Job Planner

Describe the job. Get the order.

Say what you want to run. The planner sizes it: memory, compute, the GPU classes that fit, the hours, and the cost at live open-market prices. One click turns the result into a sealed order.

Model size, billions of parameters

Training tokens, millions (dataset × epochs)

Memory needed

23 GB

Compute

1.6 × 10^18

floating-point operations

Cheapest fit

$1.57

1 × RTX 3090 for 9 h

classGPUswork timeordermedian pricecost

RTX 3090cheapest1 × 24 GB8.9 h9 GPU·h$0.175$1.57Seal this order

RTX 40901 × 24 GB3.8 h4 GPU·h$0.433$1.73Seal this order

RTX 50901 × 32 GB3.0 h4 GPU·h$0.576$2.31Seal this order

L40S1 × 48 GB3.5 h4 GPU·h$0.694$2.77Seal this order

H100 SXM1 × 80 GB1.3 h2 GPU·h$2.002$4.00Seal this order

A100 SXM41 × 80 GB4.1 h5 GPU·h$0.802$4.01Seal this order

H100 PCIe1 × 80 GB1.7 h2 GPU·h$2.590$5.18Seal this order

B200fastest1 × 192 GB34 min1 GPU·h$7.813$7.81Seal this order

H2001 × 141 GB1.3 h2 GPU·h$5.119$10.24Seal this order

A sizing estimate, not a benchmark. Compute is 2·N·T operations for inference, about 4·N·T with adapters and 6·N·T for full training; each class is counted at 35 % of its peak throughput. Prices are the live open-market median per GPU·h. Holders see the cheapest verified offers on the Desk.