SERAPHINE on hold...

...Until I can source better hardware to train her. I can run a fairly competent model on 8gb VRAM, but it is next to impossible to train her on it. Half of her design features hinge on being able to train her semi-regularly.

I am currently looking at one Nvidia Tesla v100 32gb or four Nvidia Tesla p100 16gb with NVLink... Total cost to get either up and running is similar. the V100 is newer and technically faster but smaller memory pool... I can also run the single v100 on a household circuit instead of a HV circuit.

I need thoughts and opinions on how I should pivot. My goal is minimum 32gb VRAM and under $2000 for entire home-datacenter.