Namuhla kusihlwa (ngomhla ka-25), eHot Chips, i-OpenAI iveze imininingwane eyengeziwe ngohlelo olusha lokuhlela i-Jalapeño, yaphinde yadalula okokuqala imiphumela yokuhlolwa kokusebenza. Ekuhlolweni kwe-InferenceX kwe-SemiAnalysis, i-Jalapeño idlule amaphrosesa okuhlela aqavile akhona emakethe, kokubili ngenani lama-token acutshungulwa umsebenzisi oyedwa kanye nokusebenza nge-kilowatt ngayinye.
URichard Ho, ophethe ihardware kwa-OpenAI, uthe ocingweni lwabezindaba: “Into ebaluleke kakhulu ngemiphumela yokuhlolwa ukuthi i-Jalapeño ithole ukuthuthuka okukhulu kakhulu kokusebenza uma iqhathaniswa nemikhiqizo esezingeni eliphezulu njengamanje. Ingacubungula imithwalo eminingi ye-AI ngamandla kagesi afanayo, iphinde inikeze izimpendulo ngokushesha. I-Jalapeño ifanele kokubili ukusiza abasebenzisi abaningi ngempumelelo nokugcina ukubambezeleka kuphansi kakhulu.”
I-Jalapeño ithuthukiswe ngokubambisana yi-OpenAI ne-Broadcom, futhi amamodeli e-OpenAI nawo abambe iqhaza ocwaningweni nasekuthuthukisweni kwayo. I-OpenAI inethemba lokuthuthukisa i-Jalapeño ibe inkundla yobuchwepheshe engaqhubeka izizukulwane eziningi, ukuze imikhiqizo ye-AI, amamodeli, ama-chip nememori kwakhiwe ngokubambisana kusukela ekuqaleni.
Le chip ikakhulukazi iphethe ukuhlela kwe-AI, okusho ukwenza amamodeli e-AI asevele eqeqeshiwe asebenze ngempela ukuze aqedele imisebenzi noma aqhube ama-agent e-AI.
I-OpenAI isebenzise inkundla ye-InferenceX ukuhlola ukusebenza kokuhlela kwe-Jalapeño, yayiqhathanisa nezinhlelo zama-superchip i-NVIDIA GB200 ne-GB300. I-OpenAI ithi kumamodeli amathathu, okuyi-GPT-OSS 120B, i-DeepSeek R1 ne-Kimi K2.5 1T, inani lomsebenzi we-AI elingaqedwa yi-Jalapeño nge-watt ngayinye liphindwe izikhathi ezingu-1.5 kuya kwezingu-1.9 kunalezo zezinhlelo eziqhathaniswayo, kanti ukubambezeleka kokuphela kuya ekupheleni kungamaphesenti angu-28 kuya kwangu-59 kuphela kokwezinhlelo eziqhathaniswayo.
Ukuthuthukiswa okuhambisanayo kwalo lonke uhlelo kunikeze i-OpenAI ithuba lokuthuthukisa ngokuqondile izigaba ezivame ukwehlisa ijubane kakhulu ngesikhathi sokuhlela. I-Jalapeño igxile ekwehliseni ukubambezeleka esigabeni sokugcwalisa kusengaphambili kanye nasesigabeni sokuxhumana, okuyizigaba ezimbili i-OpenAI ezibona zivame ukuba yizithiyo ezinkulu ekusebenzeni kokuhlela.
Encwadini yebhulogi ememezela imiphumela yokuhlolwa, i-OpenAI ithe i-Jalapeño yaklanywa kusukela ekuqaleni ngenhloso yokunciphisa ukuthuthwa kwedatha nokubambezeleka kwezokuxhumana. Izimo zemodeli, njengendawo yokugcina i-KV cache esetshenziswa lapho kukhiqizwa izimpendulo, zingagcinwa ngokucacile endaweni futhi zigcinwe khona; uhlelo bese lwabela izinsiza ezifanele zokubala, zememori nezokuxhumana kuye ngezigaba ezahlukene zokuhlela.
Ngokolwazi lwe-IT之家, i-OpenAI ihlela ukuqala ngokufaka i-Jalapeño “ngesilinganiso esincane” ngaphambi kokuphela kwalo nyaka, bese yandisa kancane kancane inani lokufakwa ngo-2027, kodwa ayidalulanga ukuthi mangaki ama-chip azosetshenziswa ngonyaka ozayo.
Izikhombo
Umbhalo osemthethweni we-OpenAI: Jalapeño’s first results show industry-leading speed and efficiency in AI inference