design. Please get your facts straight and not spread misinformation? k?
Titans are not crippled in FP64. The original Titan and Titan Black had full FP64, the Titan X was crippled in FP64 because Maxwell itself was crippled in FP64.
Yeah, that is pure speculation, on my part. Although it will probably not be until the Volta cards that we see the Summit name as the Summit supercomputer will be using Volta GPU's. The Titan supercomputer is built with Kepler GPU's so that's where the Titan name came from.
Looking at the configuration of this card, I'm wondering if a 1080Ti-type card based on it might not come out for a while, perhaps over a year. The 1080-type card will be based on a different chip (GP104?) and probably won't use HBM 2 at all (perhap it'll use GDDR5X?). I'm thinking HBM 2 might come to the GP104-type chip eventually, though, when NVIDIA replaces its Tesla M40 offering. I just wonder if the extra cost of HBM2 is worth it for desktop graphics at this point. But maybe it benefits mobile (notebooks) enough that they end up using it anyway?
It would be a terrible consumer card in many ways, it barely has any more cores then the current 980Ti because it's full of floating point hardware that games don't need.
It would result in roughly a 50%+ boost in performance beyond the Titan X. We saw roughly the same performance boost when we went from Titan Black (Kepler) to Titan X (Maxwell), so this is to expected. Nothing surprising.
A. This is a partially disabled GPU for the Tesla lineup, Nvidia's highest end "Brand" EG above their "Titan" cards. And this doesn't even have a release date or price. Don't expect a "1080i" or whatever until perhaps even next year.
B. 8gb of ram, that's guaranteed as you can't have 4 (games can use more, barely but it's there) and above 8 is a waste.
C. SINCE it's a "Tesla" only GPU for the moment, don't expect anything like affordable even at the high end ($1000+) until the end of this year at the earliest.
Speaking of margins, if consumer GPUs are vastly less profitable than HPC, why not just delay production of consumer Pascal and make cash-cow-HPC GPUs until AMD makes a move?
They won't bother with the 1/2 rate FP64 on the consumer dies, using 1/16 or 1/8 likely, that's a big difference. ECC memory support won't be present or will be disabled too.
Maybe it wasn't clear, but I'm not expecting the GTX1080 Ti to use the PM200 die since it's so huge. I expect a derivative die like PM202 will be made for it that's significantly smaller, has similar FP32 performance but cut down FP64 performance.
As far as I observed, SM architecture doesn't change from die to die, so I think FP64 ratio will be the same. Looking back and comparing Quadro and Tesla with consumer cards I can't see a FP64 ratio reduction.
Because it is huge HBM2 is expensive, I think GP100 will be a HPC exclusive until Q2/Q3'17 when it will released as '80 Ti.
I expect the lower sized GP104 to be released at Computex '16 @ 350-400 sqmm with GDDR5.
GTX1070 @ 5.5 TFLOPS and GTX80 @ 7 TFLOPS that replace the GM200.
"Looking back and comparing Quadro and Tesla with consumer cards I can't see a FP64 ratio reduction." Consumer cards based on GF100/110 and GK110 had reduced double precision performance. This was accomplished by either limiting clocks under double precision workloads through firmware and drivers, or fusing off FP64 cores during binning. GF100/110 is capable of 1/2 FP32, but the 580/570 were limited to 1/8. GK110 is capable of 1/3, but the 780/Ti were limited to 1/24.
True, an exception to the rule... sort of. Even in the case of GK210, it was still very much GK110. If I remember right they made some compute specific enhancements like larger caches and registers. But I think the point is that the top tier compute oriented GPU for a given architecture, like GF110 or GK110 have always been used in both consumer and pro cards. The 'big die' is never pro exclusive.
Yeah, I think we'll see a gp10p-based Titan, but the question is: will we see gp100 throughout the rest of the consumer lineup?
There are still rumors the existence of a gp102 part. I wonder if that could be a slow-dp gddr5x part that fills out most of consumer lineup above gp104.
The idea is that it'd be a much smaller die uses a less costly memory technology to get nearly the same gaming performance.
If that's the case it would be unprecedented. Nvidia hasn't done that before, and I always thought the reason for that was financial viability. It simply made more sense to invest in a single big die to serve multiple markets. Perhaps the 1/2 ratio might make it impractical as the halo gaming GPU.
