Does it ever get easier?

I thought using the network in the web would be a piece of cake. I was wrong.

After fixing some minor issues I wrote some quick code to test training. It worked… technically. It was really slow and froze my whole pc. A quick glance into my system monitor revealed the secret. All CPU cores were almost maxed out while the GPU sat idle. So after spending a bunch of time to use the GPU everything was back to the CPU again.

WebGPU is still experimental. Especially on Linux which is why instead of throwing an error when it couldn’t access my graphics card it just used all my CPU cores as a fallback. To get the GPU back into the game I started my browser with some flags and compatibility layers which makes training still slower than on native hardware but faster than with the CPU.

Hopefully this was the last major challenge and implementing the frontend will be easier.