KINGFISHER: A Framework for Fast Machine Learning Inference for Autonomous Accelerator Systems

Mod­ern par­ti­cle ac­cel­er­a­tor fa­cil­i­ties allow new and ex­cit­ing beam prop­er­ties and op­er­a­tion modes. Tra­di­tional real-time con­trol sys­tems, al­beit pow­er­ful, have band­width and la­tency con­straints that limit the range of op­er­at­ing con­di­tions cur­rently made avail­able to users. The ca­pa­bil­ity of Re­in­force­ment Learn­ing to per­form self-learn­ing con­trol poli­cies by in­ter­act­ing with the ac­cel­er­a­tor is in­trigu­ing. The ex­treme dy­namic con­di­tions re­quire fast real-time feed­back through­out the whole con­trol loop from the di­ag­nos­tic, with novel and in­tel­li­gent de­tec­tor sys­tems, all the way to the in­ter­ac­tion with the ac­cel­er­a­tor com­po­nents. In this con­tri­bu­tion, the novel KING­FISHER frame­work based on the mod­ern Xil­inx Ver­sal de­vices will be pre­sented. Ver­sal com­bines sev­eral com­pu­ta­tional en­gines, specif­i­cally com­bin­ing pow­er­ful FPGA logic with pro­gram­ma­ble AI En­gines in a sin­gle de­vice. Fur­ther­more, this sys­tem can be na­tively in­te­grated with the fastest beam di­ag­nos­tic tools al­ready avail­able, i.e. KAP­TURE and KA­LYPSO.