Release Notes

Compute Sanitizer Release Notes.

Release notes and known issues.

1. Release Notes

Updates in 2022.3

  • Added support for the NVIDIA GH100/SM 9.x GPU architecture.
  • Added support for the NVIDIA AD10x/SM 8.9 GPU architecture.
  • Added support for lazy kernel loading.
  • Added memcheck support for distributed shared memory.
  • Added new options --num-callers-device and --num-callers-host to control the number of callers to print in stack traces.
  • Fix bug on Linux ppc64le where the host stack trace was incomplete.

Updates in 2022.2

  • Added memcheck support for use-before-alloc and use-after-free race detection. See the stream-ordered race detection documentation for more information.
  • Added leakcheck support for asynchronous allocations, OptiX resources and CUDA memmap (on Linux only for the latter).
  • Added option to ignore CUDA_ERROR_NOT_FOUND error codes returned by the cuGetProcAddress API.
  • Added new public API functions to allocate and free page-locked host memory.
  • Added public API callbacks for the event management API.

Updates in 2022.1.1

  • Fixed initcheck issue where the tool would incorrectly abort a CUDA kernel launch after reporting an uninitialized access on Windows with hardware scheduling enabled.

Updates in 2022.1

  • Added support for generating coredumps.
  • Improved support for stack overflow detection.
  • Added new option --target-processes-filter to filter the processes being tracked by name.
  • Added initcheck support for asynchronous allocations. Requires CUDA driver version 510 or newer.
  • Added initcheck support for accesses on peer devices. Requires CUDA driver version 510 or newer.
  • Added support for OptiX 7 applications.
  • Added support for tracking the child processes of 32-bit processes in multi-process applications on Linux and Windows x86_64.

Updates in 2021.3.1

  • Fixed intermittent issue on vGPU where synccheck would incorrectly detect divergent threads.
  • Fixed potential hang when tracking several graph launches.

Updates in 2021.3

  • Improved Linux host backtrace.
  • Removed requirement to call cudaDeviceReset() for accurate reporting of memory leaks and unused memory features.
  • Fixed synccheck potential hang when calling __syncthreads in divergent code paths on Volta GPUs or newer.
  • Added print of nearest allocation information for memcheck precise errors in global memory.
  • Added warning when calling device-side malloc with an empty size.
  • Added separate public API device callback for cuda::memcpy_async.
  • Added new command-line option --num-cuda-barriers to override the expected number of cuda::barrier used by the target application.
  • Added new command-line options --print-session-details to print session information and --save-session-details to save it to the output file.
  • Added support for WSL2.

Updates in 2021.2.1

  • Added device backtrace for malloc/free errors in CUDA kernels.
  • Improved racecheck host memory footprint.

Updates in 2021.2

  • Added racecheck and synccheck support for cuda::barrier on Ampere GPUs or newer.
  • Added racecheck support for __syncwarp with partial mask.
  • Added --launch-count and --launch-skip filtering options. See the Command Line Options documentation for more information.
  • --filter and --exclude options have been respectively renamed to --kernel-regex and --kernel-regex-exclude.
  • Added support for QNX and Linux aarch64 platforms.
  • Added support for CUDA graphs memory nodes.

Updates in 2021.1.1

  • Fixed an issue where incorrect line numbers could be shown in errors reports.

Updates in 2021.1

Updates in 2020.3.1

  • Fixed issue when launching a CUDA graph multiple times.
  • Fixed false positives when using cooperative groups synchronization primitives with initcheck and synccheck.

Updates in 2020.3

  • Added support for CUDA memory pools and CUDA API reduced serialization.
  • Added host backtrace for unused memory reports.

Updates in 2020.2.1

  • Fixed crash when loading cubins of size larger than 2 GiB.
  • Fixed error detection on systems with multiple GPUs.
  • Fixed issue when using CUDA Virtual Memory Management API cuMemSetAccess to remove access to a subset of devices on a system with multiple GPUs.
  • Added public API to translate between sanitizer and CUDA stream handles.

Updates in 2020.2

  • Added support for CUDA graphs and CUDA memmap APIs.
  • The memory access callback of the public API has been split into three distinct callbacks corresponding to global, shared and local memory accesses.

