Résumé
Memoire Bandwidth bu kawe (HBM) mooy mémoire buñ def ci wetu GPU biy joxe done yi gëna gaaw ci RAM yi. Mooy luy dundal gaawaayu IA yi, di tere core ordinatër yu am doole yi toog rek di xaar diisaayu model yi ak done yi.
Memoire bu am yaatuwaayu bande bu kawe ab bloku tabax la bu am njeexital ci kalite model bi, njëgu infrastructure bi, latency bi, ak wóor ci escale bi.
Plongeur bu xóot
HBM saafara na jafe-jafe bu yomb: puce IA yu bees yi mën nañu def ay trillion ci operation ci segond bu nekk, waaye sudee done yi yegsi nañu ci lu gaaw. Memoire GDDR buñ miin dafay boole ci bus bu sew, waaye HBM dafay boole DRAM yu bari ci anam wu taxaw ba noppi boole leen ci ay junni fiil yu ndaw yu taxaw yu ñuy woowe through-silicon vias (TSVs). Stack yooyu ñu ngi toog ci kaw interposer silicon millimetre ci GPU bi, di joxe yoonu done bu yaatu lool, xalaat ay junni bit ci benn yoon ci barabu téemeeri. Lépp soo ko boolee mu am yaatuwaayu band bi ñuy natt ci terabyte ci segond bu nekk. Ay jamono joge ci HBM2 dem ba ci HBM2e, HBM3, ak HBM3e, bu nekk di yokk kàttan ak gaawaay. Ci modelu làkk yu mag yi, te dañu wara streaming seeni poid saa yu nekk, kàttanu HBM ak yaatuwaayu bandwidth dañu gëna am solo ci ordinatër bu ñor.
Gis-gis xarala
HBM defay yegg ci gaawaayam ci parallelism bu tar te du ci montor yu gëna kawe. Sooy stack DRAM dies yi nga boole leen ak ay junni TSV, dafay wane interface bu yaatu lool (1024 bits ci stack bu nekk ak ci kaw), kon byte yu bari dañuy toxu benn yoon. Teg stack yi ci interposer buñ bokk ci wetu GPU dafay tax fiil yi di gaawa gàtt, dagg doole ci bit bu nekk ak latency. Benn gaawaay bu melni NVIDIA H100 wala H200 dafay boole ay stack HBM yu bari ngir yegg ci terabyte yu bari ci segond bu nekk ci bandwidth mémoire bi yépp.
Jàppale mémoire bu am bandwidth bu kawe
Memoire Bandwidth bu kawe (HBM) mooy mémoire buñ def ci wetu GPU biy joxe done yi gëna gaaw ci RAM yi. Mooy luy dundal gaawaayu IA yi, di tere core ordinatër yu am doole yi toog rek di xaar diisaayu model yi ak done yi. Memoire bu am yaatuwaayu bande bu kawe ab bloku tabax la bu am njeexital ci kalite model bi, njëgu infrastructure bi, latency bi, ak wóor ci escale bi. Ngir tabax xam-xam bu xóot, jàppal Memoire Bandwidth bu kawe ni xeetu liggéey, du benn man-man: leeral njariñ yi nga bëgg, leeral xalaat yi, ba noppi tàqale li sistem bi mëna def ci anam wu wóor ak li ba leegi soxla àtteb kàngam.
Ci jëf, ekip yu am doole yiy jëfandikoo Memoire Bandwidth bu kawe dañuy gëna baaxal architecture, done, ak tànneefi infrastructure ci wàllu wóor ak njëg. Dañuy bind kritër yu leer ngir am ndam, natt leen ci done yu dëggu ak def liggéey, ba noppi ñu baamtu ci anamu ñàkka mëna seetlu, du ci benn yoon benchmark wins. Mooy barab bi xam-xam theorie bi di soppiku nekk kàttan buy yàgg ci produit yi, ci politik yi ak ci liggéey yi.
Dogal yi architecture di jël dañuy indi njariñ ak njëgu liggéey bi ay at ci ginaaw. Ci jamano jooju, Optimisation benn benchmark mën na nëbb ñakk kattan yu gëna yaatu ci sistem bi. Xeetu jëf bi gëna dëgër mooy boole gaawaayu jàngat ak disipline nguur: doxal pilote, jàpp firnde, siiwal dogal yi, ak wéy di yeesal kaaraange gi ci anam wi ñuy doxalee, li jëfandikukat bi di xaar, ak sàrti sàrt yi di jëm kanam.
njeextalu pexe
Dogal yi architecture di jël dañuy indi njariñ ak njëgu liggéey bi ay at ci ginaaw.
Dogal yi architecture di jël dañuy indi njariñ ak njëgu liggéey bi ay at ci ginaaw. Ci jëfandikoo yu am kalite bu kawe, loolu dañu koy tekki ci sàrti liggéey yuñ mëna natt, ay peggu boroom, ak ay xew-xewu xoolaat yu bari suko defee ekip yi mëna yokk wóolu seen bopp ci barabu yokk lu jaxasoo.
Njàngalem xarala yi dafay jàppale ekip yi ñu tànn li gën, te baña yam ci li gëna bees daal.
Njàngalem xarala yi dafay jàppale ekip yi ñu tànn li gën, te baña yam ci li gëna bees daal. Ci jëfandikoo yu am kalite bu kawe, loolu dañu koy tekki ci sàrti liggéey yuñ mëna natt, ay peggu boroom, ak ay xew-xewu xoolaat yu bari suko defee ekip yi mëna yokk wóolu seen bopp ci barabu yokk lu jaxasoo.
