TheGuruStudOh, yeah, windows is of no help. I just use the driver panel to set displays. Miraculously (ha), it works just fine after that.

renz496www.nvidia.com/object/accelerate-inference.html

this is nvidia tesla card made specifically for deep learning stuff. look at tesla P4 (GP104 based) and tesla P40 (GP102 based). for both product nvidia did not mention it's FP16 performance. only tesla P100 FP16 performance was listed. for this kind of product there is no reason for nvidia to artificially limit it's FP16 performance unless the performance was poor to being with. and then source from anandech:

www.anandtech.com/show/10510/nvidia-announces-nvidia-titan-x-video-card-1200-available-august-2nd

look at the spec table both Titan X 2016 and GTX 1080 FP16 (native) performance is listed as 1/64 (of FP32 performance)

renz496that prospect is interesting but physic engine is integrated on game engine level not operating system. imagine the game suddenly cannot work when being port to linux or PS4 since both did not use Microsoft direct x. i don't think developer like the idea of changing physics engine in their game based on what operating system they use because right now when it comes to physic engine the one they use will pretty much work on all platform.

the54thvoidhttp://wccftech.com/nvidia-titan-xp-titan-x-385-12-driver-update-massive-performance/

FordGT90ConceptUnfortunately doesn't explain how.

vega22so correct me if i'm wrong here, but doesn't this just confirm that nv deliberately nerfed the titans at the prosumer shit they claimed the card was meant for?

cdawallWonder if any of the FP16 driver based limitations are removed.

FordGT90ConceptHence my FP16 rant. I don't know if these programs liberally use FP16 though. Could be FP64 too for that matter.

FordGT90ConceptBut they do: Titan Xp has an FP16 rate of 1:64 where it's GP102 Quadro brother is 1:1. Only GP100 supports 2:1 like Vega does. The point is that Titan Xp is seriously gimped in everything that isn't FP32 just because it's not a professional card. Because this announcement enables more professional features in Titan Xp (likely because the threat of Vega), it gives a strong impression that NVIDIA decided to not gimp Titan Xp (or severely reduce the gimping). GP100 (what you're talking about) isn't available to consumers at all.

FordGT90ConceptIt's all three layers: hardware (GPU instructions specific to accelerating physics calculations), operating system (exposes the functionality in the hardware to software), and software (API bridges the gap between hardware manufacturers). Yes, game engines have to be updated to use the physics engine.

I think Microsoft will provide Linux and Mac OS X libraries for the physics component and allow the physics component to run independent of the graphics pipeline. If they don't, adaption rate will be low.

Game developers really want a physics API that works reliably in all scenarios. They can't build a game based on physics without that.

renz496can you give link showing that Quadro P6000 (GP102) that it's FP16 ratio is 1:1 to it's FP32? because when anantech mention FP16 (native) for GP102 and GP104 was rated at 1/64 of it's FP32 it means that the limitation was done on hardware level. just like Fiji FP64 was locked at 1/16 of it's FP32 performance be it in radeon or compute card like Firepro. and nvidia specific card designed for deep learning (tesla P40; GP102) nvidia sells them for it's INT 8 performance only. the reason for that most likely because it's FP16 performance was locked at 1/64 on hardware level.

renz496that would make stuff even more complicated than it was before. not sure if developer will be interested in that. just look at GPU accelerated Physics itself. vendor neutral solution was there for game developer to use for years but no game ever use it.

vega22so correct me if i'm wrong here, but doesn't this just confirm that nv deliberately nerfed the titans at the prosumer shit they claimed the card was meant for?

FordGT90ConceptYou have to look at P100 cards to find them retroactively reference P6000 card performance. Example here:
www.custompcreview.com/news/nvidias-quadro-gp100-announced-will-supercharge-deep-learning-design-capabilities/37421/

As long as there is no gimping involved, FP16 can usually be done at 1:1: to FP32 because it just spreads out the bits to make it work then converts back when done.

NVIDIA has given us a few answers to the question above. We have confirmation that the FP64 and FP16 rates are identical to GP104, which is to say very slow, and primarily there for compatibility/debug purposes. With the exception of INT8 support, this is a bigger GP104 throughout.

Meanwhile we have a die size for GP102: 471mm2, which is 139mm2 smaller than GP100. Given that both (presumably) have the same number of FP32 cores, the die space savings and implications are significant. This is as best of an example as we're ever going to get on the die space cost of the HPC features limited to GP100: NVLInk, fast FP64/FP16 support, larger register files, etc.

FordGT90ConceptThere's really only two: Havok and PhysX. Havok has effectively been abandonware since Intel bought them out years ago. PhysX, because it doesn't hardware accelerate on any GPU that isn't NVIDIA, can only be used for cosmetic stuff since it completely and utterly breaks on AMD/Intel GPUs when used for more than that. Havok, despite being abandonware is still pretty popular. Microsoft will include PhysX in probably DirectX 13. This effectively forces NVIDIA to play ball because if they refuse to support Microsoft's Havok-based physics engine, NVIDIA will not be able to market/sell DirectX 13 cards. I think it will only be a matter of years before Microsoft's physics engine replaces PhysX once it releases.

renz496probably they got it wrong? anandtech did asked nvidia about GP102 FP64 and FP16 support and nvidia answer is it is the same as GP104. remember anandtech is asking about the chip capabilities itself not specific card like quadro/geforce/tesla.



www.anandtech.com/show/10510/nvidia-announces-nvidia-titan-x-video-card-1200-available-august-2nd

renz496isn't that can lead to lawsuit against MS?

FordGT90ConceptNope, because DirectX is an API and to advertise as being capable of fully accelerating the API, they need to be fully compliant with it. Microsoft is not presenting a barrier to entry to NVIDIA and thus, no grounds for lawsuit.

Very few physics engines allow for GPU acceleration. That massively limits what can be done with it. This is why I singled out Havok and PhysX. Games have been doing simple CPU driven physics forever and still do.

renz496actually the titan is never meant to have pro drivers for existing professional application. the titan is meant to push developer to adopt GPGPU computing. they want developer to start building their ecosystem around GPGPU but getting tesla from the get go might be very expensive venture. so in a nutshell the titan was supposed to be poor man's tesla. not as cheap quadro replacement as many like to think. and when these developer want to roll out the real deal into the market nvidia hope they will replace that titan with true tesla which comes with full support from nvidia. that's why the titan never have this pro driver for existing pro application since the original kepler titan. maxwell titan did not have massive FP64 but it's FP16 somehow still useful for deep learning. with pascal nvidia cripple FP16 support pretty much on all their chip except for GP100. titan pascal was supposed to push INT 8 for inferencing. though nvidia probably no longer push that much in favor of tensor core. probably that's why nvidia keep GTX1080ti INT 8 performance intact.

now AMD after looking nvidia success for several years with titan probably want a piece of that pie as well. if nvidia is successful in this venture so why can't they as well? though it is not as simply throwing out a product out there. they must have something extra they can offer to make their solution is more attractive than competitor solution. hence AMD throwing in the pro driver for existing professional application except without the usual certification and extensive support that comes with true firepro card.

cdawallAh so hows this signed bios junk treating you? Its my favorite thing having to bypass checking to get a bios mod to work.

Multigpu works fine for nvidia right now. I am using it. Either it works or it doesn't in a game it isn't a game to see if the profile launches correctly like crossfire.

ImsochoboBios junk works perfect for nvidia and amd for me, no issues :) as it have since forever, I've done it since GF2 era.
I think Only RX480 needs bypass in windows?