v0.9

New features:

• Add coefficient of variance to bandwidth output statistics
• Add huge page support for host memory (disabled on Windows)
• Add option to sample pairs in device-to-device tests
• Add troubleshooting guide
• Unify multinode and single-node execution paths

Improvements:

• Improve CUDA architecture detection without requiring GPU access
• Deprecate Volta (sm_70/sm_72) support for CUDA toolkit >=13.0

Bug fixes:

• Fix JSON output aggregation

Platform:

• Skip Boost static libs on Azure Linux

v0.8

Release Notes:

Bug Fixes:

• Device Latency Test Accuracy:
  • Fixed an issue where the device_to_device_latency test was incorrectly reporting host-device latency instead of device-to-device latency.
  • Host-device latency reports now correctly reflect C2C or PCIe latency depending on the system, while device-to-device latency reports focus on NVLINK or equivalent inter-device connections.
• Other general bug fixes

v0.7

Release Notes:

• Support measurements on multinode systems
• Add tests for intra-device memory copies
• Default memcpy buffer size increased to 512MB
• Improved stability of bidirectional bandwidth measurements
• General bug fixes

v0.6

Release Notes:
- Report bidirectional bandwidth in each direction for device to device tests
- Use leading zeroes in PCI BDF reporting
- Bug fixes

nvbandwidth v0.5

v0.5
Release Notes:
• Added 2 New tests to measure CPU <-> GPU and GPU <-> GPU latencies.
• Added a more robust way to detect the GPUs on the system during build phase.
• Added an option to output the test results in JSON for machine parsing.
• Added PCI bus, device details to device listing in test output
• Better Error Handling in the codebase
• General Bug fixes.