Tanneef yu gëna baax ci wàllu ingeñër dina wàññi jafe-jafe yi ci wàllu wóor ci liggéey bi.
Tanneef yu gëna baax ci wàllu ingeñër dina wàññi jafe-jafe yi ci wàllu wóor ci liggéey bi. Ci jëfandikoo yu am kalite bu kawe, loolu dañu koy tekki ci sàrti liggéey yuñ mëna natt, ay peggu boroom, ak ay xew-xewu xoolaat yu bari suko defee ekip yi mëna yokk wóolu seen bopp ci barabu yokk lu jaxasoo.
Doxal ci àdduna dëgg
Teg fukki wala téemeeri gigabytes ci diisaay ngir modelu làkk bu mag bu jege GPU bi suko defee ñu mëna streaming ci bépp jéego bu jëm ci inference.
Fexe ba NVIDIA H100 ak H200 GPU yi mëna yegg ci terabyte yu bari ci segond bu nekk ci yaatuwaayu mémoire ngir tàggat.
Xootal clusters tàggat IA fu GPU yu bari ku nekk di wéeru ci HBM ngir moytu taxaw ci diggante liggéey matrix yi.
Jàppale nataal yu am resolusioŋ bu kawe ak modeli wideo yu wara toxal tensor yu mag yuy tàmbali ak genne ci mémoire bi ci lu gaaw.
Modèlu jëfandikoo
Memoire bu yaatu ci jëf
Teg fukki wala téemeeri gigabytes ci diisaay ngir modelu làkk bu mag bu jege GPU bi suko defee ñu mëna streaming ci bépp jéego bu jëm ci inference.
Teg fukki wala téemeeri gigabytes ci diisaay ngir benn modelu làkk bu mag bu jege GPU suko defee ñu mëna streaming ci bepp jéego bu ñuy jël. Ekip yi dañuy faral di am njariñ yu gëna baax suñu joxee thresholds yu baax ci kanam, tëye yoonu escalation nit ngir jafe-jafe yi, ak topp njuréefi produit yi ci diir bi ak e.
Memoire bu yaatu ci jëf
Fexe ba NVIDIA H100 ak H200 GPU yi mëna yegg ci terabyte yu bari ci segond bu nekk ci yaatuwaayu mémoire ngir tàggat.
Fexe ba NVIDIA H100 ak H200 datacenter GPUs yegg ci terabytes yu bari ci segond bu nekk ci bandwidth memory ngir tàggat ekip yi dañuy faral di am njariñ yu gëna baax suñu joxee threshold yu baax ci kanam, tëye yoonu escalation nit ngir jafe-jafe yi, ak topp njuréefi liggéey ak njëgu njuumte ci diir bi.
Memoire bu yaatu ci jëf
Xootal clusters tàggat IA fu GPU yu bari ku nekk di wéeru ci HBM ngir moytu taxaw ci diggante liggéey matrix yi.
Powering IA training clusters fu GPUs yu bari ku nekk di wéeru ci HBM ngir moytu taxaw ci diggante matrix operations Teams yi dañuy faral di am njariñ yu gëna baax suñu joxee threshold yu baax ci kanam, tëye yoonu escalation nit ngir jafe-jafe yi, ba noppi topp njariñu produit ak njëgu njuumte ci diir bi.
Memoire bu yaatu ci jëf
Jàppale nataal yu am resolusioŋ bu kawe ak modeli wideo yu wara toxal tensor yu mag yuy tàmbali ak genne ci mémoire bi ci lu gaaw.
Jàppale nataal yu am resolusioŋ bu kawe ak xeetu wideo yu wara toxal tensor yu mag yi ci biir ak biti ci mémoire ci lu gaaw. Ekip yi dañuy faral di am njariñ yu gëna baax suñu joxee threshold yu baax ci kanam, tëye yoonu escalation nit ngir mbir yu am solo, ak topp njariñu produit ak njëgu njuumte ci diir bi.
Risk yi ak balustrade yi
Optimize benn benchmark mën na nëbb ñakk kattan yu gëna yaatu ci sistem bi.
Njëg li ñuy fay ci infrastructure yi ak ci toppatoo dañuy faral di suufeel.
Bu sistem yi di gëna xawa jafee xam, jafe-jafe yi am ci wàllu kaaraange ak seetlu mën nañu gëna bari.
Roadmap ngir samp gi
Mandargal latency, kalite, ak njëg yi laata ngay jëfandikoo.
Mandargal latency, kalite, ak njëg yi laata ngay jëfandikoo. Japp jéego bu nekk ni buntu firnde: sudee mattul kritër yi, noppali génne gi, tëj bërëb bi, ba noppi nga yaatal jëfandikoo gi.
Benchmark ci biir sargal ak done yu dëggu.
Benchmark ci biir sargal ak done yu dëggu. Japp jéego bu nekk ni buntu firnde: sudee mattul kritër yi, noppali génne gi, tëj bërëb bi, ba noppi nga yaatal jëfandikoo gi.
Jumtukaay bi di saytu njuumte yi, derive bi ak njeextalu jëfandikukat bi.
Jumtukaay bi di saytu njuumte yi, derive bi ak njeextalu jëfandikukat bi. Japp jéego bu nekk ni buntu firnde: sudee mattul kritër yi, noppali génne gi, tëj bërëb bi, ba noppi nga yaatal jëfandikoo gi.
Waajal rollback ak yooni tontu ci jafe-jafe yi laata ngay eskale.
Waajal rollback ak yooni tontu ci jafe-jafe yi laata ngay eskale. Japp jéego bu nekk ni buntu firnde: sudee mattul kritër yi, noppali génne gi, tëj bërëb bi, ba noppi nga yaatal jëfandikoo gi.