NVIDIA Pascal Mythbusting

[Glenwing]

It’s high time to start shutting down some of these myths. There have been too many poorly written and misleading articles published on various tech “news” websites, generating hype out of nothing. Pascal is entirely focused on HPC (high-performance compute, a.k.a. supercomputers and servers) and NVIDIA hasn’t said so much as a word about gaming. But at every turn people keep trying as hard as they can to interpret every statement as "amazing for gamers!", and every time NVIDIA specifically says “compute performance” the words somehow turn into “gaming performance” in people’s minds, leading to a lot of false impressions and expectations. I'm not saying Pascal won't have amazing gaming performance. I'm saying we have no information about Pascal's gaming performance so far. Pascal might be great for gaming, it might be rubbish. NVIDIA has said nothing on the topic, basically none of what they've said so far is really applicable to gaming.

But anyway... let’s get busting!

"10× the Performance of Maxwell GM200!"

Spoiler

This is nothing but an empty myth, spread by tech "news" columnists who are either professional writers that are somehow woefully incompetent when it comes to reading comprehension, or writers that will knowingly use misleading titles and spread false rumors if it gets their website more attention and ad revenue (I'll let you decide which is more likely). In any case, a lot of people are now under the impression that Pascal's top-end GPU will have ten times the general computational power (and so 10× the gaming performance) of Maxwell GM200, the GPU used in the GeForce GTX Titan X. Is this what NVIDIA said? Well, let's just read NVIDIA's original blog post, something the tech "news" columnists apparently didn't do:

Note that this does NOT say "Pascal GPU to provide 10× speedup in everything". That's a hint right there, because if that were the case, you know NVIDIA would making that the headline instead. I'll admit though that the title is maybe a little ambiguous and could be misinterpreted by someone carelessly skimming, but hopefully any professional writer reporting on this would realize that there's an actual article that goes with that title. Reading past the title and into the article itself (take notes, tech "news" writers), we can get some more specific information.

NVIDIA's original claim was that Pascal would have ten times the overall throughput in deep learning compute workloads utilizing mixed precision compared to Maxwell, which sounds impressive until you’re made aware that Maxwell is (by design) completely horrible for many compute tasks to begin with, in some cases actually many times slower than Kepler. In double-precision compute workloads for example, Maxwell GM200’s performance in FLOPs is about 1/32 of its single-precision (normal workload) performance, while Kepler GK110 puts out 1/3 of its normal performance. If we do a little math with these numbers, we find that even Kepler GK110 is already over seven times faster than Maxwell GM200 in double precision. So ten times the performance of Maxwell in compute tasks is not necessarily as earth-shattering as it sounds; it depends on what type of computations are being discussed.

The key thing to note is that NVIDIA didn't say the GPU would be 10× as powerful. They said it would have 10× the throughput in a specific type of workload. Let's see if we can analyze what that might mean. "Throughput" is an important word, as it's not quite synonymous with "computational power". A graphics card's throughput (overall work being done) isn't always solely determined by the computational power of the GPU itself. The GPU is continuously retrieving and storing data related to its computations, and the GPU can stall if the data it needs for its calculations isn't being provided quickly enough. In that case, the full power of the GPU isn't being utilized. If you have a situation like this, increasing the speed of the GPU's data transfer subsystems will increase the GPU's overall throughput without actually increasing the amount of calculations that the GPU can do. This improvement is workload-specific though, because different types of workloads use the GPU in different ways. Some might require the GPU to do a lot of calculations with a relatively small amount of data, and some might need the GPU to do some lighter calculations but with a much larger data set. So different workloads have different balances of how taxing they are on GPU power vs data transfer. Improving data transfer rates will only help workloads that were being limited by that, and vice versa.

Now, let's put this in context. Gaming workloads are heavy on GPU computational power. The amount of data transported across the PCI Express bus (the interlink between the CPU and GPU) is very small, even "low" configurations like PCI Express 3.0 ×4 rarely affect performance at all. The amount of data transported across the GPU memory bus (the interlink between the GPU and the VRAM) is also relatively small, and isn't a major factor in performance. On the other hand, compute workloads are much heavier on memory bandwidth, memory capacity, and communication with the CPU compared to gaming. PCI Express 3.0 ×16 was a bottleneck the moment it was available. Memory bandwidth is woefully inadequate. Improving these aspects of the GPU would increase performance in compute workloads significantly while gaming performance would see very little, if any, improvement. Of course, there's no reason improvements couldn't be made to both the memory/system interconnects and the GPU core at the same time, but given a statement like "10× throughput in deep learning workloads" we have no way of knowing how much of that comes from improvements to the memory/system interconnects and how much comes from increased GPU power. So it really doesn't tell us much at all about the computational power of the Pascal GPU, because we have no way of knowing how much of that "10×" actually comes from increased GPU power itself.

