Kwenzekeni
Irejista ibika umnyakazo onwebekayo obheke kuzingxenyekazi zekhompuyutha zangokwezifiso nenye ye-AI. I-akhawunti yayo ihlanganisa i-Cerebras's modular rack rack system ye-CS-4, indima ekhulayo kaMarvell ekwakhiweni kwe-accelerator yangokwezifiso, ukuqhubeka nokusebenzisa kwe-Google ama-TPU akhethekile, kanye nokuthuthukiswa kukaWaymo kwesisheshisi sokufunda ngomshini sezimoto ezizimele. Umbiko uphinde uchaze i-Microsoft ebuyisela ngokwezifiso ezimbalwa zeWindows nezici zokuqapha ze-AI, kodwa lokho kuwuchungechunge lwesibili oluhlukile.
Umbiko we-Register's Aug. 24 report ugxile kulokho abahleli balo abakuchaza njengokudlulela okungaphezu kwendima enkulu ye-Nvidia kuma-datacenters e-AI. Abathengisi baya ngokuya behlukanisa imithwalo yemisebenzi phakathi kwezinhlobo ezahlukene ze-silicon esikhundleni sokuthembela ohlotsheni olulodwa lwe-accelerator. Iveza okucatshangwayo, noma ikhiqiza imiphumela evela kumamodeli aqeqeshiwe, njengendawo yokucindezela okukhulu: abahlinzeki bafuna amanani amathokheni aphezulu abasizi bokubhala amakhodi nezinye izinsizakalo ezisebenzisanayo kuyilapho belawula izindleko zamandla nezingxenyekazi zekhompuyutha zokunikeza izicelo. I-athikili yeRejista iwumbhalo we-podcast nokuhlaziya, ngakho-ke kufanele ifundwe njengokubika okuvela kuleyo ndawo kunokuba kubhalwe ngokuzimela amadokhumenti azo zonke izimangalo zobuchwepheshe.
I-Register ibika ukuthi i-Cerebras isuka kumasistimu amakhulu, e-monolithic iya kumklamo wayo we-CS-4 we-rack-scale. Ngaphambilini amasistimu e-Cerebras abeke isikali se-wafer-scale esikhulu, esilambele amandla ku-chassis oziqukathise uketshezi. Indlela entsha ibeka ama-chips amathathu ekuhlelweni kwe-rack ye-modular, okuvumela izingxenye zekhompiyutha ukuthi zishintshwe ngaphandle kokuthatha indawo yemishini ehlobene yokuletha amandla. IRejista ithi umklamo uphinda kabili isivinini sewashi, umkhawulokudonsa wenkumbulo kanye nesivinini sokufakwayo, kuyilapho ubeka i-silicon ephindwe kathathu ku-rack. Iphinde ithi i-Cerebras inobambiswano ne-AWS kanye ne-AMD, nokuthi isizukulwane samathokheni esisheshayo sihehe intshisekelo yabasizi bokubhala amakhodi nezinye izinsiza zokukhomba.
Umbiko uchaza i-ecosystem ebanzi ye-silicon ezungezile ezungeze ama-hyperscaler: I-Google isebenzise i-Tensor Processing Units yayo kusukela ngonyaka ka-2015 futhi ithembele ku-Broadcom ngezingxenye zomklamo wokusheshisa, kuyilapho u-Marvell eba imbangi ku-XPU yangokwezifiso nomsebenzi wokuxhumana ngemva kokwakha iphothifoliyo yobuhlakani ngokusebenzisa ubuchwepheshe. I-athikili ifaka lokhu njengendlela yokuthi izinkampani zamafu ziqhathanise abahlinzeki noma zinciphise ukuncika kuzakwethu wokuklama oyedwa, kodwa asibhali inkontileka entsha ethize ye-Google-Marvell; Ukugomela mayelana namasu ekhasimende lesikhathi esizayo kanye nengcindezi yentengo kuwukuphawula, akuyona intuthuko emisiwe. U-Waymo udalule isisheshisi sokufunda umshini ngokwezifiso esihloselwe izimoto ngenxa yokuthi ivolumu yenzwa nokubambezeleka kuyabalulekile uma kuvela isithiyo. Irejista ithi i-Waymo ithembele kakhulu kuhlelo lwamasango ahlelekayo futhi ingase ifune ukuminyana okukhulu kwekhompiyutha kanye nokuthuthukiswa okulula kwesifunda esihlanganisiwe esiqondene nohlelo lokusebenza, kodwa ayinikezi imininingwane, imiphumela yokuhlola, uhlelo lwezikhathi zokukhiqiza noma irekhodi lokuphepha. Ngokuhlukana, i-Microsoft ihlola ukuthi iyakwazi ukususwa Windows 11 amabha emisebenzi, amamenyu womongo enziwa ngokwezifiso kanye nokubonakala Kwesiphathi Somsebenzi ku-neural-processing-unit workloads.
