4 ms·

installed standard Ubuntu 14.04 w/ Caffe, Torch, Theano, BIDMach, cuDNN v2, and CUDA 7.0 whoa - are you telling me that the nVidia drivers on Linux are so stab

installed standard Ubuntu 14.04 w/ Caffe, Torch, Theano, BIDMach, cuDNN v2, and CUDA 7.0 whoa - are you telling me that the nVidia drivers on Linux are so stable that they are building a commercial deep learning system on top of that. Is this the same thing as normal graphics drivers ?

wmf 11y ago
AFAIK Linux with Nvidia GPUs is used in plenty of supercomputers and Hollywood special effects companies, so the drivers must be stable enough.
somerandomone 11y ago
The difference is between "given this particular hardware and OS setup, the driver will work correctly, guaranteed." vs. "on your (discontinued) Sony laptop with a strange hardware interface running the beta Slackware release the driver will probably work"
- wtallis 11y ago
  It's also the difference between "using our tools implementing our API on our hardware" vs. "trying to figure out the right thing when every component has a slightly different take on the API spec and the applications using the API make mistakes that we have to try to correct for with unreliable heuristics". Developers in the scientific computing world care about correctness a lot more than game developers working under unrealistic deadlines and with no commitment to long-term maintenance.
  - sandGorgon 11y ago
    that doesnt make sense to me. If the driver works and does not crash the operating system (kernel panic FTW), that's good enough at this point. nVidia uses CUDA or OpenGL - so its not quite the question of proprietary API. At this point, I'm not worried about "framerate on my linux box isnt as good as windows".. its more "it works...".
    - wtallis 11y ago
      An OpenGL driver capable of running most commercial games is horrifically more complex than a CUDA driver. An OpenGL driver that merely works according to spec is useless in practice. To achieve any practicality for non-trivial use cases an OpenGL driver has to take an attitude of "do what I mean, not what I say", much like web browsers and Windows' backwards compatibility. CUDA doesn't have those problems. NVidia never has to deal with developers complaining that their broken code worked fine on some other vendor's platform. They don't have to worry about programs relying on some esoteric decades-old feature that NVidia doesn't care about but had to implement anyways for standards compliance. And since CUDA is operating in the professional segment of the market, they can take their time when it comes to compatibility with bleeding-edge versions of other OS components.
      - p0nce 11y ago
        Also compute API makes a lot more sense than OpenGL, with an understandable mapping to actual GPU resources.
Galaxeblaffer 11y ago
Nvidias drivers for Linux are pretty darn good these days..
- yulaow 11y ago
  Those for laptops are pretty shitty, we are still using third party programs to support optimus technology (with not great results).
  - ris 11y ago
    Optimus technology is a fairly stupid idea in the first place though that exists for business reasons rather than technical.
    - StavrosK 11y ago
      What's stupid about it? Sounds pretty smart to not have to use the powerful GPU if I'm only browsing the web, to me.
      - wtallis 11y ago
        Intel's got the good fabs, but they can't get along well enough with the folks who know how to make a good GPU, so it's impossible to get an affordable desktop or mobile system at any price that has a decent GPU without actually paying for two entirely different GPUs.
      - ris 11y ago
        "The powerful GPU." You've been sold the idea that a "powerful GPU" needs to suck a lot of power all the time. There is no real reason a "powerful GPU" shouldn't be able to scale its power usage way down when doing something simple like browsing the web. The only reason NVidia weren't able to do the "low power" thing on these systems is they weren't able to be the ones putting their GPUs on the same die as the CPU like Intel (& AMD) were. But of course they still wanted part of the action, so people ended up being sold this massive engineering bodge and told it's a good thing.
pantalaimon 11y ago
the nVidia driver is the most stable and and feature-complete driver you can get on Linux.
- mwilcox 11y ago
  Yep. Godsend in comparison to AMD drivers
gcb0 11y ago
for CUDA only. and keep in mind, it is a binary blob. proprietary to the core. in fact, CUDA the protocol itself is kinda proprietary. if you base your solution on OpenCL, you can use all the vendors. nvidia, amd, imb, intel, altera, etc... but NVIDIA spends billions on marketing to convince you that only CUDA matters. the same way sony spent millions to convince you that only laser disc^H^H^H^H^H^H mini disc ^H^H^H^H^H memory stick ^H^H^H^H^H blue ray, matter.
m-i-l 11y ago
In Nvidia's deep learning lab last week [0] they described setting up on Ubuntu as "the easy way" and Windows the "not so easy way" [1], adding that their developers are on Linux so no-one has really tested it on Windows. [0] https://developer.nvidia.com/deep-learning-courses https://developer.nvidia.com/deep-learning-courses [1] http://on-demand.gputechconf.com/gtc/2015/webinar/deep-learning-course/getting-started-with-digits.pdf http://on-demand.gputechconf.com/gtc/2015/webinar/deep-learn...