...Unless of course NVIDIA breaks down exactly what goes into that 10× figure for us:
 

Uh-oh, all the big numbers on there are representing improvements to memory bandwidth and the system interconnect. These things won't significantly affect gaming performance.

The only thing related to computational power on that slide is mixed precision performance. What we know so far is that Pascal's support for mixed precision compute mode means that it can do FP16 calculations at double the rate compared to FP32 (standard calculations), which is all well and good for deep learning but not particularly useful for games.
 

Quote

Mixed-Precision Computing – for Greater Accuracy
Mixed-precision computing enables Pascal architecture-based GPUs to compute at 16-bit floating point accuracy at twice the rate of 32-bit floating point accuracy. Increased floating point performance particularly benefits classification and convolution – two key activities in deep learning – while achieving needed accuracy.

http://blogs.nvidia.com/blog/2015/03/17/pascal/

NVIDIA's CEO stated (in the video below) that Pascal would have 4× the FP16 performance of Maxwell, 2× of which is from the addition of mixed precision mode, and 2× "from the Pascal architecture itself". So at best we can predict that the highest-end Pascal chip might have up to 2× the performance of Maxwell GM200 in standard FP32 calculations, if we're feeling very optimistic. Obviously that will still make for an incredibly powerful GPU, don't get me wrong, I'm just saying... it's not 10×. Not even close.
 

 

"NVLink!"

Spoiler

NVLink is an interface for connecting multiple GPUs and/or the CPU. NVLink can be used in place of PCI Express, with about 5–12× the bandwidth of PCI Express 3.0 ×16 depending on implementation. The full implementation of NVLink, where it entirely replaces PCIe, is only for supercomputers and servers, where they will require purpose-built motherboards and CPUs with NVLink controllers instead of PCIe controllers. So far the only CPUs we know of that are looking to fully implement NVLink in the near future are custom IBM POWER processors to be used as part of the Summit supercomputer being jointly designed and built by IBM and NVIDIA. The full NVLink implementation will not be used in the general desktop market, normal GeForce and Quadro cards will continue using PCI Express.

In compute applications, like what supercomputers do, the connection between the GPU and the CPU (currently PCI Express) is heavily used and can be a significant bottleneck; removing that by replacing PCIe with NVLink will increase the overall throughput by quite a lot in those applications.

However, games do not need to use that connection very heavily, and so the communication bus between the GPU and CPU sees much less traffic in gaming compared to compute. PCI Express is nowhere near being a bottleneck for gaming. So even if full NVLink appeared on the desktop as a GPU-CPU interconnect, it wouldn’t provide any significant benefit to gaming performance compared to PCI Express.

NVLink can also be used purely as an interconnect between multiple GPUs, without an NVLink connection from the GPUs to the CPU. So basically, it might be implemented as a replacement for the SLI bridge. If this happens, it is not clear yet whether it would be done with a new external bridge, or whether it would be integrated into a proprietary connector or interface on the motherboard itself, though I think that NVIDIA is unlikely to implement it that way unless they can make the cards still compatible with PCI Express slots.

But, it is not even clear whether NVLink will appear at all on the desktop market, as NVIDIA has not mentioned anything about NVLink in the context of multi-GPU gaming or consumer GPUs, they’ve only talked about how useful it will be for supercomputers so far.
 

Quote

Since Titan, a trend has emerged toward heterogeneous node configurations with larger ratios of GPU accelerators per CPU socket, with two or more GPUs per CPU becoming common as developers increasingly expose and leverage the available parallelism in their applications. Although each of the new DoE systems is unique, they share the same fundamental multi-GPU node architecture.

While multi-GPU applications provide a vehicle for scaling single node performance, they can be constrained by interconnect performance between the GPUs. Developers must overlap data transfers with computation or carefully orchestrate GPU accesses over PCIe interconnect to maximize performance. However, as GPUs get faster and GPU-to-CPU ratios climb, a higher performance node integration interconnect is warranted. Enter NVLink.

