Ngokombiko we-DigiTimes wanamuhla, iMicron icwaninga ngamamojuli e-NAND flash aqinile kakhulu. Le nkampani ihlela ukuwafaka eduze ne-GPU, kunokuba, njengezixazululo zendabuko zokugcina idatha, iwaxhumanise ngamaphrothokholi amaningi nedamu lesitoreji elikude ne-GPU.
Ngokwemibiko, iMicron ihlola ukuthi ingayisondeza kanjani i-NAND flash ku-GPU, ngethemba lokunikeza isisombululo esiphakathi phakathi kobuningi bokugcina idatha nokuqina. Lo mkhiqizo okwamanje ubizwa ngokuthi “near-GPU NAND” (inothi: near-GPU NAND), futhi ungasetshenziselwa ukuphatha imithwalo yomsebenzi engadingi ukubambezeleka okuphansi noma umkhawulokudonsa ophezulu.
Lolu hlelo lungenza i-GPU ifinyelele ngokuqondile emachibini e-NAND anomthamo wamakhulu ama-GB. Lungafakwa ngokuqondile ngaphakathi kwephakheji ye-GPU noma ku-PCB, futhi lusebenze ngendlela efana nezixazululo ezifana ne-HBM ne-GDDR7. Uma luqhathaniswa ne-TLC noma i-QLC NAND yendabuko, lolu hlobo lwe-chip ye-NAND lunobuningi bokugcina obuphansi, lugxile ekuthuthukiseni isivinini se-I/O nomkhawulokudonsa.
Lo ngqimba lokugcina utholakala phakathi kwememori yevidiyo nesitoreji esivamile, ngakho ungasebenza njengongqimba lwebhafa olusheshayo futhi ulungele ikakhulukazi ukusebenzisa amamodeli amakhulu olimi e-AI. Uma lolu hlelo selusetshenziswa, ukucubungula iziphetho kwe-LLM ngeke kusavinjelwa umthamo wememori yevidiyo.