Yaya ne kamar yadda DDR Memory da aka bukatar su a Multi-Camera AI Vision System? (Daga cikin RK3576 Selection Table)

Tare da saurin ci gaban aikace-aikacen hangen nesa na AI, an yi amfani da hanyoyin magance kyamarori da yawa (sitiriyo, kyamarori huɗu da ma ƙarin tashoshin kyamara) a cikin al'amuran kamar tattara bayanai, sa ido kan tsaro, dubawa na masana'antu, fahimtar mutum-mutumi da ƙididdigar Yawancin ayyukan suna fada cikin matsala ta yau da kullun yayin zaɓar kayan aiki: mai da hankali ne kawai kan ƙarfin sarrafa guntu yayin da yake watsi da ƙirar ƙwaƙwalwar ajiya. Wannan a ƙarshe yana haifar da matsaloli ciki har da asarar firam, ƙaddamar da AI da aka yi da kuma tsarin tsarin.

Shin ƙwaƙwalwar ajiyar DDR mafi girma koyaushe ya fi kyau? Haka ne, daga ra'ayi na aiki da kwanciyar hankali. Duk da haka, rashin daidaito yana da mahimmanci wuya a kan walat, musamman ma da karuwar farashin DDR. Daidaita farashin da ainihin bukatun aikin ya zama mahimmanci don zaɓar mafita.

Bari mu dubi mafi yawan tarkon a DDR sizing: kullum kuma yana iya zama ƙarami a cikin halin da aka saba, amma yana iya tafiyar da kwayoyin da suka dace da saurin ayyukan da suka gudan, wanda ya saita kankara na sistem. Mahailloli mai zuwa suka yi amfani da karamin kuma ta hanyar fahimtar hardware. Don misalai, AR0234 sensor wanda aka amfani da shi da yawa: kamar yadda ake fahimta, kamera 1080P ya yi amfani da karamin kuma ta yawa a lokacin da ke cikin halin da aka saba. Amma, lokacin da ke aiki da kwayoyin da suka dace da saurin ayyukan da suka gudan—kamar yadda ake fahimta, kamar yadda ake fahimta, kamar yadda ake fahimta, kamar yadda ake fahimta—karamin kuma ta yawa ta iya tafiyar da tens MB per channel. Lokacin da ke aiki da kamera 4K, karamin kuma ta yawa per channel zai iya tafiyar da 100 MB. Kamar yadda ake fahimta, HTTP request mai zuwa ya yi amfani da karamin kuma ta yawa da tens MB per connection. Wadannan karamin kuma ta yawa suna iya zama ƙarami, amma suna iya tafiyar da kwayoyin da suka dace da saurin lokacin da kamar yadda ake fahimta da kamar yadda ake fahimta suna aiki da yawa. Wannan zai iya saita OOM (Out-of-Memory), wanda zai saita kankara na ayyukan da suka kill da sistem.

Ina karamin kuma ta yawa ta iya tafiyar da kwayoyin da suka dace da saurin ayyukan da suka gudan?

Kudin kuma a cikin ƙarfin kuma na amfani da AI ba su daga yadda ya kamata a yi amfani da su ba. Su ke amfani da su wajen gano hoto, sauya hoto, amfani da AI, ayyukan ƙasa da ayyukan bayanai. Kudin kuma ya ziyar da kowane kwayoyin hoto da aka shiga.

1. Kudin Hoto (Mafi Yawa Na Kudin Kuma) Hoto mai zuwa, rukuni na YUV da hoto mai zuwa da aka sauya daga kowane kwayoyin hoto suna yanka kudin kuma. Don AR0234 da ke shiga a 1080P@30fps, wani hoto na YUV ya zama mafi yawa da MB. Don hana hoto mai zuwa da ke fuskaci, al'ummar yana amfani da kudin hoto na ring buffer, wanda ya ziyar da kudin kuma. Aiki na hoto biyu (original, thumbnail, cropped frames) yana ziyar da kudin hoto. Kudin kuma ya ziyar da kowane kwayoyin hoto da suke shiga ta hanyar kuma.

