Shipu yinye lengashintja kubalwa kwetindleko te-AI ebhizinisini lakho

Ngamhla ngu-25 Agasti 2026, i-OpenAI yashicilela imiphumela yekucala lekaliwe yetilinganiso te-Jalapeño, shipu ye-inference leyentiwe ngalokwakhetsekile leyitfutfukise ndzawonye ne-Broadcom. Tinombolo letibikiwe: umsebenti we-AI lomnyenti kusuka ku-1,5x kuya ku-1,9x nge-watt ngayinye esivinini lesiphakeme kakhulu, kanye nekubambeka kusuka ekucaleni kuya ekugcineni lokuphansi kusuka ku-1,7x kuya ku-3,6x kunemasistimu e-NVIDIA GB200 ne-GB300 leyacatsaniswa nawo. Ngemagama lalula, gesi lofanako ukhicita lokucishe kuphindzwe kabili, futsi timphendvulo tifika ngekushesha. Yiloko kanye lokunquma sikweleti sakho senyanga nenyanga uma usebentisa umsiti we-AI nobe i-automation etikwema-model elulwimi.

Yini kahle lokwamenyetelwa

Jalapeño akusiyo shipu yekuceceshwa. Yishipu ye-inference, lokusho kutsi isebentisa ema-model ngemuva kwekuceceshwa kwawo, futsi i-inference yincenye loyikhokhela kunoma ngusiphi sicelo se-API. Lena shipu yavetwa kwekucala ngamhla ngu-24 Juni 2026, futsi tinombolo tayo letikaliwe tetfulwa enkhomfeni ye-Hot Chips, tahlolwa nge-InferenceX, silinganiso lesivulekele umphakatsi lese-SemiAnalysis, kuma-model lamatsatfu lanesisindvo lesivulekile: GPT-OSS, DeepSeek R1 ne-Kimi K2.5.

  • Gesi: cishe 700W ngeshipu ngayinye, kantsi emasistimu lacatsaniswa nawo adla cishe lokuphindzwe kabili.
  • Umkhicito: umsebenti lomnyenti kusuka ku-1,5x kuya ku-1,9x nge-watt ngayinye esivinini lesiphakeme kakhulu.
  • Kubambeka: kuphansi kusuka ku-1,7x kuya ku-3,6x kusuka ekucaleni kuya ekugcineni, kuya nge-model lehloliwe.
  • Sikhatsi: kusetjentiswa kwekucala lokulinganiselwe ekupheleni kwa-2026, kukhuliswe kancane kancane kuwo wonkhe umnyaka wa-2027.

Fundza tinombolo ngeliso lelibukisisako

Tinkinga letingaba khona tifanele kutsi tikhulunywe ngaphambi kwenjabulo. Tinombolo tivela ku-OpenAI uqobo futsi taciniswa etilabhorethri tayo. Kucatsaniswa kwentiwa netitukulwane talamuhla te-NVIDIA, hhayi lesitako. Futsi tivivinyo betigcile emisebentini lemifisha, hhayi etingcocweni letidze te-agent letinematfuba lamanyenti. Noko, umlayeto lomkhulu usamile: emandla e-inference akuselona imakethe lenemhlinzeki munye, futsi noma ngubani lotsenga lawo mandla utawutfola tinketho letinyenti ngemanani laphansi eminyakeni lemibili letako.

Loku kusho kutsini ngalokusebentako kumnikati webhizinisi

Angeke uyitsenge shipu ye-Jalapeño nobe uyibone. Utawuva umtselela wayo etindzaweni letintsatfu:

  • Tindleko: uma kusebentisa i-model kushibhile kumhlinzeki, emanani ekusetjentiswa ehla kancane kancane. Iphrojekthi ye-AI lebeyibonakala ibita kakhulu umnyaka lowendlulile ingaba nenzuzo lamuhla.
  • Sivinini: kubambeka lokuphansi kuvula tindlela tekusebentisa lebetingenteki ngaphambilini: umsiti wetwi longenako kuthula lokungakahle, kusesha masinyane etincwadzini takho, sicelo lesibuyeketwa ngesikhatsi likhasimende lisemi ekhawuntheni.
  • Bunyenti: uma ticelo seshibhile, kusebentisa i-AI kuyo yonkhe intsengiselwano esikhundleni sesampula kuba nengcondvo: kuhlela yonkhe imilayeto yemakhasimende, kuhlola sonkhe sikweleti, kufingcita yonkhe ikholi.

