Data Science

NVIDIA CUDA Toolkit Symbol Server

Decorative image of two boxes with libcuda.sym labels.

NVIDIA has already made available a GPU driver binary symbols server for Windows. Now, NVIDIA is making available a repository of CUDA Toolkit symbols for Linux.

What are we providing?

NVIDIA is introducing CUDA Toolkit symbols for Linux for an application development enhancement. During application development, you can now download obfuscated symbols for NVIDIA libraries that are being debugged or profiled in your application. This is shipping initially for the CUDA Driver ( and the CUDA Runtime (, with more libraries to be added.

For instance, when an issue appears to relate to a CUDA API, it may not always be possible to provide NVIDIA with a reproducing example, core dump, or unsymbolized stack traces with all library load information. Providing a symbolized call stack can help speed up the debug process.

We are only hosting symbol files, so debug data will not be distributed. The symbol files contain obfuscated symbol names.

Quickstart guide

There are two recommended ways to use the obfuscated symbols for each library:

  • By unstripping the library
  • By deploying the .sym file as a symbol file for the library
# Determine the symbol file to fetch and obtain it
$ readelf -n /usr/local/cuda/lib64/

# ... Build ID: 70f26eb93e24216ffc0e93ccd8da31612d277030
# Browse to to determine filename to download
$ wget

# Then with appropriate permissions, either unstrip,
$ eu-unstrip /usr/local/cuda-12.2/targets/x86_64-linux/lib/ –o /usr/local/cuda-12.2/targets/x86_64-linux/lib/

# Or, with appropriate permissions, deploy as symbol file
# By splitting the Build ID into first two characters as directory, then remaining with ".debug" extension
$ cp /usr/lib/debug/.build-id/70/f26eb93e24216ffc0e93ccd8da31612d277030.debug

Example: Symbolizing

Here is a simplified example to show the uses of symbolizing. The sample application test_shared has a data corruption that leads to an invalid handle being passed to the CUDA Runtime API cudaStreamDestroy. With a default install of CUDA Toolkit and no obfuscated symbols, the output in gdb might look like the following:

Thread 1 "test_shared" received signal SIGSEGV, Segmentation fault.
0x00007ffff65f9468 in ?? () from /lib/x86_64-linux-gnu/
(gdb) bt
#0  0x00007ffff65f9468 in ?? () from /lib/x86_64-linux-gnu/
#1  0x00007ffff6657e1f in ?? () from /lib/x86_64-linux-gnu/
#2  0x00007ffff6013845 in ?? () from /usr/local/cuda/lib64/
#3  0x00007ffff604e698 in cudaStreamDestroy () from /usr/local/cuda/lib64/
#4  0x00005555555554e3 in main ()

After applying the obfuscated symbols in one of the ways described earlier, it would give a stack trace like the following example:

Thread 1 "test_shared" received signal SIGSEGV, Segmentation fault.
0x00007ffff65f9468 in libcuda_8e2eae48ba8eb68460582f76460557784d48a71a () from /lib/x86_64-linux-gnu/
(gdb) bt
#0  0x00007ffff65f9468 in libcuda_8e2eae48ba8eb68460582f76460557784d48a71a () from /lib/x86_64-linux-gnu/
#1  0x00007ffff6657e1f in libcuda_10c0735c5053f532d0a8bdb0959e754c2e7a4e3d () from /lib/x86_64-linux-gnu/
#2  0x00007ffff6013845 in libcudart_43d9a0d553511aed66b6c644856e24b360d81d0c () from /usr/local/cuda/lib64/
#3  0x00007ffff604e698 in cudaStreamDestroy () from /usr/local/cuda/lib64/
#4  0x00005555555554e3 in main ()

The symbolized call stack can then be documented as part of the bug description provided to NVIDIA for analysis.


When you have to profile and debug applications using CUDA and want to share a call stack with NVIDIA for analysis, use the CUDA symbol server. Profiling and debugging will be faster and easier.

For questions or issues, dive into the forum at Developer Tools.

Related resources

Discuss (2)