Probably not, they would just not utilize the nvlink parts on the consumer version of GP100. Note that these still connect to the CPU via PCIe, at least on Xeon systems.
Are you sure? This is the first time I hear this. NVLink used for SLI seems overkill. Isn't NVLink physically extremely complex and thus PCIe can't be used, thus new proprietary slots, thus new motherboards, thus industry goes mad?
Those new 'flat' cards are ONLY going to be used in HPC servers, and even those still connect to the CPU via PCIe. Consumer video cards will be in the normal PCIe expansion card format like they always have been.
NVLink will not be used for SLI, it is 100+ pins just for ONE of the four lanes. The existing SLI connector is like 10 pins. NV will just use over-the-pcie like AMD did with Hawaii/XDMA.
Aren't they doing what you just said , the enteir presentation was about compute/AI/Machine learning/... . So far they have been able to use the same design for their gaming lineup. But these tesla design seams to branch away from conusmers cards.
1) AMD's Polaris has no high end chip (Polaris 11 is Pitcairn-level and Polaris 10 is Hawaii-level). Their high end HBM2-packing chips come with Vega in early 2017.
2) GP100 is an absolutely massive chip on an immature process. I doubt we'll see a fully enabled GP100 until at least mid-2017 (probably late 2017 to early 2018).
It seems it's a special math GPU for double precision computing only for Tesla cards, not for gamer cards, because it have too low CUDA cores here. If we just imagine Geforce 780Ti or 980Ti with such number of cores as GP100, it will be a small ~9-10 billion transistors chip, then raise core clock to 1.4GHz and we get the same single precision performance.
However 1.4GHz on such Teslas is great! So gamer boards will work at about 1.5GHz, which is about 30% higher then on 28nm.
I doubt such a drastic architecture change is going to occur. They practically just sliced SMs into half compared to Maxwell, doubling their count and giving them half the cores. That has certain advantages of offering more registers and supporting more warps in parallel.
Yeah, you are right. They'll probably simply cut down the number of FP64 cores in the GP104 die like they did with GK104. 64 cores/SM is definitely a welcome change for compute purposes...
Anyone know if they are using 3D vs 2.5D? I know Jen Hsun said 3D, but that would mean that the memory chips are on top of the GPU as opposed to next to it, like with Fiji. Putting memory chips on top of a very hot GPU seems like not such a great idea.
Well, mobile SoC's don;t technically use 3d stacking, because they don't use TSV's or microbumps, but they uses PoP, which is Package on Package, where it's just the memory package (die in packaging, not bare die like 3d stack would be) on top of the SoC package with regular BGA balls connecting them. They might possibly move to a true 3d stack in the future. Yes PoP is technically a "3d" stack, but the term 3d stack has a specific meaning in this case.
To all the people saying that consumer dies will have less DP cores, no NVLink, no HBM or many other things ripped out of its dies, I want to remind you that NVIDIA gave just one codename for its 2016 architecture release - Pascal.
So many architecture changes would ask for a different codename. Volta coming in 2018.
Did anyone suggest different dies would used? Everything discussed including FP64 ratios and NVLink can be accomplished through binning or some method other than manufacturing different
Dude, pretty much every past generation has had different ratios of FP64 cores in the lower Gx104, 106, etc chips. Fermi did, Kepler did, Maxwell has gimped FP64 through the whole stack so it's basically the only exception.
Could you please have all the FP16/32/64 in boost or base clocks. GP100 is all boost the Tesla K20 is at about 760Mhz while the K40 is at base and the M40 is at boost. It's a bit of a wreck. Thanks for the liveblog :)
Probably not. It's likely a GP104 or similar chip. You can't feasibly fit a GP100 in a <150W power envelope necessary for the MXM form factor to make sense.
No release date. It's not surprising. NV has contracts for super computers which they must fulfill. For those consumers price is an afterthought so such a huge die works if you get $10k per die/board. So first existing contracts are met, then the HPC market is served and then consumers might get access in form of a Titan-like card. I bet this will be earliest in Q2/17.
This also has nothing to do with AMD and everything with Intel Knights Landing. The later having the advantage of working as a main CPU too not just as add-in card.
