"Mixed precision" with Nvidia T4 GPU

[Nicholas Schafer]

Hi Peter -

Do you have a sense of whether what Nvidia calls "mixed precision" on its new T4 card will be suitable for running in mixed precision mode in OpenMM? https://www.nvidia.com/en-us/data-center/tesla-t4/

If so, it seems like the T4 might have a good price/performance ratio. Any feeling for this, or, better yet, a chance to test it?

Thanks much,
Nick

[Peter Eastman]

No, it's not suitable. What they call "mixed precision" means using half precision as much as possible, and single precision only when required. What OpenMM calls "mixed precision" means using single precision as much as possible, and double only when required. Also, they're referring to the "tensor cores", which only do one very specific operation (matrix multiply).

[Nicholas Schafer]

That's very helpful, Peter, thank you.

Cheers,
Nick