Ezinsukwini ezimbalwa ezedlule, ulwazi lukhishwe yithimba labacwaningi mayelana nokuthuthukiswa kwe okokuqala Rowhammer attack lokho iqondiswe ngempumelelo ku la Imemori yevidiyo ye-GDDR6 ye-GPU, ikakhulukazi i-NVIDIA A6000.
Ubuchwepheshe, ebizwa ngokuthi i-GPUHammer, ivumela amabhithi angawodwana ku-DRAM ye-GPU ukuthi akhonjwe indlela, ehlisa isithunzi kakhulu ukunemba kwamamodeli okufunda omshini ngokushintsha ingxenye eyodwa nje yamapharamitha awo. Lawa ma-bit flips avumela umsebenzisi we-GPU onobungozi ukuthi alawule idatha ye-GPU yomunye umsebenzisi endaweni okwabelwana ngayo, enqunyiwe yesikhathi.
Kuze kube manje, Ukusebenzisa i-Rowhammer kuzinkumbulo zevidiyo kwakubhekwa njengento engenakwenzeka ngenxa yemikhawulo eminingana yobuchwepheshe. Ukwakheka komzimba kwamaseli enkumbulo kuma-chips e-GDDR kunzima ukukudweba, ukubambezeleka kokufinyelela kuhamba kancane ngokuphindwe kane kune-DRAM evamile, futhi amanani okuvuselela aphezulu kakhulu. Okungezwe kulokhu izindlela zokuvikela zobunikazi ekulahlekelweni kwenkokhiso ngaphambi kwesikhathi, ubunjiniyela obuhlehlayo obudinga okokusebenza okukhethekile.
Ukuze unqobe lezi zithiyo, Abacwaningi bathuthukise indlela entsha yobunjiniyela eqondiswe ku-GDDR DRAMBesebenzisa ikhodi ye-CUDA esezingeni eliphansi, bahlasele ngokulungisa okuthile okuqinisa ukufinyelela kumaseli athile enkumbulo, okudala izimo ezivumela ukukhohlisa kancane. Isihluthulelo sempumelelo sisekuzuzweni kwekhompuyutha ehambisanayo ehleleke kakhulu, esebenza njengesandisi sokucindezela kumaseli aseduze.
Kusebenza kanjani lokhu kuhlasela?
Ukuhlasela isebenzisa ubuthakathaka bomzimba ku-DRAM, lapho ukufinyelela okujulile kumugqa wenkumbulo (okwaziwa ngokuthi “i-hammering”) ingadala izinguquko emigqeni eseduzeNakuba lobu bungozi buhlonzwe ngo-2014 futhi kwafundwa kabanzi kumemori ye-CPU DDR, ukukuyisa kuma-GPU kuze kube manje kube yinselelo ngenxa:
- Ukubambezeleka okuphezulu kokufinyelela kwe-GDDR6 (kufika izikhathi ezi-4 ngaphezu kwe-DDR4).
- Ubunkimbinkimbi ekwabiweni ngokomzimba kwenkumbulo.
- Ukuba khona kokuncishiswa okuphathelene nokuphathelene nokungabhaliwe kahle, njenge-TRR.
I-Rowhammer ukuba sengozini kwehadiwe lapho ukwenza kusebenze umugqa owodwa wenkumbulo ngokushesha kwethula ukuphenduka okuncane emigqeni encikene. Kusukela ngo-2014, lobu bungozi bufundwe kabanzi kuma-CPU nenkumbulo esekwe ku-CPU njenge-DDR3, DDR4, ne-LPDDR4. Kodwa-ke, njengoba i-AI ebucayi kanye nemithwalo yokufunda yomshini manje isebenza kuma-GPU ahlukene efwini, ukuhlola ubungozi benkumbulo ye-GPU ekuhlaselweni kwe-Rowhammer kubalulekile.
Naphezu kwalezi zingqinamba, i Abacwaningi bakwazile ukusebenzisa ubunjiniyela obuhlehlayo ekwabelweni kwenkumbulo ebonakalayo/yomzimba ku-CUDA, Bakha indlela yokuhlonza amabhange enkumbulo athile e-DRAM kanye nokufinyelela okuhambisanayo okulungiselelwe kusetshenziswa imicu eminingi nama-warps, okukhuphula izinga lokusando ngaphandle kokubangela ukubambezeleka okwengeziwe.
Ubufakazi bomqondo bubonise ukuthi i-single-bit flip ku-deep neural network (DNN) izisindo, ikakhulukazi kuma-exponents e-FP16, ingehlisa kanjani ukunemba okuphezulu koku-1 kwamamodeli wokuhlukaniswa kwezithombe ku-ImageNet isuka ku-80% iye ku-0,1%. Lokhu okutholakele kuyethusa ezikhungweni zedatha nezinsizakalo zamafu ezisebenzisa imithwalo yemisebenzi ye-AI ezindaweni okwabelwana ngazo nama-GPU.
Ukunciphisa kanye nemikhawulo
I-NVIDIA iqinisekise ukuba sengozini futhi incoma ukunika amandla usekelo lwe-ECC. (Ikhodi Yokulungisa Iphutha) usebenzisa umyalo nvidia-smi -e 1. Nakuba Lesi silinganiso singalungisa amaphutha i-single-bit, Lokhu kusho ukulahleka kokusebenza okungafika ku-10%. kanye nokuncipha kwenkumbulo etholakalayo ngo-6,25%. Futhi ayivikeli ekuhlaselweni okuzayo okubandakanya ukuphenduka kwamabhithi amaningi.
Siqinisekise ukuguquguquka kwebhithi ye-Rowhammer kuma-NVIDIA A6000 GPUs ngememori ye-GDDR6. Amanye ama-GDDR6 GPU, njenge-RTX 3080, awazange abonise ukuguquguquka okuncane ekuhlolweni kwethu, okungenzeka ngenxa yokuhluka komthengisi we-DRAM, izici ze-chip, noma izimo zokusebenza ezifana nezinga lokushisa. Asizange futhi sibone ukushintshashintsha kwe-A100 GPU enenkumbulo ye-HBM.
Ithimba liyakuveza lokho I-GPUHammer okwamanje iqinisekiswe kuphela ku-A6000 GPU nge-GDDR6, futhi hhayi kumamodeli afana ne-A100 (HBM) noma i-RTX 3080. Kodwa-ke, njengoba lokhu kuwukuhlasela okunwebekayo, abanye abacwaningi bakhuthazwa ukuba baphindaphinde futhi banwebe ukuhlaziywa kwezakhiwo namamodeli e-GPU ahlukene.
Okokugcina, uma ungathanda ukufunda okwengeziwe ngakho, ungabheka imininingwane ku isixhumanisi esilandelayo.