Intel's Habana Labs AI Driver With Linux 5.11 Begins Prepping For Future ASICs
Well before Intel bought out AI startup Habana Labs this company with their compelling AI training and inference accelerators has been maintaining an open-source kernel driver in the Linux kernel. That has fortunately continued under Intel's ownership to no surprise and coming up with the Linux 5.11 cycle are more improvements to this accelerator driver for both the Gaudi AI Training and Goya AI Inference products.
The Habana Labs driver updates have been queued as part of the "char-misc" area, which remains where it and similar drivers reside until there is a formal "accelerator" subsystem for the Linux kernel. The changes with this code for Linux 5.11 include:
- Changes in preparing for future, unnamed Habana Labs ASICs. The preparations around the future ASIC support include support for loading multiple types of firmware, command buffer changes, and support to put the PCI MMU page tables on host memory. No new hardware support though is enabled as part of these Linux 5.11 changes. It will be interesting to see what comes next after Gaudi and Goya given how well the hardware is already performing.
- Support for "stream sync" to synchronize between compute and network streams.
- Preparations for upstreaming the Ethernet and RDMA code with Habana Labs hardware.
- A new option so the internal SRAM/DRAM is scrubbed when a user closes a file descriptor to ensure no traces of the work is left in the memory.
- Along with the memory scrubbing on FD closure, there is support for new firmware to provide better device security. These security improvements are good considering that it was just announced Habana Labs' AI hardware will begin rolling out in the Amazon AWS public cloud.
- Support for new firmware with the Gaudi hardware where the firmware now is responsible for device resets.
- Various other fixes and improvements.
This Habana Labs driver work is in char-misc-next until the Linux 5.11 merge window opens up.
Add A Comment