Ryan Smith : If you can be so kind an take a look into this: Changed Features and Fixed Issues in the 5.1 Release of the Graphics Debugger
Dynamic shader editing support for Direct3D 12 has been added. (41101) Starting with the next release of NVIDIA Nsight, Fermi GPUs will no longer be supported. http://docs.nvidia.com/nsight-visual-studio-editio...
What I really would like to know is about consumer GPUs. When will the next generation of graphic cards with Pascal tech be available? How much will it cost? I wanna upgrade my PC for the HTC Vive (loved Toms Hardware review of the headset, by the way, did you guys read it?)
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
93 Comments
Back to Article
nathanddrews - Tuesday, April 5, 2016 - link
Does the GP100 confirm that this is the full Pascal die? 16GB of HBM2... good God! I wonder what the 1080/1080Ti will come with?Ryan Smith - Tuesday, April 5, 2016 - link
This is 36 of 40 SMXes for GP100 enabled.T1beriu - Tuesday, April 5, 2016 - link
*This is 56(!) of 60(!) SMXes for GP100 enabled.Ryan Smith - Tuesday, April 5, 2016 - link
Yes. I can't count today. Thanks!Samus - Wednesday, April 6, 2016 - link
Jesus those are some crazy specs for a not even fully enabled die. 16nm is going to make GPU's ridiculous.ImSpartacus - Wednesday, April 6, 2016 - link
But we only get it once then that's it.we're already at reticle with this one, so we won't even see bigger chips either.
The only trick is a maxwell-esque efficiency strategy, but I doubt they can get materially more efficient.
Low dp would buy time though. But that's it.
Let's hope we don't get stuck on the same process for five years again.
SunnyNW - Wednesday, April 6, 2016 - link
Could they not possibly remove the FP64 cuda cores from the SMs and use that space for more FP32 cores?Flunk - Wednesday, April 6, 2016 - link
Maybe, but I doubt they will.CrazyElf - Wednesday, April 6, 2016 - link
It would not be surprising if we got stuck at 16/20nm (it's really a hybrid process) for longer than we did at 28nm.Die shrinks are getting harder as we approach the limits of physics.
FunBunny2 - Wednesday, April 6, 2016 - link
-- Die shrinks are getting harder as we approach the limits of physics.why do so many ignore this? reality bites, yeah. but there's not a thing you can do about it.
SunnyNW - Wednesday, April 6, 2016 - link
Please excuse the ignorance, but there are 3584 Stream Processors of the FP32 variety correct? So does that mean that the FP64 CUDA Cores per SM are separate hardware-wise and therefore there are 1792 of those on chip? Also could FP64 cores, theoretically, be used for games either by having them crunch 32bit math or however else?SunnyNW - Wednesday, April 6, 2016 - link
LOL, let me just say sorry Ryan. That is my younger brother asking that question since he finds it easier to use my login than to take the minuscule amount of time to sign-up for his own. It might explain some previous comments/questions too not sure how far back tho, not that I think its a big deal. Sorry, n its just funny on a personal level because he definitely does not talk like that or type like that from what I have seen haha. He's a pretty bright kid n Im glad he enjoys the workings of tech.But his last question does seem intriguing. I'm not sure if it would take some architecture changes (maybe?) but could FP64 cores be used in conjunction with FP32 cores to crunch the same type of dataset ("gaming" data). Is the gpu designed so that you are either using the 3584 FP32 cores OR the FP64 cores but not both at the same time? Could this just be as simple as being a power or thermal limit (running ALL the cores at the same time being too power hungry)?
