2 ms·

Also could run on a more generic cloud inference or gpu site. At least to see how well it works for your use-case before spending on hardware.