I-OpenAI ibika ukuthi i-chip ye-Jalapeño yangokwezifiso ithuthukisa isivinini sokuqanjwa kwe-AI nokusebenza kahle

Kwenzekeni

I-OpenAI ishicilele imiphumela yokuhlola yokuqala ye-Jalapeño, i-chip yayo yokuqala yangokwezifiso. Le nkampani ithi uhlelo lulethe izikhathi ezi-1.5 kuye kwezingu-1.9 ngaphezulu kwe-AI emsebenzini nge-watt ngayinye ekuphumeni okuphezulu kanye nezikhathi ezi-1.7 kuya kwezingu-3.6 eziphansi zokuphela ukuya ekupheleni kunezinhlelo zokuqhathanisa kuyo yonke i-GPT-OSS 120B, DeepSeek R1 670B kanye ne-Kimi K2.51

I-OpenAI ithi i-Jalapeño iyi-chip yayo yokuqala yangokwezifiso nokuthi imiphumela isuselwe ekuhloleni i-chip kanye nesistimu eyakhelwe kuyo. Inkampani ihlole uhlelo ku-InferenceX, ibhentshimakhi yomphakathi evela ku-SemiAnalysis ekala inqubo ephelele yokunikeza isicelo se-AI. I-OpenAI ithi iqhathanise i-Jalapeño namasistimu e-AI atholakalayo kwezentengiso ngaphezu kwamaphoyinti okusebenza asukela kokusebenza okuphezulu kuya ekusetshenzisweni okusebenzisana kakhulu, ukubambezeleka okuphansi.

Ku-GPT-OSS 120B yonkana, DeepSeek R1 kanye ne-Kimi K2.5, i-OpenAI ibika izikhathi ezingu-1.5 kuya kwezingu-1.9 ngaphezulu kwe-AI ngomsebenzi ngamunye nge-watt ngayinye ekuphumeni okuphezulu kanye nezikhathi ezi-1.7 kuye kwezingu-3.6 zokubambezeleka okuphansi kokuphela kuya kokuphela kunezinhlelo zokuqhathanisa. Ngomthwalo osebenzayo osebenzisana kakhulu, ibika ukusebenza okuphezulu okuphindwe izikhathi ezingu-2.1 kuye kwezingu-4.1. Esithasiselweni, inkampani iqhathanisa i-Jalapeño ne-GB200 ye-GPT-OSS 120B kanye ne-GB300 ye-DeepSeek R1 kanye ne-Kimi K2.5. I-OpenAI ithi i-Jalapeño ifinyelele umngcele we-Pareto kuwo wonke amaphoyinti okusebenza ahloliwe, okusho ukuthi ithole inhlanganisela engcono yokuphuma, ukubambezeleka nokusebenza kahle kwamandla phakathi kwalezo ziqhathaniso.

Le nkampani ibika ukuthi i-Jalapeño inesilinganiso samandla wephakheji lama-watt angu-700, kuyilapho amandla aqinile akaliwe ahlala noma ngaphansi kwama-watts angu-550 emisebenzini ehloliwe. Iziqhathaniso ezishicilelwe zisebenzisa izilinganiso zamandla ephakheji, okuhlanganisa ama-watts angu-1,200 we-GB200 kanye nama-watts angu-1,400 we-GB300. Ku-GPT-OSS 120B, i-OpenAI ibika cishe izikhathi ezingu-1.9 ukuphakama okuphezulu kwamathokheni axutshwe ngesekhondi nge-kilowatt ngayinye kanye nokubambezeleka okuphansi kokuphela kuya ekupheleni izikhathi ezingu-1.7 kunesistimu yokuqhathanisa. Ku-DeepSeek R1, ibika cishe izikhathi ezingu-1.7 ukusebenza okuphezulu okuphezulu kwe-watt ngayinye kanye nokubambezeleka okuphansi okuphindwe izikhathi ezingu-3.6. Ku-Kimi K2.5, ibika cishe izikhathi ezingu-1.5 ukusebenza okuphezulu okuphezulu kwe-watt ngayinye kanye nokubambezeleka okuphansi okuphindwe izikhathi ezingu-3.4.

