Question about training time / computational cost

[HJkim0726]

Hello, and thank you for sharing this impressive work.

I really enjoyed reading your paper, and I was especially impressed by the idea and the strong performance of SparseDriveV2. The overall design and results are truly remarkable.

I was wondering if you could share a bit more information about the training time or overall computational cost. In the paper, the training configuration such as GPU count, batch size, and number of epochs is described, but I could not find details about the actual wall-clock training time.

If possible, could you let me know approximately:

• how long training took for the NAVSIM setting,
• how long training took for the Bench2Drive setting, and
• whether there were any particularly expensive parts of training in practice. (evaluating pdm score at training time, maybe)

Thank you again for your excellent work.

Best regards

[swc-17]

@HJkim0726
Thanks for your recognition!

1. Using 8 L20 GPUs, it takes about 15h training on NAVSIMv1 and about 10h on NAVSIMv2.
2. For Bench2Drive, we first train a perception stage1 for 20 epochs, and then train stage2 with planning module on 16 L20 GPUs for 10 epochs, which takes about 8h.
3. You are right, training on NAVSIM is heavily slowed by online PDMS computation, so we only use a small number of trajectories for metric supervision (20x20 for v1 and 20x10 for v2), but still enough to get a good result.