GPU jobs bill the CPU, memory and egress they use
A GPU job's price now counts the CPU and memory the machine actually used and the bytes it sent, so jobs that write big outputs cost more than before.
A GPU job (compute: "gpu") is now priced on what its machine used. Before, the price counted the GPU seconds and a fixed minimum of CPU and memory, and nothing for the bytes the machine sent out.
Each GPU job now bills three more things at what the run measured:
- CPU, for every second the machine used more than its minimum.
- Memory, the same way.
- Network egress, the bytes the machine sent, which for most jobs is the output on its way to storage.
The first job on a machine that just started also pays for the machine starting up and for its idle time before it shuts down, about 45 seconds of GPU time. A job that lands on a machine already running does not.
A job with a small output costs about what it did. A job that writes several gigabytes pays mostly for sending them, so it costs noticeably more. Each line shows on the job’s cost breakdown with its quantity.
CPU jobs are unchanged. Pricing has the per-plan rates.