I-OpenAI ihlanganisa imiphumela kumklamo ogcwele isitaki ohlanganisa i-chip, inkumbulo, inethiwekhi, isofthiwe kanye nesistimu yesikali esilinganisa umthwalo womsebenzi wamamodeli wolimi. Le nkampani ithi izakhiwo zigcina isimo semodeli, kufaka phakathi inqolobane ye-KV esetshenziswa ngesikhathi sokukhiqiza, indawo lapho kungenzeka khona futhi kunciphisa ukubambezeleka kokuxhumana phakathi kwama-cores nama-chips. Ichaza ukugcwalisa kuqala njengokudinga kakhulu ikhompuyutha kanye nokuqopha njengokuvinjelwa umkhawulokudonsa wenkumbulo, futhi ithi i-Jalapeño yakhelwe ukuphatha izigaba zombili. I-OpenAI iphinda ithi i-AI isize ukuhambisa i-chip isuka ekwakhiweni kokuqala iye ku-tapeout ezinyangeni eziyisishiyagalolunye futhi yasiza ekuthuthukiseni amasekhethi e-arithmetic, kodwa umthombo awunikezi i-akhawunti ezimele yalowo mugqa wesikhathi noma ukuhlukaniswa okugcwele kwawo umsebenzi wokuklama owenziwe ngokuzenzakalelayo.

Imininingwane yomthombo: openai.com ↗

Kungani kubalulekile

Imiphumela iphakamisa ukuthi ukusebenza okucatshangelwayo kungase kuncike kakhulu ekuklameni amamodeli, ama-chips, inkumbulo, inethiwekhi nesofthiwe njengesistimu eyodwa. Uma izinzuzo zibambelela esikalini sokukhiqiza, zinganciphisa amandla nehadiwe edingekayo ukuze kunikezwe izimpendulo ze-AI futhi zenze imithwalo yemisebenzi yama-ejenti esebenzisanayo iphendule kakhudlwana.

Ukubaluleka okumaphakathi wushintsho olungase lube khona endleleni ingqalasizinda ye-AI ethuthukiswa ngayo. I-akhawunti ye-OpenAI ithi isisheshisi senhloso evamile singalahlekelwa ukusebenza lapho ukubala, ukufinyelela inkumbulo nokuxhumana kuthathwa njengezinkinga ezihlukene. I-Jalapeño esikhundleni salokho ihlanganisa lezo zakhi ezizungeze isikhathi namaphethini okunyakaza kwedatha okucatshangwayo kwemodeli yolimi. Leyo ndlela ibalulekile ngoba ukunikeza impendulo kuhilela kokubili ukucubungula ukwaziswa nokukhiqiza amathokheni ngokulandelana, kuyilapho amasistimu e-ejenti angase aphinde lezo zinyathelo izikhathi eziningi emsebenzini owodwa.

Izinzuzo zokubambezeleka okubikiwe zingaba nemiphumela engokoqobo emikhiqizweni ye-AI esebenzisanayo uma isinda ekusetshenzisweni kokukhiqiza. I-OpenAI ithi ukucatshangelwa okusheshayo kungasekela ama-ejenti asabelayo futhi kwenze kube nokwenzeka ukuqhuba imithwalo yemisebenzi eshesha kakhulu ngamakhono ngaphambilini ahlotshaniswa nezindlela eziqoqwe ngokushesha noma eziningi kakhulu. Umthombo awubonisi leyo miphumela kumkhiqizo wekhasimende obukhoma, nokho. Iveza imiphumela yesilinganiso kanye nokuthunyelwa kwengqalasizinda okuhleliwe, ngakho umthelela womphakathi usalokhu ungenzeka.

Ukusebenza kahle kwamandla nakho kunokubaluleka kokusebenza. I-OpenAI ithi ukukhiqiza umsebenzi owusizo kakhudlwana ngamandla afanayo kanye nezingxenyekazi zekhompuyutha kungase kukusize ukufeza isidingo ngezindleko eziphansi futhi kuthuthukise amandla okusebenza. Leso simangalo sisobala njengenhloso yebhizinisi, kodwa umthombo awunikezi izindleko zombuzo ngamunye, izindleko zesistimu iyonke, izimfuneko zokupholisa, izindleko ezinkulu, ukuqagela kokusetshenziswa noma ubufakazi bokuthi izilinganiso zebhentshimakhi ezibikiwe zihumusha ngokuqondile amanani aphansi kubasebenzisi.

Inqubo yokuthuthukisa echazwe yi-OpenAI nayo iyaphawuleka. Le nkampani ithi amamodeli ayo asize ukuklama nokukhulisa i-Jalapeño, kuyilapho amamodeli amasha asetshenziswa ukuyithuthukisa nokuyihlela. Lisebenzisa i-Codex ne-GPT-Astra, ithimba kubikwa ukuthi lilethe amamodeli amathathu esisindo esivulekile abengeyona ingxenye yohlelo lokukhiqiza lwangempela ekusebenzeni okuphezulu phakathi nezinyanga ezimbili. I-OpenAI futhi ithi ukuqaliswa okukhiqizwa yi-AI kwakushesha izikhathi ezingu-1.5 kuya kwezingu-1.8 kunokuqaliswa okubhalwe ngumuntu kokunaka okukhethiwe kwe-GPT-OSS kanye namabhulokhi axubile ochwepheshe. Lezo zibalo zisebenza kumabhulokhi akhethiwe kuphela, hhayi amamodeli aphelele, futhi aziqinisekisi ukuthi i-AI ingaklama ngokuzimela noma ilungiselele yonke i-chip yokukhiqiza.

