Define the work
Start with independent, equal-sized tasks. State the volume and the baseline time for one task on one GPU.
Give GPU worka clear direction.
Capacity isn'tthe whole picture.
A GPU count tells you what is available, not when a workload will finish. Task duration, usable capacity and the final partial wave all shape the queue. Start with the work, not a headline number.
Make the queuemake sense.
Set the tasks, baseline time, available GPUs and expected utilization. Kernveil arranges equal, independent tasks into waves so you can inspect completion time, queue wait and theoretical throughput.
Read the scheduling model ↗A clearer wayto plan GPU work.
Map capacity
Set the GPUs available and expected utilization. The model makes those assumptions visible.
Inspect the queue
Compare wave count, completion time and the wait before the final wave. Change an input and see it move.
Keep the plan
Export the assumptions and estimate. Reopen them in Capacity Studio without connecting a wallet.
Build toward more
Provider coordination and execution are future research. Today, Kernveil is a planning prototype.