[FuckButtons]

Are you factoring in the above comment about as yet un-implemented parallel speed up in there? For on prem inference without any kind of asic this seems quite a bargain relatively speaking.