Kwenzekeni
I-Firstpost ibika ukuthi i-OpenAI yethule imiphumela yayo yokuqala yebhentshimakhi yomphakathi ye-Jalapeño, i-chip yangokwezifiso ye-AI eyakhiwe nge-Broadcom. I-OpenAI ithi i-chip isebenze kahle kakhulu kunezinhlelo zokuqhathanisa ezisekelwe ku-Nvidia ezisekelwe ku-Blackwell emisebenzini ekhethiwe yemodeli yolimi, okuhlanganisa nokuhlola okubandakanya i-GPT-OSS 120B, DeepSeek R1 670B kanye ne-Kimi K2.5 1T. Inkampani ilindele ukuthunyelwa kwangaphakathi okulinganiselwe ekupheleni kuka-2026 kanye nokukhishwa okubanzi ngo-2027.
I-Firstpost ibika ukuthi i-OpenAI yethule i-Jalapeño ngokuningiliziwe engqungqutheleni ye-Hot Chips futhi yakhipha lokho ekuchaze njengemiphumela yokuqala yebhentshimakhi yomphakathi ye-chip. I-chip yathuthukiswa nge-Broadcom njenge-accelerator yangokwezifiso ye-AI , isigaba lapho imodeli eqeqeshiwe ikhiqiza khona izimpendulo. Umbiko uchaza i-Jalapeño njengengxenye yomzamo wesikhathi eside we-OpenAI wokwakha ingqalasizinda ekhethekile kumamodeli nemikhiqizo yayo.
Ngokusho kwe-Firstpost, i-OpenAI ihlole amamodeli amathathu ku-Jalapeño: GPT-OSS 120B, DeepSeek R1 670B kanye ne-Kimi K2.5 1T. Ukuhlola kusebenzise ibhentshimakhi ye-SemiAnalysis’ InferenceX futhi kwaqhathanisa i-Jalapeño namasistimu asekelwe ekwakhiweni kwe-Blackwell ye-Nvidia. I-Firstpost ihlanganisa imiphumela ku-OpenAI futhi ayibiki ukukhiqizwa kabusha okuzimele, ukucwaningwa kwamabhuku noma ukuhlola okuqhudelanayo kwendlela yokusebenza.
I-Firstpost ithi i-OpenAI ibike ukuthi i-Jalapeño icutshungulwe phakathi kuka-1.5 no-1.9 umsebenzi omningi we-watt ngayinye kunezinhlelo zokuqhathanisa kuyilapho yehlisa ukubambezeleka sekukonke phakathi kwezikhathi eziyi-1.7 nezingu-3.6. Ngomthwalo wemisebenzi odinga ukusebenzisana njalo, i-OpenAI ibike ukuthuthukiswa kokusebenza okusuka kuzikhathi ezingu-2.1 kuye kwezingu-4.1. Lezi zibalo zichaza imiphumela yokuhlolwa ekhethiwe yenkampani, hhayi okutholakele okuvamile mayelana nemodeli ngayinye, umthwalo wokusebenza noma indawo yokusebenza.
Umbiko uphinde uchaze ukuqhathanisa endaweni yokusebenza ehlotshaniswa nesikhathi esingcono kakhulu sangaphambilini phakathi kwamathokheni akhiqiziwe. Kuleso silungiselelo, i-OpenAI ibike phakathi kuka-8.6 no-104.3 izikhathi eziningi zomsebenzi nge-watt ngayinye, kuye ngemodeli. I-Firstpost ayinikezi ulwazi olwanele embhalweni onikeziwe ukuze kunqunywe ukuthi lelo phuzu lokusebenza likhethwe kanjani, ukuthi amasistimu wokuqhathanisa acushwa kanjani, noma ukuthi umphumela umele kangakanani ukusetshenziswa kokukhiqiza okuvamile.
I-Firstpost ibika ukuthi i-OpenAI ilindele ukutshala i-Jalapeño ngamavolumu alinganiselwe ngaphakathi kwengqalasizinda yayo ekupheleni kuka-2026, okulandelwa ukukhishwa okunenjongo ngo-2027. Le nkampani futhi ithi umsebenzi ezizukulwaneni zayo ezimbili zama-chip usuqalile. Umbiko awuqinisekisi ukuthi i-chip okwamanje iyatholakala kumakhasimende angaphandle noma ukuthi uhlelo lokuphakelwa oluhleliwe luqinile.
