Improved
GPU jobs start without waiting for a free L4
A GPU ffmpeg job no longer waits when no L4 is free. It starts on an L40S or an RTX PRO 6000 instead, billed per second for the GPU it ran on.
API
An ffmpeg job with compute: "gpu" ran on an NVIDIA L4 only, so when every L4 was busy the job waited for one. One job this week waited more than seven minutes before it started.
GPU jobs now start on an L4 when one is free, and on an L40S or an RTX PRO 6000 when none is. All three encode H.264, HEVC and AV1 in hardware, so a command that runs on an L4 runs the same on either. We checked h264_nvenc, hevc_nvenc and av1_nvenc on each before turning this on.
- A job is billed per second for the GPU it actually ran on. An L40S or an RTX PRO 6000 costs more per second than an L4, and you only pay that rate on the jobs that needed it.
- Nothing changes in your requests.
compute: "gpu"andcompute: "auto"work as before.