I-OpenAI isebenzise i-GPT-5.6 ukuthuthukisa ukusebenza kwayo kokucabanga, yehlisa izindleko zesevisi ekupheleni kuye ekupheleni ngo-20%

Inkampani i-OpenAI ishicilele umbhalo izolo (29 Julayi), yamemezela ukuthi isebenzisa imodeli ye-GPT-5.6 Sol ukuthuthukisa uchungechunge lokucabanga lwe-AI lwamamodeli ochungechunge lwe-GPT-5.6 ngokwawo, kunciphise izindleko zesevisi ekupheleni kuye ekupheleni ngo-20% kakhulu.

I-OpenAI yethule uchungechunge lwayo lwamamodeli e-AI anamandla kakhulu, i-GPT-5.6, mhla ka-9 Julayi. Ekuqhubekeni kokuthuthukiswa kwesakhiwo, ithimba elisemthethweni lasebenzisa imodeli ye-GPT-5.6 Sol ukuthuthukisa ukusebenza kwayo, ukunciphisa izindleko zokubala nokwandisa ukusebenza kahle kokusetshenziswa kwama-GPU.

I-OpenAI ithe uhlelo lusabalalisa izicelo ngokwesifunda, umthamo otholakalayo nohlobo lwe-accelerator; ngaphakathi kweqoqo, imisebenzi isatshalaliswa futhi kubhekwa umthwalo, ubude bomongo nokutholakala kwe-cache. I-GPT‑5.6 Sol isebenzisa i-Codex ukuhlaziya ithrafikhi yokukhiqiza, ukuhlonza imithombo yokungalingani komthwalo, ukuhlola amasu okuqondisa nokulungisa imithetho ye-heuristic.

Esigabeni sokudlulisa phambili kwemodeli, i-GPT‑5.6 Sol ihlonza umsebenzi ongabalwa kusenesikhathi, ongagwenywa noma owenziwa ngokuhambisana, bese isebenzisa i-Codex ukuze izibhalele kabusha futhi ithuthukise ama-kernel e-GPU endaweni yokukhiqiza. Isithombe-skrini esihlobene siboniswe ngezansi:

Iqembu laf undisa imodeli ye-GPT-5.6 Sol izilimi ezimbili zokuhlela, i-Triton ne-Gluon, ukuze imodeli ikwazi ukuthuthukisa ama-kernel enjini yokucabanga. Lokhu kwenze izindleko zesevisi ekupheleni kuye ekupheleni ezibangelwa ama-kernel e-GPU endaweni yokukhiqiza zehle ngo-20% kakhulu.

I-OpenAI iphinde yasebenzisa ithuluzi lomthombo ovulekile i-FpSan (isihlanzi sezinombolo zamaphoyinti antantayo) ukuze iqinisekise ukunemba kwama-kernel abhalwe yi-GPT‑5.6 Sol.

I-OpenAI iphinde yenza i-GPT-5.6 Sol ithuthukise i-speculative decoding. Lobu buchwepheshe bokusheshisa ukucabanga busebenzisa imodeli encane esalungisa uhlaka ukuze ibikezele i-token elandelayo, bese imodeli eyinhloko iqinisekisa noma ilungise umphumela obikezelwe. Eposini lomphakathi langomhla ka-29 Julayi 2026, i-OpenAI ithe ukusebenza kahle kokukhiqizwa kwama-token kwenyuke ngaphezu kuka-15% ngemva kokuthuthukiswa.