Izixazululo ezintsha zezithombe ze-NVIDIA zivusa ilukuluku ngokusebenza kwazo ku-Geekbench

Ukungagwemeki kokumenyezelwa kwezixazululo ezintsha zehluzo ze-NVIDIA kuqinisekiswa abahlaziyi abaningi, kodwa kusenobufakazi obuncane bangempela. Njengokugcina, singacabangela imiphumela yokuhlola imikhiqizo engaqondakali yalolu hlobo e-Geekbench, esivumela ukuthi sikhulume ngokuphakama ngaphezu kwe-Tesla V100 (Volta).

Imiphumela yokuhlola imikhiqizo emibili ehlukene ye-NVIDIA ku-Geekbench 5.0.2, etholwe emuva ngo-Okthoba nangoNovemba wonyaka odlule, inengqondo ukucatshangelwa kumongo wokusebenza endaweni ye-Open CL. Umkhiqizo wokuqala onamayunithi ekhompyutha angu-118 kanye nenkumbulo engu-24 GB etholwe kulokhu kuhlolwa 184 amaphuzu. Umkhiqizo wesibili onamayunithi ekhompyutha angu-108 kanye nenkumbulo engu-48 GB 141 amaphuzu. Esimweni sokuqala, imvamisa yokukhawulela ye-GPU yayingu-1,11 GHz, okwesibili - 1,01 GHz.

Ngeke kukhishwe ukuthi kuzo zombili izimo, ama-GPU esizukulwane esilandelayo se-NVIDIA, ngokuvamile abizwa ngokuthi i-Ampere, ahlolwe. Inani lama-cores e-CUDA lingafinyelela ku-7552 nge-GPU yokuqala, kanye no-6912 kwesibili. I-accelerator ye-Tesla V100 yesizukulwane se-Volta, siyakhumbula, igcwele inkumbulo engu-16 GB ye-HBM2 nama-cores angu-5120 CUDA, nakuba kukhona ukuguqulwa nge-32 GB yememori yohlobo olufanayo.

Ezinye izilinganiso kumasayizi wenkumbulo ye-Ampere, uma lezi zingamaphrosesa ezithombe zesizukulwane esisha, zingachazwa ngokushintsha ukusuka kubhasi lememori engu-4096-bit ukuya kweyi-6144-bit. Ngaphezu kwalokho, abamele indawo I-3DCenter kholelwa ukuthi i-NVIDIA GA100 GPU ekucushweni kwayo okugcwele izoba nama-cores angu-8192 CUDA, futhi inguqulo yayo encishisiwe ibonwe kusizindalwazi se-Geekbench. Ukwehliswa ngokulinganayo kwenani lama-cores asebenzayo ama-GPU amakhulu ejiyomethri akunakugwenywa, njengoba amathuba okuba nokonakala kwe-chip eba phezulu.

Isivinini sewashi lama-GPU okukhulunywa ngawo akude kakhulu nowamanqamu, kodwa nakulokhu kulungiselelwa ayashesha ukwedlula ifulegi le-NVIDIA elikhona ngokufika ku-19-40%, ngokusekelwe emiphumeleni ye-Geekbench. Akunakwenzeka ukuthi le mikhiqizo ye-NVIDIA enenkumbulo ye-HBM2 ithole indlela eya ezinhlelweni zabathengi, ngakho-ke lesi siqhathaniso sifaneleka ukwenza ngokuyinhloko ngombono wokuqondanisa kwamandla engxenyeni yeseva.

Source: 3dnews.ru