8 video VLMs benchmarked on a single RTX 3090: TTFT and throughput compared

SkyLordOmega · reddit · 2026-09-19

As part of a NIST/TREC visual QA submission, the author measured inference speed of 8 video VLMs of varying parameter sizes and architectures on a single RTX 3090, covering two tasks: returning ten ranked textual answers per video query, and ranking candidate answers by likelihood.

Results are given in two tables plus an environment photo. Caveats the author stresses:

The author notes this is an anecdotal reference rather than a rigorous benchmark.

Original post →

More from Infra

Infra channel →