# GPU and memory

## `min_vram`

Minimum GPU memory per GPU, in GB. Machines with less are never chosen. This is
the most important setting: set it from what your model actually needs, not
from the largest GPU you know of.

## `gpu`

Pin a GPU model. Leave it out to let RunCompute choose; pinning usually costs
more.

| Model | Memory | Relative speed |
| --- | --- | --- |
| `L4` | 24 GB | 0.35 |
| `RTX4090` | 24 GB | 0.55 |
| `A6000` | 48 GB | 0.50 |
| `A100` | 80 GB | 1.00 |
| `H100` | 80 GB | 1.60 |
| `H200` | 141 GB | 1.90 |

Relative speed is throughput compared with one A100 and is used to turn
`work_units` into hours.

## `gpu_count`

GPUs on one machine, 1 to 8. Hours are divided by the count; the hourly price
is multiplied by it.

## `work_units`

How much compute the job needs, in A100-hours. If a run took 3 hours on one
A100, use `3`. If you only know the time on another GPU, multiply by that GPU's
relative speed: 10 hours on an RTX 4090 is about `5.5`.

The estimate only affects which machine is picked and the quote. Underestimating
does not stop a job early; the budget does.