I-Google yethula i-Gemma 3: i-AI yayo yomthombo ovulekile ethuthuke kakhulu ye-GPU eyodwa

I-Google isimemezele ngokusemthethweni ukwethulwa kwe-Gemma 3, inguqulo entsha yemodeli yobuhlakani bokwenziwa yomthombo ovulekile, yakhelwe ukusebenza kahle ekhadini lemifanekiso elilodwa (GPU). Ngalokhu kuthuthukiswa, inkampani ifuna ukunikeza i-AI efinyeleleka kalula futhi evumelana nezimo kubathuthukisi nezinkampani eziyidingayo izixazululo ezithuthukisiwe ngaphandle kwesidingo sengqalasizinda ebizayo.

Umndeni wakwaGemma uye wabhekana nokuziphendukela kwemvelo okuqhubekayo. Ngemva kokukhishwa kwezinguqulo zokuqala ngoFebhuwari wonyaka odlule, okwakunamamodeli anamapharamitha ayizigidi eziyizinkulungwane ezimbili nezingu-2.000, kanye nokufika okwalandela kweGemma 7.000 ngoMeyi namapharamitha ayizigidi eziyizinkulungwane ezingama-2, I-Google manje inweba uchungechunge nge-Gemma 3, imodeli eklanyelwe ukulinganisa amandla nokusebenza kahle ezindaweni ezahlukene..

Izici eziyinhloko ze-Gemma 3

I-Gemma 3 yakhelwe ngamasayizi amane: 1, 4, 12 kanye namapharamitha ayizigidi ezingama-27, ukuyivumela ukuthi ivumelane nezingxenyekazi zekhompuyutha ezahlukene kanye nezidingo zokusebenza. Ukusebenzisa kahle kwayo kuyivumela ukuthi isebenze kumadivayisi ane-GPU eyodwa, emboza yonke into esuka umakhalekhukhwini up izindawo zokusebenza ezithuthukisiwe.

Esinye sezici zayo eziphawuleka kakhulu Ukusekelwa kolimi okungama-140, ngokusekelwa ngaphandle kwebhokisi ngezilimi ezingaphezu kuka-35, okwenza kube lula ukwakha izinhlelo zokusebenza zezilimi eziningi ngaphandle kokucushwa okwengeziwe.

Ukusebenza okuthuthukisiwe nokuqhathanisa namanye amamodeli

Ngokusho kweGoogle, iGemma 3 inezinhlobo ezincintisanayo ezisebenza kahle kakhulu ezivivinyweni zokusebenza ezifana nalezi I-Meta's Llama-405B, i-OpenAI's DeepSeek-V3, kanye ne-o3-mini, okubonisa ukusebenza kahle kwayo ekwenzeni imisebenzi eyinkimbinkimbi nge-GPU eyodwa.

Imodeli ilungiselelwe ukusebenza ikakhulukazi nayo amakhadi wehluzo we-nvidia, kodwa futhi ingasebenza nayo Ama-TPU kanye nezinye izinkundla zehadiwe, okuqinisekisa ukuhlanganiswa kwazo ezindaweni eziningi zokuthuthuka.

Izici Ezithuthukile Nezinhlelo Zokusebenza

Phakathi kwentuthuko ebaluleke kakhulu ye-Gemma 3, yayo iwindi lomongo elinwetshiwe lamathokheni angu-128K, okuvumela ukucutshungulwa nokuhlaziywa kwenqwaba yolwazi. Iphinde ibe namandla okwenza hlaziya umbhalo, izithombe, namavidiyo amafushane, okulenza libe ithuluzi eliyinhloko lobuhlakani bokwenziwa obukhiqizayo.

Ngaphezu kwalokho, ihlanganisa imisebenzi efana izingcingo zokusebenza nokuphumayo okuhleliwe, okukuvumela ukuthi wenze imisebenzi ngokuzenzakalelayo futhi uthuthukise abasizi abahlakaniphile ngaphakathi kwezinhlelo zokusebenza ezahlukahlukene.

I-ShieldGemma 2: Isixazululo Sezokuphepha se-AI

Kanye ne-Gemma 3, i-Google nayo yethuliwe I-ShieldGemma 2, isiqinisekisi sokuvikeleka esisekelwe ku-AI esidizayinelwe ukuhlaziya nokulebula izithombe ngokwezigaba ezintathu: Okuqukethwe okuyingozi, izinto ezibekela ingcaca ucansi nodlame. Leli thuluzi livumela abathuthukisi ukuthi benze amasistimu abo okuvikela afanele izidingo zohlelo lokusebenza ngalunye.

Ukutholakala nokuhlanganiswa ezisekelweni ezahlukene

I-Gemma 3 isiyatholakala futhi ingahlanganiswa ezindaweni eziningi zokuthuthuka. Ukufinyelela kunikwe amandla Ubuso Obugonayo, i-Kaggle, i-Google AI Edge, ne-Vertex AI, ngaphezu kokuba nokusekelwa ku I-Google AI Studio ukwenza lula ukuqaliswa kwayo.

Lokhu kwethulwa kuqinisa ukuzibophezela kwe-Google ekuthuthukiseni amamodeli e-AI avulekile nafinyelelekayo. Nge-Gemma 3, inkampani ihlose ukunikeza abacwaningi nabathuthukisi ithuluzi elinamandla, elilungiselelwe, neliguquguqukayo lezinhlelo zokusebenza eziningi ze-AI.