Nvidia Spark

[RobertJasiek]

Nvidia Spark uses an ARM CPU, an RTX GPU, Windows, allows all Nvidia libraries including CUDA and TensorRT, and promises to execute all Windows softwares.

1) Can it emulate KataGo for x64-Windows at all?

2) Relatively how fast is this emulation?

3) Should there also be a native KataGo for Nvidia Spark?

Currently Nvidia Spark is overpriced, and designed for LLMs and AI agents. In the forseeable future, this computer architecture might become another valid alternative for deep learning AIs though.

[xela]

Interesting questions, but of little practical importance. Robert, please do not let your quest for optimal KataGo performance take away all the time that you could have spent using KataGo to gain new insights about go theory! The advice you've already shared elsewhere about KataGo setup is valuable and a good enough foundation.

[RobertJasiek]

Oh, I have spent most of the previous three years on using KataGo, gaining new insight and can publish such in a couple of months, I hope.

Currently, I am a happy user of x64 / Nvidia RTX. However, it is unclear whether the RTX line of graphics cards continues as usual except for a delay or whether Nvidia goes all in on ARM / Spark. Such developments will matter some time for a then hopefully much faster computer, presuming native software.

[xela]

Great, I will look forward to seeing what you publish!

I'm sure that commodity GPUs (i.e. not priced in the many thousands of dollars) will continue to be available in some form. If Nvidia go crazy and stop making them available, it just leaves a gap in the market that someone else will fill.