Imininingwane yomthombo: theregister.com ↗
Kungani kubalulekile
Ukushintsha kungenza ingqalasizinda ye-AI incike kancane ekilasini elilodwa le-Nvidia GPU futhi igcizelele kakhulu ihadiwe eklanyelwe imithwalo yemisebenzi ethile. Izikhathi zokuphendula ezisheshayo, izindleko eziphansi zokusebenza, ukusebenza kahle kwamandla okungcono kanye nokuhlanganiswa okuqinile nezinzwa yimigomo esebenzayo echazwe yiRejista. Umbiko awuqinisekisi ukuthi noma iyiphi enye indlela eyedlule ngobubanzi i-Nvidia ekusetshenzisweni komhlaba wangempela.
Ukubaluleka okusebenzayo ukuthi ukusebenza kwe-AI akunqunywa imodeli kuphela. Umbiko weRejista ubonisa abathengisi abathuthukisa ihadiwe ngaphansi kokucatshangelwa kwemikhawulo ethile: isivinini sesizukulwane samathokheni samathuluzi asebenzisanayo, amandla nokupholisa kwama-datacenter racks, kanye nokubambezeleka kokulawula imoto. Uma leyo miklamo isebenza ngezinga eliphezulu, izinkampani zingakhetha ama-accelerator ahlukene ukuze aqeqeshwe, i-volume high , izicelo ze-premium low-latency kanye nemithwalo yemisebenzi enqenqemeni. Lokho kungenza imakethe yengqalasizinda ye-AI ikhetheke kakhulu futhi inganikeza amakhasimende ezinye izindlela zokusebenzisa izinhlelo ze-GPU ezihloselwe ezijwayelekile.
Ukuklanywa kabusha kwe-rack okubikiwe kwe-Cerebras kugqamisa inkinga yokusebenza engaphuthelwa ezimangalweni zokusebenza kwesihloko. IRejista ithi amasistimu angaphambili adle cishe ama-kilowatts angu-15 ku-chip futhi ahlanganiswe ikhompuyutha nemishini yokuletha amandla ku-chassis enkulu. I-rack ye-modular ingenza ukulungiswa nokuthuthukiswa kungaphazamisi kangako, noma ngabe i-silicon ihlala ilambile amandla ngendlela engavamile. Kubasebenzisi be-datacenter, ukusebenza kahle, ukufaka ikhebula, ukupholisa kanye nekhono lokushintsha ingxenye eyodwa ehlulekile kungase kubaluleke njengomphumela omkhulu. Umthombo awuqinisekisi ngokuzimela amandla abikiwe noma izibalo zokusebenza, ngakho lezo zinzuzo zihlala ziyizimangalo ezidinga ukubuyekezwa kobuchwepheshe.
Isisekelo esibanzi sabahlinzeki singathinta amandla okuxoxisana nezinqumo zokutshala izimali. I-Register iphakamisa ukuthi izinkampani zamafu zivame ukusebenzisa amafemu angaphandle ezindawo zobuhlakani ukuthola izingxenye ezihlukile zama-chips angokwezifiso kuyilapho zigxile kwezobunjiniyela bangaphakathi ezicini ezihambisana namasevisi azo. Ukuncintisana phakathi kwe-Broadcom ne-Marvell kunganikeza ama-hyperscaler omunye umzila wokuklama, kodwa i-silicon yangokwezifiso futhi idala izibopho zesofthiwe nokusekela. Ukusebenza kahle kwe-chip kumthwalo owodwa kungase kube nzima ukukuhlela, kube nzima ukukuthola noma ukungafaneleki komunye. Umbiko awunikezi ukuqhathaniswa kwezindleko nezinhlelo ze-Nvidia futhi abukho ubufakazi bokuthi amakhasimende ashintsha ngezinga elibanzi. Isibonelo sika-Waymo sixhuma i-Hardware ye-AI yangokwezifiso ekusetshenzisweni okuzwelayo ukuphepha esikhundleni sezomnotho ze-datacenter kuphela: ukubambezeleka okuphansi kuyabaluleka lapho imoto iphendula okokufaka kwenzwa, futhi ukungenelela komuntu okukude akuyona indawo efanelekile yokucubungula ngokushesha. Lokho akubonisi ukuphepha okuthuthukisiwe; akukho ukuhlola okuzimele, ukuqhathanisa izinga lokuhluleka noma ukuhlola okulawulayo. Isiphetho esincane siwukuthi izimoto ezizimele zinikeza isizathu sokuklama i-silicon ngesikhathi sokucubungula kanye nesikhathi sokuphendula, kuyilapho umthelela womphakathi uncike ekuqinisekisweni kwezimo zokusebenza zangempela.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Ukuhlola okubalulekile imibhalo yezobuchwepheshe yomphakathi, amabhentshimakhi azimele kanye nobufakazi bokusetshenziswa okulinganiselwe kwamakhasimende. Buka ukuthi idizayini entsha ye-rack ye-Cerebras, izisheshisi ngokwezifiso ezisekelwa i-Marvell kanye ne-silicon yemoto ka-Waymo iyafinyelela yini ekukhiqizweni, nokuthi izinzuzo zazo zingaphezu kwezindleko zesofthiwe, ukuthengwa kwempahla kanye nezindleko zokuyilungisa. Izinguquko ze-Windows Microsoft kufanele futhi zahlulelwe ngokutholakala komphakathi nokuthi ziyathuthukisa yini ukulawula komsebenzisi kunokumane zengeze ukuqapha kwesistimu okwengeziwe.