NVLink is an energy-efficient, high-bandwidth path between the GPU and the CPU at data rates of at least 80 gigabytes per second, or at least 5 times that of the current PCIe Gen3 x16, delivering faster application performance. NVLink is the node integration interconnect for both the Summit and Sierra pre-exascale supercomputers commissioned by the U.S. Department of Energy, enabling NVIDIA GPUs and CPUs such as IBM POWER to access each other’s memory quickly and seamlessly. NVLink will first be available with the next-generation NVIDIA Pascal GPU in 2016.

In addition to speeding CPU-to-GPU communications for systems with an NVLink CPU connection, NVLink can have significant performance benefit for GPU-to-GPU (peer-to-peer) communications as well. A second new NVIDIA white paper focuses on these peer-to-peer benefits from NVLink, showing how systems with next-generation NVLink-interconnected GPUs are projected to deliver considerable application speedup compared to systems with GPUs interconnected via PCIe.

[...]

The paper examines five multi-GPU algorithms and applications important to HPC: exchange and sort, FFT, AMBER Molecular Dynamics (PMEMD), ANSYS Fluent Computational Fluid Dynamics (CFD), and QUDA Lattice Quantum Chromodynamics (LQCD).

http://devblogs.nvidia.com/parallelforall/how-nvlink-will-enable-faster-easier-multi-gpu-computing/

 

Quote

[NVLink] allows data sharing at rates 5 to 12 times faster than the traditional PCIe Gen3 interconnect, resulting in dramatic speed-ups in application performance and creating a new breed of high-density, flexible servers for accelerated computing.

Designed to solve the challenges of exascale computing, NVLink is a fundamental ingredient of the U.S. Department of Energy's next-generation supercomputers. These new systems—"Summit" at Oak Ridge National Laboratory and "Sierra" at Lawrence Livermore National Laboratory—will be three times faster than today's fastest supercomputer.

http://www.nvidia.com/object/nvlink.html

 

"8-Way SLI!"

Spoiler

8-way GPUs will only be possible over NVLink, meaning it will only appear in specialized systems, and unless NVLink replaces the SLI implementation we have now in consumer cards, 8-way GPUs likely won’t see usage outside of NVIDIA’s devboxes, which they announced would have up to 8 GPUs in the future with Pascal, the only mention of 8-way GPUs thus far (and it was never stated that they would be running in SLI, which is not required for multi-GPU compute applications that devboxes are used for). Various tech “news” websites as usual chose to interpret that as “8-way SLI will now be a thing!!!!!”
 

Quote

The addition of NVLink to Pascal will let data move between GPUs and CPUs five to 12 times faster than they can with today’s current standard, PCI-Express. This is greatly [sic] benefits applications, such as deep learning, that have high inter-GPU communication needs.

NVLink allows for double the number of GPUs in a system to work together in deep learning computations. In addition, CPUs and GPUs can connect in new ways to enable more flexibility and energy efficiency in server design compared to PCI-E.

http://blogs.nvidia.com/blog/2015/03/17/pascal/

Even if NVLink turns out to be used in the desktop market as a multi-GPU interconnect, 8-way SLI doesn’t really have any impact on most consumers. Sure it will be fun to marvel at the one or two showcase builds that at least one or two people will surely build for conventions, but realistically hardly anyone is going to use this feature anyway even if it appears on the desktop. So it is a bit funny to see people say “You should wait for Pascal because...” and list 8-way SLI as a bullet point, as if it’s a useful consideration for regular people.

But anyway, this all depends on whether NVLink will be available as a GPU interconnect in regular graphics cards anyway. We don't know enough at this time.
 

"32 GiB of Memory!"

Spoiler

*UP TO 32 GiB of memory. The highest end Pascal-based GPU will be equipped with 32 GiB of memory. This will be the highest-end out of ALL Pascal cards, including Quadros and Teslas. It's extremely unlikely any gaming card will have more than 16 GiB, and even that will probably only be seen on a Titan-class card while the regular lineup will probably max out at 8 GiB. Still quite a lot and unlikely to make any difference for gamers, but all the same it still isn't 32 GiB, so saying “Pascal’s gonna be amazing for gaming! It’s gonna have 32 GiB of memory which is way more than AMD’s next gen gaming cards will have!” is a bit misleading when 32 GiB won’t be found outside of workstation cards that gamers don’t care about.

