[madmaze]

This is quite interesting, though keep in mind that the RPi is talking to another machine to do its GPU computations. There are actually combination Nvidia GPU & ARM board like SECO's CARMA DevKit (http://shop.seco.com/carma-devkit.html) Which will actually run CUDA locally with out virtualization

[tinco]

So.. it's just a publicity stunt? If it's not actually using the graphics hardware on-chip then there's nothing special about it that's relevant to the raspberry pi. Apparently there exists software that can give you virtual gpu's, that's cool enough news for me, a bit weird that they market it through the pi.

[madmaze]

its likely for publicity, especially with Nvidia's GTC coming up next week

[giggle_liao]

But it's too expensive. And as far as I know, the GPU virtualization breaks the barrier of the number of GPU cards inserted on the board. That is, if the library allows, we can dispatch the workload to as many as GPUs as possible.

[madmaze]

True, but why not just run it on the host of the GPUs then? I guess the advantage would be on-demand GPUs.