Updates in 2020.1.2

  • Added sanitizer stream API. This fixes tool crashes when per-thread streams are being used.

Updates in 2020.1.1

  • Added support for Windows Hardware-accelerated GPU scheduling
  • Added support for tracking child processes spawned by the application launched under the tool via the --target-processes CLI option.

Updates in 2020.1

  • Initial release of the Compute Sanitizer (with CUDA 11.0)

Updates to the Sanitizer API :

  • Added support for per-thread streams
  • Added APIs to retrieve the PC and size of a CUDA function or patch
  • Added callback for cudaStreamAttachMemAsync
  • Added direction to memcpy callback data
  • Added stream to memcpy and memset callbacks data
  • Added launch callback after syscall setup
  • Added visibility field to allocation callback data
  • Added PC argument to block entry callback
  • Added incoming value to memory access callbacks
  • Added threadCount to barrier callbacks
  • Added cooperative group flags for barrier and function callbacks

Updates in 2019.1

  • Initial release of the Compute Sanitizer API (with CUDA 10.1)

2. Known Limitations

  • Applications run much slower under the Compute Sanitizer tools. This may cause some kernel launches to fail with a launch timeout error when running with the Compute Sanitizer enabled.
  • Compute Sanitizer tools do not support device backtrace on Maxwell devices (SM 5.x).
  • Compute Sanitizer tools do not support device backtrace on Windows Server 2016 for devices in WDDM mode.
  • Compute Sanitizer tools do not support device backtrace and coredumps on WSL2.
  • Compute Sanitizer tools do not support CUDA/Direct3D interop.
  • Compute Sanitizer tools do not support CUDA/Vulkan interop.
  • The memcheck tool does not support CUDA API error checking for API calls made on the GPU using dynamic parallelism.
  • The racecheck, synccheck and initcheck tools do not support CUDA dynamic parallelism.
  • CUDA dynamic parallelism is not supported when Windows Hardware-accelerated GPU scheduling is enabled.
  • Compute Sanitizer tools cannot interoperate with other CUDA developer tools. This includes CUDA coredumps which are automatically disabled by the Compute Sanitizer. They can be enabled instead by using the --generate-coredump option.
  • Compute Sanitizer tools do not support IPC memory pools. Using it will result in false positives.
  • Compute Sanitizer tools are not supported when SLI is enabled.
  • The racecheck and synccheck tools do not support distributed shared memory.

3. Known Issues

  • The racecheck tool may print incorrect data for "Current value" when reporting a hazard on a shared memory location where the last access was an atomic operation. This can also impact the severity of this hazard.
  • With some versions of Windows Server 2016, programs built with some configurations might hang when used with the Compute Sanitizer. A workaround for this issue is to use the Computer Sanitizer with --show-backtrace device or --show-backtrace no options.
  • On QNX, when using the --target-processes all option, analyzing shell scripts may hang after the script has completed. End the application using Ctrl-C on the command line in that case.
  • The initcheck tool might report false positives for device-to-host cudaMemcpy operations on padded structs that were initialized by a CUDA kernel. The #pragma pack directive can be used to disable the padding as a workaround.

4. Support

Information on supported platforms and GPUs.

4.1. Platform Support

Table 1. Platforms supported by Compute Sanitizer
Platform Support
Windows Yes
Linux (x86_64) Yes
Linux (ppc64le) Yes
Linux (aarch64sbsa) Yes
Linux (aarch64) Yes

4.2. GPU Support

Table 2. GPU architectures supported by Compute Sanitizer
Architecture Support
Kepler No
Maxwell Yes
Pascal Yes
Volta Yes
Turing Yes
Ampere Yes




Information furnished is believed to be accurate and reliable. However, NVIDIA Corporation assumes no responsibility for the consequences of use of such information or for any infringement of patents or other rights of third parties that may result from its use. No license is granted by implication of otherwise under any patent rights of NVIDIA Corporation. Specifications mentioned in this publication are subject to change without notice. This publication supersedes and replaces all other information previously supplied. NVIDIA Corporation products are not authorized as critical components in life support devices or systems without express written approval of NVIDIA Corporation.


NVIDIA and the NVIDIA logo are trademarks or registered trademarks of NVIDIA Corporation in the U.S. and other countries. Other company and product names may be trademarks of the respective companies with which they are associated.