I think the 40x overhead is a case of comparing throughput overhead (from what I know, FHE based secure inference protocols have poor latency, but can process m

I think the 40x overhead is a case of comparing throughput overhead (from what I know, FHE based secure inference protocols have poor latency, but can process many predictions in parallel, improving throughput)