User visibility on the RAS Engine for ECC MBU dumping and proactive logging

Regarding the RAS Engine introduced in the Blackwell architecture, please clarify:

  • User Accessibility: Is the RAS Engine accessible to user-space software or drivers?

  • Error Dumping: Can we use the RAS Engine to dump hardware error records (specifically ECC multi-bit upsets) during debugging?

  • Proactive Reporting: Can the engine be configured to proactively print/report error messages when a fault occurs?

Are there any APIs or documentation available for these use cases?

RAS contains several features. In regards to “RAS Repair – GPU Memory”, which was introduced with Blackwell, the following statistics are user accessible:

Beside nvidia-smi also have a look at the options of SMBPBI

(I only got a “bad” link to it, be careful - or wait for official replies: https://max.book118.com/html/2024/0123/8031120105006030.shtm)