Jensen Huang, CUDA da hawan NVIDIA
Gabatarwa
Lokacin da Jensen Huang ya kafa NVIDIA tare da wasu a shekarar 1993, manufarta ita ce kawo zane-zanen 3D zuwa wasannin kwamfuta da kafofin watsa labarai. A yau, ana amfani da sarrafa bayanan wannan kamfanin don horarwa da gudanar da samfurori na hazaka ta wucin gadi (AI). Don fahimtar wannan canji, ya fi kyau a fara da tarihin aikin injiniya na Huang da dandalin software da NVIDIA ya gina, CUDA. [1] [1]
1. Aikin injiniya na Huang
An haifi Huang a Taiwan, ya samu digirin farko a fannin injiniyan lantarki daga Jami'ar Jihar Oregon a shekarar 1984. Ya yi aiki a matsayin mai zana microprocessor a AMD daga 1984 zuwa 1985, sannan ya rike mukamai da dama a LSI Logic daga 1985 zuwa 1993. Ya zama darektan Coreware, sashen da ke da alhakin dabarar kamfanin ta haɗa abubuwa da dama na tsarin kwamfuta a kan guntu ɗaya. A cikin wannan lokaci, ya kuma samu digiri na biyu a fannin injiniyan lantarki daga Jami'ar Stanford a shekarar 1992. [2] [1]
A ranar 5 ga Afrilu, 1993, Huang ya kafa NVIDIA tare da Chris Malachowsky da Curtis Priem. Ya yi aiki a matsayin shugaba, babban jami'in gudanarwa da mamba a hukumar gudanarwa tun lokacin da aka fara kamfanin. A shekarar 1999, NVIDIA ta gudanar da fara sayar da hannayen jarin jama'a (IPO). A wannan shekarar, ta gabatar da GeForce 256, wanda ya zama mahimmin abin tarihi a tarihin na'urorin sarrafa zane (GPU) na kamfanin. [1] [1]
2. Menene CUDA?
NVIDIA ta gabatar da CUDA a shekarar 2006 a matsayin dandalin software da tsarin shirye-shirye don gudanar da lissafi da yawa a lokaci guda a kan GPU nata. GPU ita ce kayan aikin da ke gudanar da lissafi. CUDA tana ba da kayan aikin shirye-shirye da masu haɓakawa ke amfani da su don gudanar da lissafinsu a kan wannan kayan aikin. Wannan ya buɗe GPU da aka ginawa don zane zuwa aikace-aikacen binciken kimiyya, kwamfuta mai ƙarfi da hazaka ta wucin gadi. [1] [1]
Ka yi tunanin cewa kana haskaka hoto mai ɗauke da piksal miliyan ɗaya. Lissafin piksal ɗaya na iya ci gaba da kansa ba tare da na sauran ba. Shirin CUDA na iya raba wannan aikin zuwa ƙananan lissafi kuma ya gudanar da su a kan GPU. Ana kiran aikin da ke gudana a kan GPU da suna kernel. Threads da dama na gudanar da shi a lokaci guda, kowannensu yana aiki a kan ɓangaren bayanai da aka ba shi. Wannan ka'ida ta raba aiki zuwa sassa da za a iya gudanarwa a lokaci guda tana amfani da lissafin AI kamar haɗa manyan ma'auni (matrix multiplication). [3] [1]
3. Yadda CUDA ta taimaka wa masana'antar AI
Lokacin horar da hanyar sadarwa ta jijiya (neural network), samfurin yana fara yin hasashe. Tsarin horo yana auna kuskure kuma yana sabunta sigogin samfurin. Yana maimaita wannan tsari sau da yawa. GPU na saurin ayyuka kamar haɗa ma'auni, suna bawa masu bincike lokaci don gudanar da ƙarin gwaje-gwaje kuma suna sa horar da manyan samfurori ya zama abu da za a iya yi. Haka kuma ana amfani da GPU bayan horo, yayin inference, lokacin da samfurin kamar samfurin harshe ke samar da amsa ga mai amfani. [4] [1]
