[adastra22]

Which means that inference would be approximately the same speed (but compute offloaded) as the suggested CPU inference engine.