If you want to play that card, you could do the same kind of thing with generations gone by and say that AMD's Hawaii GPU was equipped with 16 GiB (wow!) of memory or that Kepler GK110 cards had up to 12 GiB of memory by including Firepros and Quadros in your tally, but pretty much anyone will agree that as far as gamers are concerned Hawaii had up to 8 GiB at the very most and 4 GiB normally at the high end, while Kepler GK110 had 6 GiB at the very most and 3 GiB normally. So once again, the 32 GiB figure here is unlikely to apply to any cards gamers are interested in, just like the 16 GiB Hawaii cards and 12 GiB Kepler cards, these amounts will only be found in high-end Quadros and Teslas.

Yes a Pascal card of some kind will have 32 GiB of memory, but it’s probably not a good idea to get hyped about having 32 GiB cards on the market if the only market you pay attention to is the gaming market, because it’s unlikely you’ll see any 32 GiB cards appearing there.
 

"HBM2 (3D Memory)!"

Spoiler

HBM2 is an advanced type of memory which will replace GDDR5 on the new high-end Pascal cards (low and mid-tier cards will use GDDR5X). It can be scaled to much higher densities and much higher bandwidth than GDDR5 while taking less space and using less power. These benefits will be seen on both compute and gaming cards alike, in the form of smaller cards and higher power efficiency. But as far as performance goes, the benefit is marginal at best for gaming since they are rarely limited by memory bandwidth.

Compute applications are a lot heavier on memory bandwidth than gaming, so they’ll benefit a lot more. Gaming performance may see some small improvements, but it’s not going to be much. We’ve already seen the impact of massive memory bandwidth increases with the AMD R9 Fury X; it does help the card pull a little bit ahead at higher resolutions, but it’s not a game changer by any means. With HBM, the R9 Fury X has certainly pushed well beyond any kind of bandwidth bottleneck at this point, so adding even more with Pascal won't really help beyond that. The diminishing returns barrier has been crossed into the no-returns zone. For gaming, anyway.

We’ve heard so far from NVIDIA that Pascal will have a Terabyte per second of memory bandwidth in its full form (twice as much as the Fury X).
 

The things NVIDIA has actually claimed specifically:

• NVLink will be useful as a replacement for PCI Express in supercomputers, and will have 80 GB/s of total bandwidth shared between the CPU and the number of GPUs in the system. (NVIDIA developer blog)
• NVLink can be used as a GPU-GPU interconnect without replacing PCIe as the system interconnect, which provides great benefit for HPC and multi-GPU accelerated computing algorithms. No word on what this means for regular desktop cards and multi-GPU gaming. (NVIDIA developer blog)
• 2× the power efficiency (performance per watt) in SGEMM operations compared to Maxwell GM200; that’s a bit out of my depth so I don’t know how relevant that is to gaming applications, but I do think it’s important to notice that NVIDIA was very specific not to claim 2× power efficiency just as a general statement, so it probably won’t be, otherwise they would have just said that. (slide from keynote)
• 4× FP16 performance in mixed precision mode compared to Maxwell GM200; not really relevant for gaming, though it could mean the FP32 performance of Pascal is 2× that of Maxwell, this is only speculation so I don't know why I'm even mentioning it in this section (slide from keynote)
• Very roughly 10× the overall throughput for deep learning problems compared to Maxwell GM200 (slide from keynote)
• Up to 32 GiB of 3D memory (HBM) in highest-end cards (and not necessarily highest-end GeForce cards) (slide from keynote)
• Up to 3× the memory bandwidth of Maxwell GM200 (so, ≈1 TB/s) (slide from keynote)
• 8 GPUs in a future Pascal-based deep learning devkit (SLI not mentioned) (end of: https://www.youtube.com/watch?v=vdNRqZSRgfA)

So, I don’t mean to dampen the mood or say Pascal won’t be great, not at all; the reality is we have absolutely no idea how Pascal will be for gaming, there’s been no information about that topic yet. It might be only a marginal improvement, it might be totally amazing. Everything we’ve heard so far is about compute capabilities. As much as the sensationalist “news” sites across the web want to make it seem like all these things are applicable to gaming, they simply aren’t. So far the Pascal architecture seems to be entirely centered around high-performance compute and accelerated computing.

