Runway Characters: Runway’s Real-Time Video Avatars Now Run on Four H100 GPUs Instead of Eight

A live, responding video avatar can never fall behind. That is the constraint Runway describes in a research post published on November 5, alongside a technical paper: how to serve Characters, its conversational personas generated from a single image, using an approximately 57-billion-parameter video model.

A Clock That Never Stops

Traditional video generation works at the model’s own pace: you enter a prompt, you wait, and you get a clip. In contrast, a conversational character must keep up with the person speaking. At 24 frames per second, the model must generate a four-frame block at 704 x 1088 every 166.7 ms, for the entire duration of the exchange. Each frame carries its share of audio: a delayed block freezes the face and cuts off the voice at the same time.

Furthermore, it is impossible to batch computations across users as most AI services do: each session is a private, one-on-one stream. All performance gains therefore had to come from within each compute step.

Cutting GPUs per Session in Half

At launch, Characters required eight H100s per session. On four, the original software took 240.9 ms per step—far too slow. Runway discovered that 29% of GPU time was spent reprocessing a cache that barely changed; by doing this work only once, as each frame is written, cache management dropped from 69.7 ms to 6.2 ms per step. A kernel fusing computation and inter-GPU communication, alongside a dedicated GPU for frame decoding, rounded out the architecture. The result: 159.3 ms per step, or 24.5 frames per second on four H100s, and 147.4 ms with the fused kernel. The system also runs on three B200 GPUs.

To ensure quality did not degrade, Runway compared lip synchronization across 51 clips: discrepancies remained on the same order as variations between two identical runs. The post also recounts outages: memory leaks following a software update, and a simple monitoring call that lengthened a step from 168.5 to 278 ms.

The Real Bottleneck: Voice

Measured across 51 conversational turns, the median latency between the end of the user’s sentence and the character’s response was 2.13 s. Runway’s video pipeline accounts for just 0.49 s of that; the voice agent takes 1.42 s. Between humans, Runway notes, this conversational gap averages around 208 ms.