Okokuqala, bheka izinto zobuchwepheshe eziyisisekelo ezivela kwaCerebras, Waymo, Google, Marvell noma amakhasimende abo. Ubufakazi obuwusizo buzobandakanya imibhalo yezakhiwo, izilinganiso zamandla, ukusekelwa kwesofthiwe, isimo sokukhiqiza kanye nokuhlolwa komthwalo womsebenzi okuchazwe ngokucacile. Umbiko weRejista unikeza imephu ebalulekile yezindaba yezinkampani ezihilelekile, kodwa awukutholi ukutholakala, amanani, umthamo wamakhasimende noma ukusebenza okuzimele. Leyo mininingwane engekho inquma ukuthi ukuthuthukiswa kuwunyaka wemboni noma ikakhulukazi izinhlelo zobunjiniyela.
Okwesibili, qhathanisa njengokufana nokufana. Izibalo zethokheni ngesekhondi zingahluka ngosayizi wemodeli, ukunqwabelana, ukunemba, ubude bomongo kanye nenani labasebenzisi ngesikhathi esisodwa. I-accelerator yangokwezifiso ingase ibe yinhle kakhulu kuphethini eyodwa encane ye- futhi ingasebenziseki kangako emithwalweni evamile yomsebenzi. Buka izivivinyo ezizimele zokulinganisa okuphumayo, ukubambezeleka kwempendulo, amandla ngethokheni ngayinye ekhiqiziwe, ukusetshenziswa, ukwethembeka kanye nenani eliphelele lezindleko zobunikazi ngokumelene nezinhlelo ze-GPU zamanje. Ngaphandle kwalowo mongo, izimangalo ze-AI esheshayo noma eshibhile kuhlala kunzima ukuzihlola. Okwesithathu, landela ubudlelwano bezohwebo: Irejista ibika ubudlelwano be-Cerebras ne-AWS kanye ne-AMD futhi ichaza u-Marvell njengongena emkhakheni owawubuswa i-Broadcom kanye namaqembu angaphakathi e-hyperscaler. Ubufakazi obulandelayo obubalulekile kuzoba ukuthi lobo budlelwano bukhiqiza izinsizakalo ezifinyeleleka ngokuvamile, ukuthunyelwa okuqanjwe amagama noma ama-oda aphindayo. Ukulindela kombiko ukuthi i-AWS noma izinsiza ze-AMD zingasebenzisa i-Cerebras hardware kuwukuphawula okubheke phambili, hhayi ukukhishwa okuqinisekisiwe, futhi akufanele kuthathwe njengento eyodwa.
Okokugcina, qapha ukusetshenziswa kwemoto ka-Waymo kanye nesofthiwe ye-Microsoft ishintsha ngokuhlukana. Ku-Waymo, imibuzo enengqondo ukuthi ingabe i-accelerator ifakiwe ezimotweni zokukhiqiza, ukuthi iyisingatha kanjani imithwalo yemisebenzi yezinzwa, yiziphi izinhlelo zokubuyela emuva ezikhona nokuthi yikuphi ukuqinisekiswa kokuphepha okuqediwe. OkweWindows, iRejista ithi ukunyakaza kwebha yomsebenzi kusesiteshini Sokubuka Kukakhululiwe futhi ukubonakala kwenqubo ye-NPU okuthuthukisiwe kuyathuthukiswa, kodwa isikhathi somphakathi nokuziphatha kokugcina kungashintsha. Alukho uchungechunge okufanele lwethulwe njengoluphelele kuze kube yilapho izinkampani zishicilela ukutholakala okuzinzile futhi abasebenzisi abazimele bangahlola imiphumela.