Umtsetfo lobaluleke kakhulu: ungayiboshi sistimu yakho kumtsengisi munye

Sifundvo sangempela semjaho wemashipu e-inference akusiko kutsi ngubani lonqobako. Kutsi lonqobako uyashintja njalo ngemuva kwetinyanga letimbalwa. Noma ngubani lowakha sistimu lamuhla ngendlela lenta kutsi kushintja i-model kungenteki ukhokha kabili: kanye esikweletini, nakanye ekusebenteni lokuncono angeke akutfole.

Sincumo sebunjiniyela lesikongela imali akusiko kukhetsa i-model lencono lamuhla. Kwakha sistimu lekuvumela kutsi ushintje i-model kusasa ngaphandle kwekubhala kabusha i-application.

Ngalokusebentako loku kusho lungcweshana loluncane emkhatsini we-application yakho nemhlinzeki we-model, ema-prompt nematfulusi lobhalwe ngendlela lengancikile kunoma ngumuphi umtsengisi munye, imininingwane yakho nelwati kugcinwe kusistimu yakho hhayi ku-platform yemhlinzeki, kanye neluhlu lwetivivinyo telizinga lolusebentisa kunoma nguyiphi i-model lensha ngaphambi kwekuyamukela.

Indlela lesikwenta ngayo loku e-Origami

Uma sakhela likhasimende sistimu lesebentisa i-AI, sihlala sehlukanisa umcondvo webhizinisi ne-model leyiwentako. Lwati nemininingwane kuhlala kusizindalwati selikhasimende uqobo, ema-prompt abhalwa kutsi akhone kudluliselwa, futsi kushintja i-model kungushintjo lwekuhlela, hhayi iphrojekthi. Ngetulu kwaloko silinganisa tindleko ngakunye kusebenta, hhayi ngenyanga kuphela, kute wati kahle kutsi nguyiphi incenye yesistimu ledla sabelomali, nguyiphi lefanele i-model lenemandla kakhulu, nekutsi nguyiphi lenelisekile ngalokuphelele nge-model lencane lesheshako.

Luhlangotsi lwemtsetfo alushintji nentengo. I-inference leshibhile ayishintji lutfo etibophweni takho ngaphansi kweMtsetfo Wekuvikelwa Kwemininingwane Yemuntfu (PDPL) waseSaudi Arabia. Nanoma ngumiphi imininingwane yemakhasimende ledlula ku-model yangaphandle isadzinga sisekelo lesisemtsetfweni, kulawulwa kwekungena kanye nemarekhodi ekuhlolwa, nobe tindleko tekusebenta tingehla kangakanani.

Tinyatselo letisebentako tale liviki

  • Buyeketa sikweleti sakho se-AI samanje bese ubala tindleko ngakunye kusebenta lokuwusito, hhayi inani lenyanga kuphela.
  • Hlonza imisebenti lengentiwa yi-model lencane nalesheshako bese uyidlulisela khona; ngalokuvamile incenye yekusetjentiswa kwakho ayidzingi i-model lenemandla kakhulu.
  • Bhala luhlu lolufishane lwetivivinyo telizinga lolumele timo takho tebhizinisi tangempela, bese uyalusebentisa kunoma nguyiphi i-model lenye ngaphambi kwekushintja.
  • Buyeketa tibopho tesikhatsi lesidze nemtsengisi munye ngaphambi kwekutivuselela; imakethe ihamba ivuna umtsengi.
  • Buyela emicabangweni lowayala ngenca yetindleko nobe kubambeka: leminye yayo seyiyenteka nyalo.

Sifinyeto: loko lokwenteka ngamhla ngu-25 Agasti akusiyo ncesi lemayelana neshipu. Kuluphawu lwekutsi lungcweshana loluncika kulo emasistimu akho e-AI seluyashibha, seluyashesha futsi seluncintisana kakhulu. Emabhizinisi lakhe emasistimu lavumelana nesimo atawutfola lokwentiwa ncono ngalokutentakalelako. Emabhizinisi labotjelwe kunketho yinye atawukubuka kudlula.

Imitfombo