CUDA library too large to deploy

Accelerated Computing CUDA CUDA Programming and Performance
1

CUDA has a great advantage in inference speed,but CUDA+CUDNN library size more than 4G, Is there a solution to this problem? For example, only single computing power GPU is supported.

1 Like
3

Are you sure you don’t need just runtime packages, like e.g. a compiled TensorFlow binary package and the CUDA driver?

CUDNN is a set of primitives that can be used in source code to build applications, and that depends on developer packages (CUDA compiler etc) which of course is big.

4

It is, indeed, very annoying - the size that the CUDA distributable has gotten too.

It is high time NVIDIA offered:

  1. A minimal CUDA toolkit distribution, with some core libraries, no IDEs and no frills.
  2. An installer which lets you select the components you want, then downloads and installs just them.

It’s not difficult to write nor difficult to maintain. Come on, NVIDIAns, just go do it.

edit: Apparently, a “network installer” is available for Windows which does just this. Also, if you’re ok with your Linux distribution’s version, you can use apt-get install / yum install / etc. , depending on your distribution. It’s only the non-windows runfile which is huge.

1 Like
5

(1) looks like a non-starter, as N users will have N different ideas as to what comprises a minimal useful distribution.
(2) seems useful, provided that automatic checking for (and pulling of) dependencies is incorporated into such a scheme.

So, are you planning to file an RFE for this?

6

Well… everyone at NVIDIA who’s ever installed or packaged CUDA knows this is an incredible annoyance, so I shouldn’t have to file it, but - why not, sure. I’ll file an issue.

(1) looks like a non-starter, as N users will have N different ideas as to what comprises a minimal useful distribution.

Well, right now the N users get nothing. With an offering, most of these N users will have something that’s not-minimal-but-pretty-small, and a few will not be satisfied by it, and will want option (2.)

(2) seems useful, provided that automatic checking for (and pulling of) dependencies is incorporated into such a scheme.

Sure, but IIANM, dependencies can be limited to other selections within the installer.

7

Filed: