Why not CUDA?
category: offtopic [glöplog]
I wonder, why demoscene didn't catch the CUDA? Is it because it is one vendor technology only?
Or does the problem lie in shaders being more 'fancy' for demos even though it has some limitations for computation?
Or does the problem lie in shaders being more 'fancy' for demos even though it has some limitations for computation?
When I checked CUDA, I concluded it was totally kickass and suited my needs at 100%. But then, I checked the licence agreement, and absolutely every algorithm you develop belongs to NVidia apparently as well for free. So it's not only vendor specific, it's a damn good hover for all kind of new ideas for free for NV. That's a problem to me, on top of the vendor specific bits.
Well it's nVidia only, the executable is bigger, and there are not many things (specifically graphic effects) that you can do in CUDA which can't be done with DirectX/OpenGL. So in my opinion, the disadvantages win.
Only CUDA demo I know was Megademo 1 by Hard Coders. Of course I've never seen it because it runs only on NVidia hardware but from the screen grab it doesn't look like it was putting CUDA to good use.
Some OpenCL stuff would be cool I suppose, done with some style. Currently that stuff is a bit of a gimmick.
Some OpenCL stuff would be cool I suppose, done with some style. Currently that stuff is a bit of a gimmick.
maq - we should. You're right. Apologies. We are using direct-compute though, which is rather equivalent.
sounds like maq is going to make a cuda demo!
Vendor locked. It makes sense for projects not intended for distribution though, and you get a better feel for the hardware. It's actually quite nice coding with it, but I haven't tried OpenCL or direct compute and I'm guessing they're similar if not for a little more abstraction.
Cuda, Direct-compute, DX-HLSL, GLSL, AMD Stream...aww maaaan....Thats why i drink alcohol.
It's vendor locked, does very little that opencl doesn't, requires like opencl large libs and maybe troublesome installation, so unsuitable for intros.
DirectCompute (or actually just compute shaders in dx11 - we need the graphics connection anyway) may do less than the others, but just works across vendors (saw a couple of nv compiler bugs but that's about it), ofcourse integrates nicely with dx, and is already there on the machine. It's even possible to do decent 1k stuff with compute shaders.
I've released 3 CS based intros for now, not sure if there are others.. michigan, gda and traskogen.
DirectCompute (or actually just compute shaders in dx11 - we need the graphics connection anyway) may do less than the others, but just works across vendors (saw a couple of nv compiler bugs but that's about it), ofcourse integrates nicely with dx, and is already there on the machine. It's even possible to do decent 1k stuff with compute shaders.
I've released 3 CS based intros for now, not sure if there are others.. michigan, gda and traskogen.
why you people still care about being vendor specific (or not)???
i mean, if you're going to make a demo for the latest generation of graphics cards, it will be mostly watched on youtube anyway, so why care about being tied to vendor (actually, to one vendor instead of two of them)?
this has no sense.
you better just do something impressive - you will have your audience then..
do it however you like it, if you want microsoft's platform more that actual hw manufacturer's one - fine. you can have your own vision and use whatever you like, but this 'one vendor' argument is just too obvious and actually ridiculous.
i mean, if you're going to make a demo for the latest generation of graphics cards, it will be mostly watched on youtube anyway, so why care about being tied to vendor (actually, to one vendor instead of two of them)?
this has no sense.
you better just do something impressive - you will have your audience then..
do it however you like it, if you want microsoft's platform more that actual hw manufacturer's one - fine. you can have your own vision and use whatever you like, but this 'one vendor' argument is just too obvious and actually ridiculous.
you need an entire library to upload textures to the GPU, ping-pong between them and redirect fragment output to an offscreen FBO?
overkill for demos.
overkill for demos.
Quote:
i mean, if you're going to make a demo for the latest generation of graphics cards, it will be mostly watched on youtube anyway, so why care about being tied to vendor
Name 3 groups that really do that (release a demo for the latest generation of graphics cards). Most demos released nowadays run perfectly smooth on my >2 year old laptop. Even the majority of the raytracing 4KB intros run OK.
One more con: you have to use CUDA's compiler for your code.
I prefer OpenCL. It's cross verndor: AMD even made a CPU driver for it. (It comes handy when you work on a low end pc/netbook or in a virtual machine)
Also, you only need to link to a lib, same as with OpenGL.
I prefer OpenCL. It's cross verndor: AMD even made a CPU driver for it. (It comes handy when you work on a low end pc/netbook or in a virtual machine)
Also, you only need to link to a lib, same as with OpenGL.
xernobyl i don't get your point. i only know now that you've got a good laptop.
what i'm talking about is that people mostly watch recorded demos usually nowadays and not running them anyway, and less and less people care about having a mighty gpu on their laptop.
what i'm talking about is that people mostly watch recorded demos usually nowadays and not running them anyway, and less and less people care about having a mighty gpu on their laptop.
Quote:
...mostly watch recorded demos usually nowadays and not running them anyway...
will someone burn him at the stake, or do i have to do it myself?
@ton: It just doesn't fit the demoscene mentality. Also, pc is not a platform where that attitude would be common.
Demos aren't just about the capabilities of the latest hardware, but rather about the capabilities of the creators.