extide - Tuesday, April 5, 2016 - link
Not sure what you are asking, but yes this is "Big Pascal" although they are currently only able to sell partially disabled versions because of the defect rate in the 16nm process.Samus - Wednesday, April 6, 2016 - link
Considering how functional the die is (over 90%) yields appear surprisingly good. They might have a fully enabled die to market in a year, if there is even any competition to push them to do so. Here's hoping AMD pulls off something good.jasonelmore - Tuesday, April 5, 2016 - link
not much more because they are already pushing the boundaries on die size.. unless they are counting the HBM in those figures, or they are counting multiple layers.Krysto - Tuesday, April 5, 2016 - link
Didn't they promise up to 32GB HBM2 GPUs, though?T1beriu - Tuesday, April 5, 2016 - link
You read so many made-up "leaks" that you started to believe them. :)prtskg - Wednesday, April 6, 2016 - link
It'll come later.extide - Tuesday, April 5, 2016 - link
Also, to add, GTX 1080 will probably be GP104. GP100 may end up in a GTX 1080 Ti, in a partially disabled form, and then probably also in a future Titan successor (possibly called Summit).ddriver - Tuesday, April 5, 2016 - link
Fat chance. It will not be the same chip, even the titans are crippled in terms of FP64Pinn - Tuesday, April 5, 2016 - link
No, they don't do new chips to cripple FP64. They have other ways.ZeDestructor - Tuesday, April 5, 2016 - link
None of the Titans are crippled. It's just that GM200 has very poor FP64 performance by design. Please get your facts straight and not spread misinformation? k?maxxbot - Tuesday, April 5, 2016 - link
Titans are not crippled in FP64. The original Titan and Titan Black had full FP64, the Titan X was crippled in FP64 because Maxwell itself was crippled in FP64.ImSpartacus - Tuesday, April 5, 2016 - link
Summit? I haven't heard that rumor. Where does "Summit" come from?ZeDestructor - Tuesday, April 5, 2016 - link
More suppositino than rumour: Titan was named after the Titan supercomputer, and they may well call the next one Summit after the Summit supercomputerImSpartacus - Wednesday, April 6, 2016 - link
That's an interesting piece of speculation. I've been wondering where the Titan brand might go.extide - Wednesday, April 6, 2016 - link
Yeah, that is pure speculation, on my part. Although it will probably not be until the Volta cards that we see the Summit name as the Summit supercomputer will be using Volta GPU's. The Titan supercomputer is built with Kepler GPU's so that's where the Titan name came from.Yojimbo - Tuesday, April 5, 2016 - link
Looking at the configuration of this card, I'm wondering if a 1080Ti-type card based on it might not come out for a while, perhaps over a year. The 1080-type card will be based on a different chip (GP104?) and probably won't use HBM 2 at all (perhap it'll use GDDR5X?). I'm thinking HBM 2 might come to the GP104-type chip eventually, though, when NVIDIA replaces its Tesla M40 offering. I just wonder if the extra cost of HBM2 is worth it for desktop graphics at this point. But maybe it benefits mobile (notebooks) enough that they end up using it anyway?siriq - Tuesday, April 5, 2016 - link
Will be rev 2 from pascal.Dribble - Tuesday, April 5, 2016 - link
It would be a terrible consumer card in many ways, it barely has any more cores then the current 980Ti because it's full of floating point hardware that games don't need.madwolfa - Tuesday, April 5, 2016 - link
It's still 5x faster in FP16 and 2.5x faster in FP32.ImSpartacus - Wednesday, April 6, 2016 - link
It would result in roughly a 50%+ boost in performance beyond the Titan X. We saw roughly the same performance boost when we went from Titan Black (Kepler) to Titan X (Maxwell), so this is to expected. Nothing surprising.Frenetic Pony - Tuesday, April 5, 2016 - link
A. This is a partially disabled GPU for the Tesla lineup, Nvidia's highest end "Brand" EG above their "Titan" cards. And this doesn't even have a release date or price. Don't expect a "1080i" or whatever until perhaps even next year.B. 8gb of ram, that's guaranteed as you can't have 4 (games can use more, barely but it's there) and above 8 is a waste.
C. SINCE it's a "Tesla" only GPU for the moment, don't expect anything like affordable even at the high end ($1000+) until the end of this year at the earliest.
ImSpartacus - Wednesday, April 6, 2016 - link
How do you expect Nvidia to eventually do a "1080 ti" based on a new Titan?Historically, Titans have more vram (useful or not). So I'm wondering if the Titan keeps the egregious 16gb and 1080 ti gets 8gb.
