I-OpenAI iveza i-Jalapeño, i-chip yayo ye-AI esebenza kahle kakhulu kune-Nvidia ngokusebenza kahle nangesivinini

I-OpenAI ithathe isinyathelo esibalulekile esu layo lokuzimela kwezobuchwepheshe ngokwethulwa kwemiphumela yokuqala evela kuphrosesa yayo ye-AI eyenziwe ngokwezifiso, ebizwa ngokuthi i-Jalapeño. Yakhiwe ngokubambisana ne-Broadcom, le chip egxile ekucabangeni iphumelele kakhulu kune-Nvidia GB200 kanye ne-GB300 accelerators ngama-metric amabili abalulekile: amandla okucubungula nge-watt ngayinye kanye nesivinini sokuphendula. Idatha, ekhishwe engqungqutheleni ye-Hot Chips futhi yabikwa yizitolo ezifana ne-Bloomberg kanye ne-The Verge, ibeka i-OpenAI ngendlela engakaze ibonwe emjahweni wehadiwe ye-AI.

Lesi akusona isimemezelo esincane. Kuze kube manje, i-OpenAI ithembele cishe ngokuphelele kubaprosesa be-Nvidia ukuthi basebenzise amamodeli ayo, kokubili ezigabeni zokuqeqesha kanye nezokucabanga—okungukuthi, lapho uhlelo luphendula imibuzo yabasebenzisi. Nge-Jalapeño, inkampani ihlose ukunciphisa izindleko zayo zokusebenza nokuthola ukulawula ingqalasizinda yayo, yize ingayeki ngokuphelele i-Nvidia. Ngokusho kukaRichard Ho, inhloko yehadiwe ye-OpenAI, i-chip inikeza lokho akubiza ngokuthi "okungcono kakhulu kuzo zombili izindawo": amandla aphezulu okusebenzela amakhasimende amaningi ngesikhathi esisodwa kanye nokubambezeleka okuphansi kwezinhlelo zokusebenza ezisebenzisanayo.

I-athikili ehlobene:

I-OpenAI kanye nozakwethu we-Broadcom kuma-AI chips angokwezifiso

Iyini i-Jalapeño futhi isebenza kanjani?

I-Jalapeño iyi-application-specific integrated circuit (ASIC), eyenzelwe kusukela phansi umsebenzi owodwa: ukusebenzisa amamodeli olimi aqeqeshwe kusengaphambili. Ngokungafani nama-GPU e-Nvidia, angama-processor ajwayelekile, le chip ilungiselelwe isigaba sokuphetha, lapho uhlelo lucubungula isicelo futhi lukhiqize impendulo. Ayihloselwe amamodeli okuqeqesha, umsebenzi ohlala ungowe-GPU ye-Nvidia futhi i-OpenAI ezoqhubeka nokuwusebenzisa ngaleyo njongo.

Umklamo we-chip ugxile ekunciphiseni izithiyo ezivame ukwenzeka ngesikhathi sokuqagela, njengokunyakaza kwedatha phakathi kwememori namayunithi e-compute. Ngokusho kwenkampani, i-Jalapeño ihlanganisa isigaba esikhulu se-compute eceleni kwamabhulokhi ayisithupha e-high-bandwidth memory (HBM), okuvumela idatha ukuthi igcinwe endaweni futhi igweme ukudluliselwa okungadingekile. Ngaphezu kwalokho, i-chip ingasebenzisa inhlanganisela efanele yezinsiza zesigaba ngasinye sokucubungula, ithuthukise ukusebenza kwe-AI kanye nokusebenza kahle kwamandla.

I-athikili ehlobene:

I-Anthropic kanye ne-Samsung baxoxisana ngokudala i-chip yabo ye-AI ukuze baphonsele inselelo ukubusa kwe-Nvidia.

Imiphumela yesilinganiso uma iqhathaniswa neNvidia

Ukuhlolwa kwenziwe kusetshenziswa ipulatifomu yomphakathi ye-SemiAnalysis' InferenceX, elinganisa ukusebenza kwezinhlelo ze-AI ngaphansi kwezimo zangempela. I-OpenAI iqhathanise i-Jalapeño namaphuzu amahle kakhulu atholwe yi-Nvidia GB200 kanye ne-GB300 superchips, okwakuyizilinganiso ezitholakalayo ngesikhathi sokuhlolwa. Imiphumela ibonise ukuthi i-Jalapeño ifinyelela phakathi kwe-1,5 kanye ne-1,9 izikhathi zomsebenzi we-AI ngaphezulu nge-watt ngayinye, kanye ne-latency kusukela ekupheleni kuya ekugcineni phakathi kwezikhathi ezingu-1,7 kanye ne-3,6 eziphansi. Amamodeli asetshenziswe ku-test suite kwakuyi-GPT-OSS 120B, i-DeepSeek R1, kanye ne -Kimi K3 , ingxube esukela ezisindweni ezivulekile kuya kumamodeli asemngceleni.