Ongakubuka ngokulandelayo

I-OpenAI ihlela ukuqala ukusebenzisa i-Jalapeño kwingqalasizinda yayo yekhompyutha ekupheleni kuka-2026, kodwa ithi iziqu zokukhiqiza, ukuvuthwa kwesofthiwe, ukusebenza kwezikali kanye nokuqinisekiswa kwemodeli ebanzi kusaqhubeka. Umthombo awusunguli ukuqinisekiswa okuzimele, izindleko zokusebenza, umthamo wokukhiqiza noma ukuthi i-chip izosebenza kanjani kuwo wonke umthwalo wokusebenza ngale kwaleyo ehloliwe.

Ukuhlola kokuqala ukuthi ingabe i-Jalapeño iyafinyelela kungqophamlando ehleliwe yokusatshalaliswa. I-OpenAI ithi ihlose ukuqala ukusebenzisa i-chip ngaphakathi kwengqalasizinda yayo yekhompyutha ekupheleni kuka-2026, kuyilapho ukufaneleka kokukhiqiza nomsebenzi we-software uqhubeka. Ukuthunyelwa kuzoqinisekisa ukuthi isistimu idlulele ngale kokuboniswa kwebhentshimakhi, kodwa umthombo awusicacisi isikali sokuqala, izindawo, umthwalo wokusebenza noma ukuthi abasebenzisi bazolubona yini ushintsho lwezinga lomkhiqizo.

Ubufakazi obengeziwe buyadingeka ekukhiqizeni kabusha kanye nendlela yokuqhathanisa. I-OpenAI isebenzise ibhentshimakhi esesidlangalaleni, kodwa umthombo awunikezi idatha yebhentshimakhi ephelele, inani lama-chips noma amasistimu ahloliwe, izinguqulo zesofthiwe, amazinga okusetshenziswa, amandla okupholisa kanye nendawo, noma ukuqinisekiswa okuzimele kwemiphumela yokuqhathanisa. Ngenxa yokuthi izilinganiso zezihloko zenziwa ngendlela evamile kusetshenziswa izilinganiso zamandla e-chip eshicilelwe kuyilapho amandla aqhubekayo e-Jalapeño nawo abikwa ngokuhlukana, abafundi kufanele bahlukanise ukusebenza kahle kwezinga lephakheji ekusebenzeni kahle okuphelele kwesikhungo sedatha.

Ukufakwa kwamamodeli kuzoba okubalulekile. I-OpenAI ibika imiphumela ye-GPT-OSS 120B, DeepSeek R1 670B kanye ne-Kimi K2.5 1T, futhi ithi ukuhlola kwangaphakathi kubonise inzuzo ebanzi kumamodeli we-OpenAI asemngceleni. Umthombo awuwahlonzi lawo mamodeli angaphakathi noma ushicilele imiphumela yawo. Iphinde ithi umndeni ngamunye wemodeli eyengeziwe udinga ama-kernels amasha kanye nokwenza kahle kwemodeli ethile, okusho ukuthi ukuguquguquka kwe-chip kungase kuncike ekuqhubekeni komsebenzi wesofthiwe kune-hardware yodwa.

I-OpenAI ithi i-Gen 2 ijulile ngentuthuko futhi i-Gen 3 iyathuthuka, kuyilapho igcizelela ukuthi izoqhubeka nokukhipha ama-accelerator avela ku-NVIDIA nabanye ozakwethu ukuze bathole ukuqeqeshwa nokuqondiswa. Ngakho-ke, okubalulekile okungaziwa akukhona nje ukuthi i-Jalapeño iyashesha yini ekuhlolweni okubikiwe, kodwa ukuthi ingena kanjani ohlelweni lwengqalasizinda exubile, ukuthi izomela umthamo ongakanani, nokuthi izizukulwane ezizayo ziyayigcina yini ibhalansi phakathi kokusebenza, ukubambezeleka, ukusebenza kahle nokusekelwa ekushintsheni kwezakhiwo zamamodeli.

Imihlahlandlela ehlobene nemibuzo