MoQ, machine-vision codecs, and tensor compression collide
This article highlights recent developments in video technology, including Media over QUIC (MoQ) for low-latency streaming at scale, new video compression approaches for machine vision (VCM, FCM, LCEVC), and the surprising application of video codecs to compress tensors for large language models (VcLLM). It also introduces MediaMolder, an early-stage Go-based framework reimagining FFmpeg, and a new tool, SLC Bitrate Explorer, for contextual codec comparison beyond BD-Rate.
Key Takeaways
- nanocosmos said it is already serving "low hundreds of thousands" of concurrent users with Media over QUIC in production.
- MPEG’s VCM keeps video in pixel format, while FCM compresses neural-network features and claims 90% bitrate reduction versus H.264 and 67% versus VVC.
- V-Nova’s LCEVC testing with Intel showed 30-50% lower decode-and-analytics time, but not the compression gains claimed for FCM.
- The VcLLM paper says H.264 and H.265 GPU encoders can reduce LLaMA-3-70B memory needs by 5.5× and communication bandwidth by 4.5×.
- MediaMolder is a Go-based framework from Tom Vaughan, the founder of x265, and its GitHub repo shows 57 stars and 4 forks.
Why It Matters
The immediate implication is that streaming infrastructure is broadening beyond conventional delivery: MoQ is now in live demos, machine-vision codecs are targeting AI workflows, and video codec hardware is being repurposed for LLM tensor compression. That pulls streaming codec expertise into CDN design, edge delivery, and AI infrastructure planning at the same time. The ecosystem signal is especially strong because Bitmovin, Cloudflare, Oracle, Broadpeak, and nanocosmos all have visible MoQ implementations, while Tom Vaughan is trying to make MediaMolder a more usable orchestration layer around FFmpeg’s libraries. Watch whether MoQ gains browser-side support beyond JavaScript libraries and whether MediaMolder’s GitHub traction moves past 57 stars and 4 forks.
Read full article at linkedin.com