2. Kudin ISP da Kudin Sauna Hoto ISP mai saurin kayan aiki ne mai zaman kansa wanda ke aiwatar da rage amo, gyaran launi, gyaran karkatarwa, juyawa, yankewa da sauya tsari. ISP kawai yana da iyakantaccen ƙwaƙwalwar ajiya a kan guntu ba tare da babban ajiyar firam ba. Saboda haka dole ne a sanya buffers na wucin gadi a cikin DDR don ayyukan da yawa. Ƙarin tashoshin kyamara yana nufin nauyin aiki na ISP da kuma buƙatar ƙarfin DDR.

3.Memory don AI Model Inference AI inference yana cinye ƙwaƙwalwar ajiya a sassa biyu. Na farko shine nauyin samfurin: manyan sigogin samfurin suna haifar da amfani da ƙwaƙwalwar ajiya mafi girma. Abin farin ciki, ana ɗora nauyin sau ɗaya a farkon samfurin; ƙwaƙwalwar ajiyar su ba ta girma tare da ƙarin tashoshin kyamara ko al'amuran yayin gudu.

Yana kuma yana da tsarrawa a lokacin gudanar da bayani? Ba ta hanyar kamar yadda zai iya. Wannan ya sa mu zuwa babban abu na biyu: bayanin wani abu na farko wanda ke samu a lokacin gudanar da bayani. Kullum jihun da kuma kuma yana bukata wani karamin gudanar da bayani, wanda ke saita karamin RAM. Gudanar da bayani na AI da kuma yana bukata karamin gudanar da bayani na farko da kuma daga cikin jihun da kuma. Don haka, idan aka baraya karamin jihun da kuma, karamin RAM na gudanar da bayani zai taimaka.

4. Sistam, Draiwa da Midilwea Tsarrawa Don fahimtar, za mu yi misalai na RK3576 vision solution. Karamin sistam na farko ana amfani da shi ne da Linux kernel, RK media drivers, RKMPP media framework, NPU drivers, system logs da kuma ayyukan da suka run. A lokacin da kuma akwai mahaillan da kuma, media frameworks da kuma video encoding (H.264 / H.265 recording) suna taimaka karamin RAM. Idan aka rufe multi-channel encoding da kuma recording, karamin RAM zai taimaka sosai.

Bayani mai mahimmanci: Wannan yana nufin karamin aiki (DDR RAM), ba eMMC storage ba. eMMC yana nufin karamin da suka store shigar da kuma video files.

5.Aikace-aikace - Layer Kasuwancin Memory Maganar aikace-aikacen ciki har da sakamakon sakamako, bin diddigin abubuwa da yawa, rahoton bayanai, caching na gida da ajiyar hoto yana cinye ƙwaƙwalwar ajiya. Alal misali, algorithms na bin sawu suna adana bayanai masu yawa. Ƙarin kyamarori da abubuwan da aka bi suna haifar da manyan bayanai da kuma amfani da ƙwaƙwalwar ajiya mafi girma.

Ayyukan ƙwaƙwalwar ajiya na aiki akan dandamali na RK3576

RK3576 babban mai sarrafa AI ne na gida wanda ke ba da 6 TOPS na aikin NPU, wanda ya dace da aikace-aikacen hangen nesa na AI da yawa. Kamfaninmu yana da samfuran samar da kayan masarufi da yawa waɗanda ke tallafawa tashoshin kyamara 1 zuwa 6 don dacewa da buƙatun aikin daban-daban. 4 GB DDR shine babban tsari na samar da taro, yayin da ake tallafawa madadin girman DDR don nau'ikan aiki daban-daban.

Tebur da ke ƙasa ya shafi na'urori masu auna na'urori masu auna 1080p-4K a kasuwa:

Idan ba ka da tabbacin yadda za ka saita ƙwaƙwalwar ajiyar da ta dace, sai ka tuntuɓi [email protected]. Za mu samar da daidaitaccen bayani kimantawa bisa ga yawan tashar kyamarar ku, ƙuduri, ma'anar kasuwanci da aiki na aiki.