Demos aren't just about the capabilities of the latest hardware, but rather about the capabilities of the creators.
guys, i know your pain.
but the topic is: why not cuda. and the most common reason is: one vendor thing, which basically means: possibly not as much viewers as you'd want.
and the answer is: that's ridiolous, because if you want to make something remarkable with cuda it will be mostly watched on youtube, as most of the people just have laptops with lousy 'gpu's nowadays and to show off something cool you need something like 460+ card.
if you don't care about most of the people, and that's actually what fits the scene mentality, then i don't see anything wrong with cuda excepting that it's not in the drivers. and it's *actually* better than poor opencl that's being developed by a bunch of people with different goals right now.
but the topic is: why not cuda. and the most common reason is: one vendor thing, which basically means: possibly not as much viewers as you'd want.
and the answer is: that's ridiolous, because if you want to make something remarkable with cuda it will be mostly watched on youtube, as most of the people just have laptops with lousy 'gpu's nowadays and to show off something cool you need something like 460+ card.
if you don't care about most of the people, and that's actually what fits the scene mentality, then i don't see anything wrong with cuda excepting that it's not in the drivers. and it's *actually* better than poor opencl that's being developed by a bunch of people with different goals right now.
I forgot that Cuda demo that when it came I said, wow demo with new technology but I can't run it. Now you reminded it me, I have a GTS450, and I run and it says "cudart.dll not found". Shit! But I have installed drivers and stuff. I installed now Cuda SDK and Cuda drivers and I still get it. I found some cudart.dll in a folder put it, and got a dynamic resolve something mistake too because it was not the same cudart version as intended.
Arghh..
I was thinking about doing a demo using OpenCL once but if things are like this and none will be able to view my demo then I will leave it for the future when these libs mature.
Arghh..
I was thinking about doing a demo using OpenCL once but if things are like this and none will be able to view my demo then I will leave it for the future when these libs mature.
Optimus: come on. you can write demo using DirectX which won't run without a proper .DLL. it happend to me many times when DirectX was not on board.
Regarding vendor specific problem - isn't it similar to Direct Compute limiting us to Windows only?
nystep, you wrote: "But then, I checked the licence agreement, and absolutely every algorithm you develop belongs to NVidia apparently as well for free."
Can you be more specific about this?
Can you be more specific about this?
Funny enough, after trying installing CUDA sdk, CUDA toolkit, even old CUDA toolkit 2.3 as instructed by someone on the thread of that demo, and copying the cudart.dll and other versions to the folder of the demo nothing still worked. And later I realized I have fucked up my system and got this http://www.youtube.com/watch?v=S7f0mf5xc10 which is irrelevant with CUDA but something went terribly wrong :P
Quote:
if you don't care about most of the people, and that's actually what fits the scene mentality
I may disagree on this, at least on my side i always try to do my best to acheive both performance and compatibility, and it's not always easy when you do the stuff in the garage you know. :) We don't have as much $ for dev as game companies have for sure. We can't test every single configuration out there. We don't even know clearly what type of hardware our audience or potential audience has. I think we're doing actually quite good considering what means we have.
@maq, well, let's install the latest CUDA sdk together. My install is cudatoolkit_4.0.17_win_64.msi .
Let's now check the licence agreement together.
Quote:
7.1 Feedback. Notwithstanding any Non-Disclosure Agreement executed by and between the parties, the parties agree that in the event Licensee or NVIDIA provides Feedback (as defined below) to the other party on how to design, implement, or improve the SOFTWARE or Licensee’s product(s) for use with the SOFTWARE, the following terms and conditions apply the Feedback:
7.1.2 Residual Rights. Licensee agrees that NVIDIA shall be free to use any general knowledge, skills and experience, (including, but not limited to, ideas, concepts, know-how, or techniques) (“Residuals”), contained in the (i) Feedback provided by Licensee to NVIDIA; (ii) Licensee’s products shared or disclosed to NVIDIA in connection with the Feedback; or (c) Licensee’s confidential information voluntarily provided to NVIDIA in connection with the Feedback, which are retained in the memories of NVIDIA’s employees, agents, or contractors who have had access to such Residuals. Subject to the terms and conditions of this Agreement, NVIDIA’s employees, agents, or contractors shall not be prevented from using Residuals as part of such employee’s, agent’s or contractor’s general knowledge, skills, experience, talent, and/or expertise. NVIDIA shall not have any obligation to limit or restrict the assignment of such employees, agents or contractors or to pay royalties for any work resulting from the use of Residuals.
7.2 Freedom of Action. Licensee agrees that this Agreement is nonexclusive and NVIDIA may currently or in the future be developing software, other technology or confidential information internally, or receiving confidential information from other parties that maybe similar to the Feedback and Licensee’s confidential information (as provided in Section 7.1.2 above), which may be provided to NVIDIA in connection with Feedback by Licensee. Accordingly, Licensee agrees that nothing in this Agreement will be construed as a representation or inference that NVIDIA will not develop, design, manufacture, acquire, market products, or have products developed, designed, manufactured, acquired, or marketed for NVIDIA, that compete with the Licensee’s products or confidential information.
For me, the current licence agreement is at least clarifying the obscure points of the former versions which i had access to.
Offtopic: How do I opencl from visual studio express?
AMD A8-3850 (to DDR3-1866)/Radeon 6670 (to DDR3-1600 !? ) here
AMD A8-3850 (to DDR3-1866)/Radeon 6670 (to DDR3-1600 !? ) here