Lezi zinombolo zibaluleke kakhulu ngoba ukusebenza kahle kwamandla sekungenye yezinselelo ezinkulu embonini. Izikhungo zedatha ze-AI zisebenzisa ugesi omningi, futhi ukunciphisa ukusetshenziswa ngeyunithi ngayinye yomsebenzi kungaholela ekongeni okukhulu. Ngokusho kukaRichard Ho, iJalapeño isebenza cishe kuma-watts angu-700, okuyisibalo esiphansi kakhulu se-accelerator ye-inference, futhi inganciphisa izindleko zokusebenza cishe ngo-50% uma kuqhathaniswa nama-GPU avamile, ngokusho kwe-CEO ye-Broadcom uHock Tan.

I-athikili ehlobene:

Isiphepho esiphelele se-silicon: kungani ukushoda kwe-AI chip kulapha ukuze kuhlale

Ukusebenzisana ne-Broadcom kanye nohlelo lokuqondisa

Ukuthuthukiswa kweJalapeño kuwumphumela wokubambisana okuhlelekile phakathi kwe-OpenAI ne-Broadcom, okwamenyezelwa ngo-Okthoba 2025. I-Broadcom inikela ngobuchwepheshe bayo ekuklanyweni kwe-silicon, ukuxhumana, kanye nokukhiqiza, kuyilapho i-OpenAI isebenzise amamodeli ayo e-AI ukusheshisa inqubo yokuklama. I-chip isuke ekuklanyweni kokuqala yaya ekukhishweni kwe-tape, isigaba sokugcina ngaphambi kokukhiqiza, ezinyangeni eziyisishiyagalolunye nje kuphela—isikhathi esirekhodiwe se-semiconductor ethuthukile.

I-OpenAI isivele isebenza esizukulwaneni sesibili seprosesa futhi inesithathu esisathuthukiswa. Le nkampani ihlela ukufaka izinhlelo zokuqala ngeJalapeño ngamanani amancane kakhulu ekupheleni kuka-2026 futhi yandise kakhulu ukusetshenziswa kwayo phakathi no-2027. Ngokusho kweBloomberg, ukufakwa kuzokwenzeka ezikhungweni zedatha zeMicrosoft nabanye ozakwethu, futhi isilinganiso kulindeleke ukuthi sifinyelele kuma-gigawatts ayi-10 wamandla okusebenzisa ikhompyutha ngo-2029. Kodwa-ke, imiphumela eyethulwe ingeyokuqala futhi ayikaqinisekiswa endaweni yokukhiqiza enkulu.

Imiphumela emakethe nakubasebenzisi

Ukufika kweJalapeño akusho ukuthi i-OpenAI iyayishiya iNvidia. Le nkampani igcizelele ukuthi iNvidia isalokhu ingumlingani obalulekile, ikakhulukazi ekuqeqeshweni kwamamodeli, nokuthi isidingo samandla ekhompyutha siphezulu kakhulu kangangokuthi sidinga abathengisi abaningi. Eqinisweni, i-OpenAI isebenza ne-AMD kanye ne-Cerebras, futhi isu layo lihilela ukwakha ingqalasizinda ehlukahlukene lapho umthwalo ngamunye unikezwa khona ihadiwe esebenza kahle kakhulu.

Kubasebenzisi nabathuthukisi, lezi zindaba zinemiphumela eqondile. Uma iJalapeño igcwalisa izithembiso zayo, ithokheni yezindleko ngesilinganiso ngasinye ingase yehle phakathi kuka-30% no-50% ezinyangeni ezingu-18 ezizayo, okuvumela izinkampani ezintsha nezinkampani ukuthi zinikeze izinsizakalo ze-AI ezingabizi kakhulu ngokubambezeleka okuphansi. Ngaphezu kwalokho, ukuncintisana emakethe ye-chip ye-AI kuyaqina, okungase kuhunyushwe ngamanani angcono kanye nezinketho eziningi kubasebenzisi bokugcina.

Okokugcina, kubalulekile ukuqaphela ukuthi izivivinyo azizange zifake ukuqhathaniswa nokwakhiwa okusha kwe-Vera Rubin kwe-Nvidia, okusanda kuqala ukuthunyelwa. Ngakho-ke, inzuzo ye-Jalapeño ingancipha lapho i-Nvidia ikhipha isizukulwane sayo esilandelayo. Noma kunjalo, intuthuko ye-OpenAI ikhombisa ukuthi imboni ye-AI ingena esigabeni esisha, lapho ihadiwe yangokwezifiso izodlala indima ebaluleke kakhulu. Umjaho wokusebenza kahle kanye nesivinini usanda kuqala.

I-athikili ehlobene:

I-Amazon izohlanganisa ubuchwepheshe be-Nvidia kuma-chip ayo e-AI futhi yethule amaseva amasha e-Trainium