Alternatively, Nvidia could remove a stack of hbm to get to 6gb, but I'm doubtful that will happen because of marketing reasons.
mckirkus - Tuesday, April 5, 2016 - link
Speaking of margins, if consumer GPUs are vastly less profitable than HPC, why not just delay production of consumer Pascal and make cash-cow-HPC GPUs until AMD makes a move?Southrncomfortjm - Tuesday, April 5, 2016 - link
Looks like the plan since they didn't say a word about consumer Pascal, although this is a developer-focused show, so maybe that isn't surprising.olde94 - Tuesday, April 5, 2016 - link
Normaly they don't talk about the consumer ;) This is a "future of the tech" show.nevirin - Tuesday, April 5, 2016 - link
Is the die any different between consumer and HPC models?frostyfiredude - Tuesday, April 5, 2016 - link
They won't bother with the 1/2 rate FP64 on the consumer dies, using 1/16 or 1/8 likely, that's a big difference. ECC memory support won't be present or will be disabled too.T1beriu - Tuesday, April 5, 2016 - link
When did any GPU manufacturer make different dies for servers (HPC) and consumers?frostyfiredude - Tuesday, April 5, 2016 - link
Maybe it wasn't clear, but I'm not expecting the GTX1080 Ti to use the PM200 die since it's so huge. I expect a derivative die like PM202 will be made for it that's significantly smaller, has similar FP32 performance but cut down FP64 performance.frostyfiredude - Tuesday, April 5, 2016 - link
Oops used Samsung's naming; GP100 and GP102 not PM200 and PM202T1beriu - Tuesday, April 5, 2016 - link
As far as I observed, SM architecture doesn't change from die to die, so I think FP64 ratio will be the same. Looking back and comparing Quadro and Tesla with consumer cards I can't see a FP64 ratio reduction.Because it is huge HBM2 is expensive, I think GP100 will be a HPC exclusive until Q2/Q3'17 when it will released as '80 Ti.
I expect the lower sized GP104 to be released at Computex '16 @ 350-400 sqmm with GDDR5.
GTX1070 @ 5.5 TFLOPS and GTX80 @ 7 TFLOPS that replace the GM200.
GM204 will live for another season.
dragonsqrrl - Tuesday, April 5, 2016 - link
"Looking back and comparing Quadro and Tesla with consumer cards I can't see a FP64 ratio reduction."Consumer cards based on GF100/110 and GK110 had reduced double precision performance. This was accomplished by either limiting clocks under double precision workloads through firmware and drivers, or fusing off FP64 cores during binning. GF100/110 is capable of 1/2 FP32, but the 580/570 were limited to 1/8. GK110 is capable of 1/3, but the 780/Ti were limited to 1/24.
T1beriu - Wednesday, April 6, 2016 - link
It seems I haven't done my research properly. Thanks for the explanation.extide - Thursday, April 7, 2016 - link
Actually the lower tier dies typically have less FP64 h/w anyways. GK110 had 1/3 rate FP64, but GK104 and down had less than that.ZeDestructor - Tuesday, April 5, 2016 - link
Actually, they'll use the same die with FP64 shut off most likely, like the 780 and 780Ti.ImSpartacus - Tuesday, April 5, 2016 - link
GK210 in the K80?Wasn't GK210 only ever used in the K80 and nothing else?
dragonsqrrl - Tuesday, April 5, 2016 - link
True, an exception to the rule... sort of. Even in the case of GK210, it was still very much GK110. If I remember right they made some compute specific enhancements like larger caches and registers. But I think the point is that the top tier compute oriented GPU for a given architecture, like GF110 or GK110 have always been used in both consumer and pro cards. The 'big die' is never pro exclusive.ImSpartacus - Wednesday, April 6, 2016 - link
Yeah, I think we'll see a gp10p-based Titan, but the question is: will we see gp100 throughout the rest of the consumer lineup?There are still rumors the existence of a gp102 part. I wonder if that could be a slow-dp gddr5x part that fills out most of consumer lineup above gp104.
The idea is that it'd be a much smaller die uses a less costly memory technology to get nearly the same gaming performance.