Imininingwane yomthombo: firstpost.com ↗
Kungani kubalulekile
Imiphumela iphakamisa ukuthi i-OpenAI iphishekela ukulawula okukhulu ku-hardware esetshenziselwa ukunikeza amamodeli ayo, ikakhulukazi njengoba ama-agent ezinyathelo eziningi ze-AI andisa isidingo sokucabanga. Ukusebenza okuphezulu nge-watt ngayinye kungase kuthinte izindleko zokusebenza kanye nomklamo wesistimu, kodwa izibalo ezibikiwe ziyizimangalo ze-OpenAI ezivela ekuhlolweni kwebhentshimakhi futhi azikakaqinisekiswanga ngokuzimela emthonjeni.
Ukubaluleka kwamanje kungamasu kanye nobuchwepheshe. I-Firstpost ibika ukuthi i-OpenAI ifuna inkundla yezingxenyekazi zekhompuyutha ezikhiqiza izizukulwane eziningi lapho amamodeli wesikhathi esizayo, imikhiqizo, ama-chips kanye nezinhlelo zememori zingaklanywa ndawonye. Uma lolo hlelo luphumelela, i-OpenAI ingaba nomthelela owengeziwe endleleni amamodeli ayo ahlinzekwa ngayo futhi athuthukiswe ngayo esikhundleni sokuthembela ngokuphelele kuma-accelerator enhloso evamile ahlinzekwa ngezinye izinkampani.
Ukusebenza kahle kokucatshangelwa kubalulekile ngoba ukunikeza imodeli kuhlanganisa ukubala okuphindaphindiwe kuyo yonke impendulo. I-Firstpost ibika ukuthi i-OpenAI iklame i-Jalapeño ukuze ikwazi ukuthwala umsebenzi lapho ukubambezeleka kunganqwabelana kuzo zonke izinyathelo eziningi ezilandelanayo, njengama-ejenti e-AI abiza amamodeli ngokuphindaphindiwe kuyilapho esebenzisa amathuluzi, ehlola imiphumela futhi enquma ukuthi yini okufanele ayenze ngokulandelayo. Ngakho-ke ukusebenza komuntu ngamunye okusheshayo kungase kuthinte ukusabela kwemisebenzi ye-ejenti esebenza isikhathi eside, nakuba umthombo unganikezi izilinganiso ezizimele zokugeleza komsebenzi komenzeli okuphelele.
Izinzuzo zokusebenzisa amandla kahle ezibikiwe zingase zibe nendaba ezindlekweni nesikali somzimba samasevisi e-AI. Umsebenzi owengeziwe nge-watt ngayinye ungase unciphise isidingo sikagesi somsebenzi othile noma uvumele isistimu ukuthi inikeze ukucabangela okwengeziwe ngaphakathi kwesabelomali samandla esingaguquki. Leyo mithelela ihlala inemibandela: umthombo awunikezi amanani okuthenga, izindleko zokusebenza eziphelele, izinzuzo zokukhiqiza, izidingo zokupholisa, amanani okusebenzisa noma imiphumela evela endaweni yesikali sokukhiqiza.
Izinketho zedizayini ezibikiwe zika-Jalapeño ziqondisa izingqinamba ezaziwayo. I-Firstpost ithi i-architecture ikhuluma nesiteji sokugcwalisa kuqala, umkhawulokudonsa wememori ngesikhathi sokukhiqiza amathokheni kanye nokuxhumana phakathi kwamayunithi okucubungula. Iphinde ibike ukuthi isistimu igcina isimo semodeli, okuhlanganisa inqolobane ye-KV esetshenziswa ngesikhathi sokukhiqiza impendulo, eduze nalapho idingeka khona futhi ihlanganisa izinsiza zekhompuyutha, inkumbulo nezokuxhumana. Umthombo awuqinisekisi ngokuzimela ukuthi lezo zinketho ziletha izinzuzo ezifunwayo ngaphandle kokuhlolwa okubikiwe.
