Runway details Characters real-time video serving: ~57B model hits 24 fps on four H100s

On November 5 Runway published how Characters serves a roughly 57-billion-parameter autoregressive video diffusion transformer in live conversation at 24 frames per second—a new four-frame chunk every 166.7 ms. Launch used eight H100s per session; on four cards the original software took 240.9 ms per diffusion step. Non-fused work such as doing cache upkeep once at write time cut that to 159.3 ms (about 24.5 fps sustained), enough for real time on four H100s. Fusing the tensor-parallel all-reduce into the FP8 matrix multiply then reached 147.4 ms. Over seven weeks of telemetry, 99.5% of sessions that streamed at least a second averaged above 23.5 fps.

Written by dalao.ai: the title, summary and why it matters are our own words. The original reports are the record; their publishers keep authorship.