An NVIDIA engineer is on revision seven of a patch that runs under one hundred lines.
Andrea Righi has been working the Linux scheduler problem on the Vera CPU and its Olympus cores.
The patch has gone through seven revisions on its way toward the mainline kernel, and revision seven is the one carrying real numbers.
The change teaches the scheduler to prefer the right SMT sibling threads, honors asymmetric SMT priority when idle threads are selected, and adds a knob, sched_smt_asym_packing, for firmware that cannot communicate the preference on its own.
OpenBLAS throughput moved from 7.12 to 7.35 TFLOP/s, up 3.20 percent.
NVPL went from 9.65 to 10.30 TFLOP/s, up 6.73 percent.
Standard deviation dropped in both cases, so workloads settled on the lower numbered sibling consistently instead of drifting between them.
That last detail is the one teams renting compute should sit with.
This gain did not come from a purchase order.
It came from someone willing to send the same patch seven times, and from the plain fact that the compute you already pay for is the cheapest compute you will ever get.
RTX 3090s at $0.15/hr, with Buy Now Pay Later.
Zero upfront, payment after one month.
Which small optimization last saved you from buying more hardware?
https://ai.bhkcloud.com/?utm_source=linkedin&utm_medium=social&utm_campaign=bhk_social_2026w40&utm_content=evening
Originally posted on linkedin