We’ll see what kind of gaming performance Pascal brings to the table in due time. For now, just sit back, relax, and be careful of what you read on the Internet.

Full GTC 2015 Keynote:

Spoiler

 

 

GTC 2016 Update:
https://devblogs.nvidia.com/parallelforall/inside-pascal/

NVIDIA has released specifications for the GP100 GPU ("Big Pascal"), as well as the NVIDIA Tesla P100, a compute acceleration module which will use the GP100 GPU.

 

Specifications of the GP100 GPU include:

• 3840 FP32 cores, 1920 FP64 cores (1/2 FP32)
• Cores are arranged in to 60 groups of cores called SMs, with 64 FP32 cores and 32 FP64 cores in each group
• 240 texture units (4 per SM)
• 4096-bit HBM2 memory interface (8 × 512-bit)
• 4 MiB L2 cache
• 15.3 billion transistors
• 610 mm² die area
• Manufactured by TSMC on a 16 nm fabrication process

Specifications for the Tesla P100 compute accelerator include:

• A GP100 GPU with 4 SMs disabled (56 out of 60 enabled), for a total of 3584 FP32 cores active (1792 FP64 cores)
• A 1328 MHz base frequency and 1480 MHz boost frequency
• 224 texture units enabled
• Up to 16 GiB of HBM2 DRAM on a full 4096-bit memory interface
• 300 watt TDP
• 5.3 FP64 TFLOPs at boost frequency (3× compared to Kepler GK110's 1.7 TFLOPs, and 25× compared to Maxwell GM200's pathetic 0.2 TFLOPs)
• With FP64 = 1/2 FP32, this implies 10.6 TFLOPs in theoretical FP32 performance, which is 16% faster than the theoretical performance of a Maxwell GM200 GPU at equal frequency (1.4 GHz), and 72% faster than a 1.0 GHz Maxwell GM200 GPU (TITAN X stock frequency). Also keep in mind that this is with only 56 out of 60 SMs enabled.

 

[PlayStation 2]

Nicely done. A well needed topic to keep the ignorance away.

[pagani123]

good clearing the air around pascal 

[MageTank]

Absolutely amazing. Thank you for taking the time to do this. I've been trying to explain the same things multiple times to multiple people, but now i have something concrete to link to!

[The Toaster]

Very informative. Cleared up a lot of things I didn't quite understand or was confused about.

[cesrai]

@Glenwing this will be a nice read and you should do a rebadge vs refresh thread as I'm lost.

[Clyne]

I'm interested in the FP64 performance........is it actually a performance increase over Maxwell. Maxwell had shit FP64 performance compared to Hawaii/Fiji. 

If AMD barely gets any HBM2 supply for their cards, then i doubt Nvidia would get some too. 

Please Nvidia bring back a card like Fermi again..........just keep the raw power and all the features for ultimate performance and don't care about the power consumption...or at least bring back the Voodoo name but i doubt that's going to happen. 

[Samfisher]

@Glenwing this will be a nice read and you should do a rebadge vs refresh thread as I'm lost.

Rebadge = taking something that already exists and wholesame naming it something new with hardly any changes.

 

Refresh = Same tech generation, usually faster in some way (although very little) with more under the hood benefits not apparent at first look.

[Glenwing]

@Glenwing this will be a nice read and you should do a rebadge vs refresh thread as I'm lost.

A rebadge is taking an old product and giving it a new name to make it seem like a new product when it isn't. A refresh on the other hand is basically a revision; based on the same package but with hardware improvements of some kind (so, just changing the stock frequency doesn't count as a refresh).

[JustCallMeVlad]

I'm more excited for 16nm

[Zah]

Thank you so much for this thread.

[Sniper Xls]

The question remains when will it becoming? 

[Lethal Seraph]

Thanks for clarifying. People need to read this before hyping crap.

[D2ultima]

This is great and I wish we could write one about DirectX 12. Holy meowmix, people are tossing all sorts of stuff around: "well DX12 will extend the life of my AMD CPU!" and "DX12 will let me use AMD and nVidia GPUs together as one so I'll get both of their technologies!" and "DX12 will mean all these demanding games (already released) will run perfectly!" and "DX12 means we'll get vRAM doubling in SLI!" and "nVidia can't use DX12!".

 

I cri.

[Briggsy]

Good work putting this together, kudos.

And compute capability, yeah we've seen how well that's worked out for AMD in dx11, having mountains of compute capability. And who really knows yet how much compute shaders dx12 games will use.