wmf 11y ago

AFAIK Linux with Nvidia GPUs is used in plenty of supercomputers and Hollywood special effects companies, so the drivers must be stable enough.

somerandomone 11y ago

The difference is between "given this particular hardware and OS setup, the driver will work correctly, guaranteed." vs. "on your (discontinued) Sony laptop with a strange hardware interface running the beta Slackware release the driver will probably work"

wtallis 11y ago
It's also the difference between "using our tools implementing our API on our hardware" vs. "trying to figure out the right thing when every component has a slightly different take on the API spec and the applications using the API make mistakes that we have to try to correct for with unreliable heuristics". Developers in the scientific computing world care about correctness a lot more than game developers working under unrealistic deadlines and with no commitment to long-term maintenance.
- sandGorgon 11y ago
  that doesnt make sense to me. If the driver works and does not crash the operating system (kernel panic FTW), that's good enough at this point. nVidia uses CUDA or OpenGL - so its not quite the question of proprietary API. At this point, I'm not worried about "framerate on my linux box isnt as good as windows".. its more "it works...".
  - wtallis 11y ago
    An OpenGL driver capable of running most commercial games is horrifically more complex than a CUDA driver. An OpenGL driver that merely works according to spec is useless in practice. To achieve any practicality for non-trivial use cases an OpenGL driver has to take an attitude of "do what I mean, not what I say", much like web browsers and Windows' backwards compatibility. CUDA doesn't have those problems. NVidia never has to deal with developers complaining that their broken code worked fine on some other vendor's platform. They don't have to worry about programs relying on some esoteric decades-old feature that NVidia doesn't care about but had to implement anyways for standards compliance. And since CUDA is operating in the professional segment of the market, they can take their time when it comes to compatibility with bleeding-edge versions of other OS components.
    - p0nce 11y ago
      Also compute API makes a lot more sense than OpenGL, with an understandable mapping to actual GPU resources.

Galaxeblaffer 11y ago

Nvidias drivers for Linux are pretty darn good these days..

yulaow 11y ago
Those for laptops are pretty shitty, we are still using third party programs to support optimus technology (with not great results).
- ris 11y ago
  Optimus technology is a fairly stupid idea in the first place though that exists for business reasons rather than technical.
  - StavrosK 11y ago
    What's stupid about it? Sounds pretty smart to not have to use the powerful GPU if I'm only browsing the web, to me.
    - wtallis 11y ago
      Intel's got the good fabs, but they can't get along well enough with the folks who know how to make a good GPU, so it's impossible to get an affordable desktop or mobile system at any price that has a decent GPU without actually paying for two entirely different GPUs.
    - ris 11y ago
      "The powerful GPU." You've been sold the idea that a "powerful GPU" needs to suck a lot of power all the time. There is no real reason a "powerful GPU" shouldn't be able to scale its power usage way down when doing something simple like browsing the web. The only reason NVidia weren't able to do the "low power" thing on these systems is they weren't able to be the ones putting their GPUs on the same die as the CPU like Intel (& AMD) were. But of course they still wanted part of the action, so people ended up being sold this massive engineering bodge and told it's a good thing.

pantalaimon 11y ago

the nVidia driver is the most stable and and feature-complete driver you can get on Linux.

mwilcox 11y ago
Yep. Godsend in comparison to AMD drivers

gcb0 11y ago

for CUDA only. and keep in mind, it is a binary blob. proprietary to the core. in fact, CUDA the protocol itself is kinda proprietary. if you base your solution on OpenCL, you can use all the vendors. nvidia, amd, imb, intel, altera, etc... but NVIDIA spends billions on marketing to convince you that only CUDA matters. the same way sony spent millions to convince you that only laser disc^H^H^H^H^H^H mini disc ^H^H^H^H^H memory stick ^H^H^H^H^H blue ray, matter.

m-i-l 11y ago

In Nvidia's deep learning lab last week [0] they described setting up on Ubuntu as "the easy way" and Windows the "not so easy way" [1], adding that their developers are on Linux so no-one has really tested it on Windows. [0] https://developer.nvidia.com/deep-learning-courses https://developer.nvidia.com/deep-learning-courses [1] http://on-demand.gputechconf.com/gtc/2015/webinar/deep-learning-course/getting-started-with-digits.pdf http://on-demand.gputechconf.com/gtc/2015/webinar/deep-learn...