Kunye neprosesa entsha engundoqo I-ARM yazisa iprosesa yemizobo eyenzelwe isizukulwana esilandelayo senkqubo yetshiphu enye. I-Mali-G77, engafanele kudideke kunye neprosesa entsha yokubonisa , iphawula utshintsho ukusuka kwi-ARM Bifrost architecture ukuya eValhall.
I-ARM ivakalisa ukwanda okubonakalayo kwimizobo ye-Mali-G77 - nge-40% xa kuthelekiswa nesizukulwana sangoku seMali-G76. Oku kwaphunyezwa ngenkqubo yobugcisa kunye nokuphuculwa koyilo lwezakhiwo. I-Mali-G77 ingaba ne-7 ukuya kwi-16 cores (ukunyuka ukusuka kwi-1 ukuya kwi-32 kunokwenzeka kwixesha elizayo), kwaye ngamnye wabo uphantse ulingana ne-G76. Ngenxa yoko, ii-smartphones zokugqibela ziya kuba nenani elifanayo lee-GPU cores.
Kwimidlalo, unokulindela ukuphuculwa komsebenzi phakathi kwe-20 kunye ne-40%, kuxhomekeke kuhlobo lomsebenzi wegraphic. Ngokujonga iziphumo zovavanyo oludumileyo lweManhattan GFXBench, ukongama okubalulekileyo kweGPU entsha kwisizukulwana sangoku kuya kunyanzela imbangi iQualcomm ukuba ixhalabe malunga nophuculo olubonakalayo ekusebenzeni kwemizobo yeAdreno.
Ngokwayo, ulwakhiwo olutsha lweMali-G77 lunikezela nge-avareji yeepesenti ezingama-30 ekuphuculeni amandla okanye ukusebenza kakuhle, utshilo i-ARM. Isizukulwana sesibini se-ARM Valhall scalar architecture ivumela i-GPU ukuba yenze imiyalelo ye-16 kumjikelo ngamnye ngokuhambelana ne-CU, xa kuthelekiswa nesibhozo kwi-Bifrost (Mali-G76). Olunye utshintsho lubandakanya ukucwangciswa komyalelo oqhutywa ngokupheleleyo wehardware kunye neseti entsha ngokupheleleyo yomyalelo ngelixa ugcina ukuhambelana nomva kunye neBifrost. Inkxaso yefomathi yoxinzelelo ye-ARM AFBC1.3 kunye nezinye izinto ezintsha (i-FP16 yonikezelo ekujoliswe kuyo, unikezelo olunemaleko kunye neziphumo ze-vertex shader) nazo zongeziwe.
I-Bifrost CU iqulethe iinjini ze-3 zokubulawa, nganye kuzo iquka i-cache yomyalelo, irejista, kunye neyunithi yokulawula i-Warp. Ukusasazwa kuzo zonke ezi njini zintathu kuvumele ukuba imiyalelo ye-24 ye-FMA iphunyezwe kwi-32-bit yokuchaneka kwendawo edadayo (FP32). E-Valhall, i-CU nganye ine-injini yokubulawa enye kuphela, yahlulahlulwe phakathi kweeyunithi ezimbini ze-computing ezikwazi ukucubungula imiyalelo ye-Warp ye-16 ngewotshi nganye, okukhokelela ekuphumeni okupheleleyo kwe-32 ye-FMA FP32 imiyalelo nge-CU nganye. Ngombulelo kolu tshintsho lwezakhiwo, iMali-G77 inokwenza isinye kwisithathu ukubala kwezibalo kwizibalo ezifanayo xa kuthelekiswa neMali-G76.
Ukongeza, nganye kwezi CUs iqulethe iibhloko ezimbini ezintsha zemathematika. I-injini entsha yokuguqula (CVT) ilawula inani elipheleleyo, ingqiqo, isebe, kunye nemiyalelo yokuguqula. IYunithi yoMsebenzi oKhethekileyo (SFU) ikhawulezisa uphinda-phindo olupheleleyo, ulwahlulo, i-square root, iilogarithms, kunye neminye imisebenzi edityanisiweyo epheleleyo.
Ibhloko ye-FMA eqhelekileyo inezicwangciso ezininzi ezixhasa imiyalelo ye-16 FP32 kumjikelo ngamnye, i-32 ye-FP16, okanye i-64 ye-INT8 Dot Product. Olu lungiso lunokubonelela ukuya kuthi ga kwi-60% yokuphucula ukusebenza kwizicelo zokufunda koomatshini.
Olunye utshintsho olungundoqo kwi-Mali-G77 kukuphinda-phinda ukusebenza kwe-injini ye-texture, ngoku iqhuba ii-texel ezi-4 ze-bilinear ngewotshi nganye xa kuthelekiswa nezimbini zangaphambili, ii-texel ezi-2 ze-trilinear ngewotshi nganye, eyenza ukuba i-FP16 ikhawuleze ukuhluza kunye ne-FP32.
I-ARM yenze ezinye iinguqu ezininzi, kunye neMali-G77 kunye neValhall zithembisa ukuphuculwa komsebenzi obalulekileyo womdlalo kunye nomsebenzi wokufunda ngomatshini. Okubalulekileyo, ukusetyenziswa kwamandla kunye nendawo ye-chip igcinwa kumanqanaba e-Bifrost, ethembisa izixhobo eziphathwayo ezinomgangatho ophezulu wokusebenza ngaphandle kokunyuswa kokusetyenziswa kwamandla, ukutshatyalaliswa kobushushu kunye neemfuno zobukhulu.
umthombo: 3dnews.ru