# Runway — 60x Faster Cold Starts: Treating Peer GPUs as Weight Servers

- Company: Runway (runway.com)
- Announced: 2026-05-04
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://runway.com/news/engineering/60x-faster-cold-starts-treating-peer-gpus-as-weight-servers
- Record: https://forck.live/items/7886-60x-faster-cold-starts-treating-peer-gpus-as-weight-servers
- Subject: Gen / Aleph

Runway built a system called NCCLBack that reduces model cold-start times from minutes to seconds by broadcasting weights between GPUs instead of downloading them from storage.

## Evidence

Verbatim from https://runway.com/news/engineering/60x-faster-cold-starts-treating-peer-gpus-as-weight-servers:

> How we cut model cold-start times from minutes to seconds by broadcasting weights between GPUs instead of downloading them from storage, and the sharp edges we found along the way.

---

Record: https://forck.live/items/7886-60x-faster-cold-starts-treating-peer-gpus-as-weight-servers
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md