dragonsqrrl - Wednesday, April 6, 2016 - link
If that's the case it would be unprecedented. Nvidia hasn't done that before, and I always thought the reason for that was financial viability. It simply made more sense to invest in a single big die to serve multiple markets. Perhaps the 1/2 ratio might make it impractical as the halo gaming GPU.dragonsqrrl - Tuesday, April 5, 2016 - link
I wouldn't be surprised if it's as little as 1/32 if the current trend continues.extide - Tuesday, April 5, 2016 - link
Probably not, they would just not utilize the nvlink parts on the consumer version of GP100. Note that these still connect to the CPU via PCIe, at least on Xeon systems.HighTech4US - Tuesday, April 5, 2016 - link
NVlink will be used for SLI so yes NVlink will be on consumer Pascals.T1beriu - Wednesday, April 6, 2016 - link
Are you sure? This is the first time I hear this. NVLink used for SLI seems overkill. Isn't NVLink physically extremely complex and thus PCIe can't be used, thus new proprietary slots, thus new motherboards, thus industry goes mad?extide - Thursday, April 7, 2016 - link
Those new 'flat' cards are ONLY going to be used in HPC servers, and even those still connect to the CPU via PCIe. Consumer video cards will be in the normal PCIe expansion card format like they always have been.NVLink will not be used for SLI, it is 100+ pins just for ONE of the four lanes. The existing SLI connector is like 10 pins. NV will just use over-the-pcie like AMD did with Hawaii/XDMA.
dragonsqrrl - Tuesday, April 5, 2016 - link
Same die, binned for different purposes. Wouldn't be unusual for consumer cards to have most FP64 cores fused off during manufacturing.plopke - Tuesday, April 5, 2016 - link
Aren't they doing what you just said , the enteir presentation was about compute/AI/Machine learning/... . So far they have been able to use the same design for their gaming lineup. But these tesla design seams to branch away from conusmers cards.extide - Tuesday, April 5, 2016 - link
Uhh, that is EXACTLY what they are doing...ImSpartacus - Tuesday, April 5, 2016 - link
That's literally the plan.1) AMD's Polaris has no high end chip (Polaris 11 is Pitcairn-level and Polaris 10 is Hawaii-level). Their high end HBM2-packing chips come with Vega in early 2017.
2) GP100 is an absolutely massive chip on an immature process. I doubt we'll see a fully enabled GP100 until at least mid-2017 (probably late 2017 to early 2018).
xrror - Tuesday, April 5, 2016 - link
I think that's kinda what they did here ;)Lanskuat - Tuesday, April 5, 2016 - link
It seems it's a special math GPU for double precision computing only for Tesla cards, not for gamer cards, because it have too low CUDA cores here. If we just imagine Geforce 780Ti or 980Ti with such number of cores as GP100, it will be a small ~9-10 billion transistors chip, then raise core clock to 1.4GHz and we get the same single precision performance.However 1.4GHz on such Teslas is great! So gamer boards will work at about 1.5GHz, which is about 30% higher then on 28nm.
hammer256 - Tuesday, April 5, 2016 - link
Yeah, I would not be surprised if the consumer Pascal runs at 96 or 128 FP32 cores and 8 (4?) FP64 cores per SM. That would be rather interesting...nevcairiel - Tuesday, April 5, 2016 - link
I doubt such a drastic architecture change is going to occur. They practically just sliced SMs into half compared to Maxwell, doubling their count and giving them half the cores. That has certain advantages of offering more registers and supporting more warps in parallel.nevcairiel - Tuesday, April 5, 2016 - link
Infact the number of cores per SM has been gradually going down, from 192 in Kepler, to 128 in Maxwell and 64 now in Pascal.hammer256 - Tuesday, April 5, 2016 - link
Yeah, you are right. They'll probably simply cut down the number of FP64 cores in the GP104 die like they did with GK104. 64 cores/SM is definitely a welcome change for compute purposes...extide - Tuesday, April 5, 2016 - link
Anyone know if they are using 3D vs 2.5D? I know Jen Hsun said 3D, but that would mean that the memory chips are on top of the GPU as opposed to next to it, like with Fiji. Putting memory chips on top of a very hot GPU seems like not such a great idea.T1beriu - Tuesday, April 5, 2016 - link
He was talking about HBM.extide - Thursday, April 7, 2016 - link
Yeah, so am I...ImSpartacus - Tuesday, April 5, 2016 - link
Yeah, isn't that more of a mobile-friendly sort of thing?extide - Thursday, April 7, 2016 - link
Well, mobile SoC's don;t technically use 3d stacking, because they don't use TSV's or microbumps, but they uses PoP, which is Package on Package, where it's just the memory package (die in packaging, not bare die like 3d stack would be) on top of the SoC package with regular BGA balls connecting them. They might possibly move to a true 3d stack in the future. Yes PoP is technically a "3d" stack, but the term 3d stack has a specific meaning in this case.Fusion_GER - Tuesday, April 5, 2016 - link
First GPU using GPC and TPC (TPC was gone since the GTX480)Fusion_GER - Tuesday, April 5, 2016 - link
http://www.anandtech.com/show/2549/2T1beriu - Tuesday, April 5, 2016 - link
To all the people saying that consumer dies will have less DP cores, no NVLink, no HBM or many other things ripped out of its dies, I want to remind you that NVIDIA gave just one codename for its 2016 architecture release - Pascal.So many architecture changes would ask for a different codename. Volta coming in 2018.