AlexNet ta ba da shaida ta aikace-aikace a shekarar 2012. Bayan horar da ita a kan GPU na NVIDIA, ta lashe gasar rarrabe hotuna ta ImageNet. Aikinta ya nuna yadda za a iya amfani da sarrafa zane da ake amfani da su a wasannin kwamfuta don horar da samfurori da ke gane hotuna. [1] [1]
4. Yadda CUDA ta ba NVIDIA fa'ida mai dawwama
NVIDIA ta gina dakunan karatu (libraries) da kayan aiki da aka inganta don lissafi da masu haɓakawa ke buƙata akai-akai. cuBLAS na saurin ayyukan ma'auni, cuDNN tana goyon bayan ayyukan hanyar sadarwa ta jijiya na yau da kullun, NCCL tana kula da sadarwa tsakanin GPU da dama, kuma TensorRT tana taimaka wa samfurori da aka horar su gudana da sauri. Tsarin kamar PyTorch na goyon bayan CUDA, don haka masu haɓakawa na iya amfani da waɗannan ƙarfin ba tare da rubuta kowanne aikin GPU daga farko ba. [5] [1]
Ƙarshen magana
Wannan yanayin software ya ƙarfafa kasuwancin kayan aiki na NVIDIA. Ƙungiya na iya komawa daga ƙaramin gwaji zuwa tsarin da ke gudana a kan GPU da dama yayin da take ci gaba da amfani da software da kayan aiki da ta saba da su. Canza kayan aiki na iya buƙatar ƙungiya ta sake dubawa buƙatunta na CUDA, ta daidaita wasu sassa na software nata kuma ta sake gwada aiki. Ta hanyar zuba jari a software tare da kayan aiki shekaru da shekaru, NVIDIA tana da kayan aiki a shirye don masu haɓakawa waɗanda sun riga sun san yadda za su yi amfani da su lokacin da buƙatar kwamfuta ta AI ta ƙaru. [1] [1]
Majiya
- https://www.nvidia.com/en-us/about-nvidia/corporate-timeline/
- https://s201.q4cdn.com/141608511/files/doc_financials/annual/2013/NVDA_2013AR_Proxy_10K-FINAL.pdf
- https://docs.nvidia.com/cuda/cuda-programming-guide/02-basics/writing-cuda-kernels.html
- https://docs.nvidia.com/cuda/cuda-programming-guide/index.html
- https://developer.nvidia.com/machine-learning
Tambayoyi game da Jensen Huang, CUDA da hawan NVIDIA
Menene CUDA kuma me ya sa NVIDIA ta ginata?
CUDA dandalin software ne da tsarin shirye-shirye da NVIDIA ta gabatar a shekarar 2006 don gudanar da lissafi da yawa a lokaci guda a kan GPU nata. GPU ita ce kayan aikin; CUDA tana ba da kayan aikin shirye-shirye, waɗanda suka buɗe kayan aikin zane ga binciken kimiyya, kwamfuta mai ƙarfi da AI.
Yadda CUDA ta taimaka wa masana'antar AI?
Horar da hanyar sadarwa ta jijiya na nufin maimaita hasashe, auna kuskure da sabunta sigogi sau da yawa. GPU na saurin ayyuka kamar haɗa ma'auni, suna barin masu bincike su gudanar da ƙarin gwaje-gwaje kuma su horar da manyan samfurori. AlexNet, wadda aka horar a kan GPU na NVIDIA, ta tabbatar da wannan ta hanyar lashe gasar ImageNet ta 2012.
Me ya sa CUDA take ba NVIDIA fa'ida mai dawwama?
NVIDIA ta gina dakunan karatu da aka inganta waɗanda masu haɓakawa ke buƙata akai-akai – cuBLAS don ayyukan ma'auni, cuDNN don ayyukan hanyar sadarwa ta jijiya, NCCL don sadarwar GPU da dama da TensorRT don inference mai sauri – kuma tsarin kamar PyTorch na goyon bayan CUDA, don haka ƙungiyoyi na iya faɗaɗa daga ƙaramin gwaji zuwa GPU da dama ba tare da sake rubuta software nasu ba.
Powered By K7