Is there a way to run rapids time series modules (ARIMA, ESM) on multi-gpu?

Question

I have a node with 2 Tesla P100 gpu's on it.

When i run rapids.tsa.ARIMA (or ESM), it will only utilise one of the GPU's. Is there a way to utilise multi-gpu's for training the models? Like as in rapids-dask-xgboost ?

score 2 · Accepted Answer · answered Jul 06 '20 at 20:30

2

Not according to cuml documentation. For multi-node multi-GPU check the column Notes on the table.

answered Jul 06 '20 at 20:30

rpanai

12,515
2
42
64

score 1 · Answer 2 · answered Jul 13 '20 at 16:46

1

A multi node multi gpu implementation of the time series models, Arima and Holtwinters is not present. I would recommend creating a feature request on cuML's github repository (https://github.com/rapidsai/cuml).

answered Jul 13 '20 at 16:46

saloni

296
1
7

Hi @saloni it's actually their goal to have as many multi node multi gpu algorithms as possible. – rpanai Jul 13 '20 at 18:13

Is there a way to run rapids time series modules (ARIMA, ESM) on multi-gpu?

2 Answers2