Indaba futhi ikhombisa imikhawulo yokuqhathanisa kwebhentshimakhi yesihloko. Izinhlelo ze-Nvidia zeBlackwell ziyizithenjwa ezivivinyweni ezibikiwe, kodwa i-Firstpost iphawula ukuthi ihadiwe eliqhudelanayo kungenzeka liqhubekele phambili ngaphambi kokuba i-Jalapeño ifinyelele ekusetshenzisweni okubanzi. Ngakho-ke umphumela uqondwa kangcono njengesiqhathaniso sakuqala esibikwe yinkampani esikhundleni sezinga elixazululiwe lengqalasizinda ye-AI. I-OpenAI futhi ilindele ukuqhubeka nokusebenzisa izisheshisi ze-Nvidia nezingxenyekazi zekhompuyutha ezivela kwabanye ozakwethu, ngakho-ke i-chip ayichazwa njengokuthatha indawo ngokushesha yabahlinzeki bayo abakhona.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Ukuhlolwa okubalulekile kuzoba ukuthi i-Jalapeño ingakhiqizwa futhi isetshenziswe esikalini, nokuthi iqhathaniswa kanjani ne-Nvidia entsha kanye nama-accelerator ayimbangi atholakalayo uma lokho kwenzeka. Okungaziwa okubalulekile kufaka phakathi umthamo wokukhiqiza, izindleko, ukwethembeka, amasayithi okuthunyelwa, ukufinyelela kwamakhasimende kanye nokuthi izinzuzo ezibikiwe ziyaqhubeka yini kuyo yonke ingxubevange yomthwalo obanzi we-OpenAI.
Ingqopha-mlando yokuqala esebenzayo ukuthi i-OpenAI iqala yini ukusetshenziswa kwangaphakathi okulinganiselwe ekuchaze ekupheleni kuka-2026. Ukuqapha leso sinyathelo kufanele kubandakanye ukuthi mangaki amasistimu afakiwe, amamodeli noma amasevisi awasebenzisayo, noma ngabe ukuthunyelwa kungokokuhlola noma kubheke ekukhiqizeni, nokuthi i-OpenAI ibika imiphumela yokusebenza ngale kwezibalo zebhentshimakhi. Ayikho kuleyo mininingwane esungulwe umbiko we-Firstpost.
Ukukhishwa kuka-2027 kuzobonisa ukuthi i-Jalapeño iyinkundla esebenzayo yini esikhundleni sephrojekthi yobunjiniyela yesizukulwane esisodwa. Ubufakazi obubalulekile buzobandakanya isikali sokukhiqiza, ukutholakala kwememori edingekayo kanye nezingxenye zenethiwekhi, ukwethembeka kwesistimu, ukusetshenziswa kanye nezindleko ngeyunithi ngayinye yokucabanga okuwusizo. Umthombo awunikezi umthamo wokukhiqiza, ukuhlukaniswa kwabahlinzeki ngale kwendima yokuthuthukisa ye-Broadcom, amanani entengo noma ulwazi lwezinga lesevisi.
Ukuqhathanisa okuzayo kuyodingeka kubhekane nehadiwe etholakalayo ngesikhathi sokuthunyelwa. I-Firstpost iphawula ngokucacile ukuthi izinhlelo ezintsha ze-Nvidia namanye amaphrosesa ayimbangi angase abe semakethe ngo-2027. Ngakho-ke ukuhlola okunengqondo kufanele kuqhathanise amamodeli afanayo, izidingo zekhwalithi yokuphendula, osayizi beqoqo, okuhloswe ukubambezeleka, ukuqagela kwamandla kanye nesitaki sesofthiwe kuzo zonke izinhlelo zesimanje kunokuthembela kusisekelo sanamuhla se-Blackwell.
Kubalulekile futhi ukubheka ukuthi izinzuzo ezibikiwe zivame ukudlula amamodeli amathathu ashiwo esihlokweni. I-GPT-OSS 120B, DeepSeek R1 670B kanye ne-Kimi K2.5 1T imele imodeli ethile nokukhetha komthwalo womsebenzi, kodwa umthombo awusho ukuthi i-Jalapeño isebenza kanjani kuwo wonke amamodeli okuphathelene we-OpenAI, amasistimu we-multimodal, amamodeli amancane esicelo sombhalo omude, amathuluzi asebenza ngombhalo omude. Lokho okungaziwa kukhawulela lokho okungacatshangwa mayelana nomthelela obanzi womkhiqizo.
Ekugcineni, ukuqhubeka kwe-OpenAI ukusebenzisa i-Nvidia nabanye abahlinzeki bazoveza ukuthi inkampani ibhalansisa kanjani isiko nehardware yezohwebo. Isu eliyingxubevange lingavumela i-OpenAI ukuthi isebenzise i-Jalapeño emisebenzini eminingi lapho ukwakheka kwayo kunenzuzo kuyilapho igcina ama-accelerator angaphandle ukuze aguquguquke noma umthamo. Umbiko awusho ukuthi i-Jalapeño izoke idayiswe ngaphandle, ilayisensi, noma isetshenziswe kuphela ngaphakathi kwengqalasizinda ye-OpenAI.