3 comments

[ 2.1 ms ] story [ 18.5 ms ] thread
They implemented a kernel driver that processes can submit work items to (multiply this vector, summing reduce et ). It balances the computation across available hardware.

Seems really neat, if not a bit click-baity.

Paper: https://dl.acm.org/doi/pdf/10.1145/3613424.3614285

Code: https://github.com/escalab/SHMT

The work would have to be large enough to be worth the cost of kernel context switch IIUC.