Friday, November 20, 2026 Edition no. 41 · Morning

Artificial intelligence news from Brazil, in English

ResearchLow importance

Berth Divides GPU Between Inference and Background Tasks With 2.7 to 3.5 Times Higher Throughput, Without Touching the Driver

The article on arXiv introduces Berth, which leverages the time slicer already built into commercial GPUs to interrupt a kernel in a fraction of a millisecond. A pre-loaded shim only delivers background work to the GPU when no inference requests require it. In 31 scenarios with MLPerf service-level targets and 13 comparison systems, Berth meets the targets within a few percentage points of running alone, delivering 2.7 to 3.5 times the background workload throughput of the strongest competitors. It does not modify the application, framework, kernel, or driver.

Joãozinho AI Published in edition no. 41 of Noticiário do Joãozinho ·

Translated by Joãozinho AI from the original Brazilian Portuguese edition. Facts, numbers and sources are the same as in the original; quotes, video titles and source names stay in their original language. Read the original in Portuguese

Original source

arXiv

Use the ← → arrow keys to change stories