[Statecat]

Yeah this is why I was cool with getting the 980 Ti. I don't do any CAD or crazy editing work so I wouldn't need a pascal card. Pascal could definitely be amazing for the work world though. Just imagine the possibilities for game development!

[ShanaSakai]

Thank you for this, I was wondering whether I should wait b/c I'm worried I would need a new motherboard, guess not, so I'll just continue with my new build!

[Metro]

I expect low end and mid end Pascal to be Q2 2016, then high end Pascal in Q2 2017

 

One is the 16nm fab is new...and untested...and you can't go large and power hungry on first generation chips. Don't believe how unreliable a new fab is? Look at how long it took Nvidia to pick between Samsung and TSMC and finally siding with TSMC due to their longer working history and reliability.

Second point is...they will kill their own sales by offering their best card right at the release of the new generation. They will not be able to give any reason at all to their enthusiast market to upgrade their card within 3 years of launch. Whereas when you look at Kepler, people bought the 680, then the Titan came out, then the 780, and most people who bought the 680 eventually upgraded to those cards. And then when Maxwell launched, they started with just the 750 and 750ti. Then released the 980 that some bought, and finally after a while, the Titan X and 980ti which gave people (enthusiast market, again) a reason to upgrade.

Nvidia likes incremental upgrades. They don't want you to buy a card now, and launch another card in a year that completely destroys the card you bought. They put out one card, then 12 months later they bring out another card that is generally just 10-20% faster, so it's better than what they offered before, but not enough to **** off anyone who bought their last gen cards. And then another 6-12 months later they put out an even better card, with 50%+ better performance than their older cards, and people start upgrading again.

If on day one of the Pascal launch, if they come out with their absolute best card, they are going to have nothing interesting to bring to market for 2 to 3 years until Volta comes out. And that would be silly. Even if it were possible with the new 16nm fab. In terms of business, you need to offer a product that is a bit better than your competition, without being too much better/too costly for you. So they just need to put out a slight performance increase, but sell the card on much lower power consumption, wait for AMD to release something else, and then launch a bigger die version themselves, and back/forth they go. Just a quick reference:

FERMI

GTX 580 = 520mm2

KEPLER

GTX 680 = 294mm2
GTX Titan = 551mm2

MAXWELL

GTX 750ti = 148mm2
GTX 980 = 398mm2
GTX Titan X = 601mm2

Do you see the pattern? Small, Medium, Big, restart. Don't think about it based on die size. Because it's really just about transistor count. Pascal at 16nm, even a 294mm2 sized die like the GTX 680, along with HBM2, would result in performance close to the Titan X...and perhaps even higher due to lower heat/power consumption allowing higher clocks.
Hope that helps.


Also, NVIDIA is launching the GTX 990, so it would not make sense to release high end Pascal in Q2 2016

We also know HBM2 is going to have problems with amount of stacks as there was not enough HBM1 for AMD, so if NVIDIA and AMD are using HBM2, there will not be enough (Titan O, will need 16GB HBM2) this is why NVIDIA will have to wait for Samsung HBM2 (which starts production in Q2 2016) so it will take till 2017 for there to be enough and for NVIDIA to implement it on the GTX 1080ti and Titan O

[Coaxialgamer]

BUT ... BUT ....

I was riding the hype train !... You destroyed my dreams of a killer gpu ...

On the other hand it wouldn't make much sense for nvidia to make a gpu that would destroy it's older lineup.

[MacNite]

Where die you geht the information that Nvidia is launching a GTX990?

I just saw this as an April's fools joke over a year ago.

A source would be nice

Greets

[Metro]

Where die you geht the information that Nvidia is launching a GTX990?

I just saw this as an April's fools joke over a year ago.

A source would be nice

Greets

 

It is here

 

http://wccftech.com/nvidia-launching-dual-gpu-flagship-graphics/

[MacNite]

Thanks!

An now I think I know what I need next

[SoloDolo]

Fantastic thread OP! Finally some solid information. 

[ybriK]

Interesting information but Nvidia also forgot to claim how they'll be disabling the possibilities of multi-GPU use. Well let's be honest here they will do it and I mark my words.

 

 

Remember PhysX?

[RnkPa]

Thanks for the information.. I wud really like a jump in energy efficiency in pascal cards.. Save some electricity bills and money on psu's