dragonsqrrl - Tuesday, April 5, 2016 - link
Did anyone suggest different dies would used? Everything discussed including FP64 ratios and NVLink can be accomplished through binning or some method other than manufacturing differentvFunct - Tuesday, April 5, 2016 - link
FP ratios aren't changed by binning. That's intrinsic to the design.dragonsqrrl - Tuesday, April 5, 2016 - link
FP64 cores can be fused off. Clocks can also be limited to effect the ratio.Michael Bay - Thursday, April 7, 2016 - link
What would be the point of having NVLink on consumer videocard if no chipset provider supports it?T1beriu - Thursday, April 7, 2016 - link
Consumer dies always had extra bits disabled that were enabled in professional dies. NVLink will be disabled.extide - Thursday, April 7, 2016 - link
NVLink will not be used on the consumer cards, there is no point.extide - Thursday, April 7, 2016 - link
Dude, pretty much every past generation has had different ratios of FP64 cores in the lower Gx104, 106, etc chips. Fermi did, Kepler did, Maxwell has gimped FP64 through the whole stack so it's basically the only exception.extide - Thursday, April 7, 2016 - link
Plus there is no point to NVLink in consumer cards. HBM will be on the GP100 consumer card, but not any of the lower ones, like GP104 and down.testbug00 - Tuesday, April 5, 2016 - link
Could you please have all the FP16/32/64 in boost or base clocks. GP100 is all boost the Tesla K20 is at about 760Mhz while the K40 is at base and the M40 is at boost. It's a bit of a wreck. Thanks for the liveblog :)testbug00 - Tuesday, April 5, 2016 - link
That should be "P100" not "GP100".Pinn - Tuesday, April 5, 2016 - link
8-way? Voodoo's revenge?poohbear - Tuesday, April 5, 2016 - link
Ok excuse my ignorance, but is this TESLA GPU gonna power TESLA cars' onboard computer, or is it totally unrelated?Zertzable - Tuesday, April 5, 2016 - link
Unrelated :)ImSpartacus - Tuesday, April 5, 2016 - link
Probably not. It's likely a GP104 or similar chip. You can't feasibly fit a GP100 in a <150W power envelope necessary for the MXM form factor to make sense.beginner99 - Wednesday, April 6, 2016 - link
No release date. It's not surprising. NV has contracts for super computers which they must fulfill. For those consumers price is an afterthought so such a huge die works if you get $10k per die/board. So first existing contracts are met, then the HPC market is served and then consumers might get access in form of a Titan-like card. I bet this will be earliest in Q2/17.This also has nothing to do with AMD and everything with Intel Knights Landing. The later having the advantage of working as a main CPU too not just as add-in card.
ImSpartacus - Wednesday, April 6, 2016 - link
Yeah, that's how they can afford to go so big.siriq - Wednesday, April 6, 2016 - link
Ryan Smith : If you can be so kind an take a look into this: Changed Features and Fixed Issues in the 5.1 Release of the Graphics DebuggerDynamic shader editing support for Direct3D 12 has been added. (41101)
Starting with the next release of NVIDIA Nsight, Fermi GPUs will no longer be supported.
http://docs.nvidia.com/nsight-visual-studio-editio...
Badelhas - Wednesday, April 6, 2016 - link
What I really would like to know is about consumer GPUs. When will the next generation of graphic cards with Pascal tech be available? How much will it cost?I wanna upgrade my PC for the HTC Vive (loved Toms Hardware review of the headset, by the way, did you guys read it?)
Cheers
HollyDOL - Wednesday, April 6, 2016 - link
Guess we all would, but from what I have read nV was strictly tight-lipped considering consumer gpus...