Binciken Lissafi Na Gaba Da Juna Na GPU Akan Matlab

Teburin Abubuwan Ciki

• 1. Gabatarwa
• 2. Tsarin GPU
  • 2.1 Kwatanta GPU da CPU
  • 2.2 Fa'idodin GPU
  • 2.3 Shirye-shiryen Da Suka Dace Da Lissafin GPU
• 3. Hanyar Gwaji
• 4. Sakamako da Bincike
• 5. Tsarin Fasaha
• 6. Ayyuka Na Gaba
• 7. Nassoshi

1. Gabatarwa

Ana amfani da Matlab sosai a cikin lissafin kimiyya amma yana fama da ƙarancin ingancin lissafi idan aka kwatanta da yaren C. Wannan takarda tana binciken hanzarin GPU ta hanyar Akwatin Kayan Aikin Lissafi Na Gaba Da Juna na Matlab don haɓaka aiki ba tare da buƙatar haɓaka kayan aiki ko sake rubuta code ba.

2. Tsarin GPU

An ƙera tsarin GPU don sarrafa ayyuka tare, yana da ɗimbin raka'o'in aiwatarwa waɗanda aka inganta don ayyukan bayanai masu kama da juna.

2.1 Kwatanta GPU da CPU

GPUs suna amfani da ƙarin masu canza wutar lantarki don raka'o'in aiwatarwa maimakon sarrafa dabaru, suna ba da damar yin ayyuka masu yawa tare amma an rage inganci don ayyuka na bi da bi.

2.2 Fa'idodin GPU

Manyan fa'idodi sun haɗa da mafi girman aikin ma'auni mai iyo da kuma bandeji na ƙwaƙwalwar ajiya. GPU na yanzu suna samun bandeji na 40-142 GB/s idan aka kwatanta da 32 GB/s na ƙwaƙwalwar DDR3.

2.3 Shirye-shiryen Da Suka Dace Da Lissafin GPU

Mafi kyawun aikace-aikacen GPU sune masu zurfin lissafi, masu yawan aiki tare, sun haɗa da ayyuka masu sauƙi, kuma suna sarrafa manyan bayanai.

3. Hanyar Gwaji

Gwaje-gwajen da aka gudanar sun haɗa da FFT, ninka matrix, quicksort, da kuma kwaikwayon code na Hamming a cikin tashar BSC. An auna aikin ta amfani da ma'aunin hanzari: $Hanzari = rac{T_{CPU}}{T_{GPU}}$

4. Sakamako da Bincike

GPU ya nuna gagarumin haɓaka don ayyukan gaba da juna: sau 15 don babban ninka matrix ($2048 imes 2048$), sau 8 don FFT. Duk da haka, ayyukan ma'ana sun kasance sau 2-3 a hankali akan GPU.

5. Tsarin Fasaha

Hankalin Asali: Wannan bincike ya fallasa ainihin ciniki a cikin lissafin GPU - ƙarfin gaba da juna na danye da iyakokin dabaru na bi da bi. Marubutan sun gano daidai cewa hanzarin GPU ba maganin gama gari ba ne amma kayan aiki na musamman.

Kwararar Ma'ana: Takardar ta bi hanyar gwaji bayyananna: gano nau'ikan lissafi → aiwatar da kwatancin CPU/GPU → nazarin alamu na aiki. Wannan hanyar tana nuna inda jarin GPU ya yi amfanin gaske.

Ƙarfi & Kurakurai: Ƙarfin yana taƙaice a cikin ingantaccen tabbatarwa a cikin ayyuka daban-daban. Duk da haka, binciken ba shi da zurfi a cikin nazarin matsayi na ƙwaƙwalwar ajiya kuma bai magance sabbin tsare-tsaren GPU kamar NVIDIA's Tensor Cores waɗanda zasu iya canza yanayin aikin ba.

Hankali Mai Aiki: Ya kamata masu bincike su yi bayanin aikace-aikacen don abun ciki na gaba da juna kafin aiwatar da GPU. Don ayyukan aiki gauraye, hanyoyin haɗin gwiwar CPU-GPU (kamar yadda aka gani a cikin tsarin shirye-shiryen CUDA na NVIDIA) sau da yawa suna haifar da sakamako mafi kyau.

Bincike Na Asali

Wannan bincike yana ba da ingantaccen shaida na zahiri ga fagen da ke girma na lissafin kimiyya mai hanzari na GPU. Binciken ya yi daidai da ka'idoji da aka kafa a cikin tsarin lissafi na gaba da juna, musamman Dokar Amdahl wacce ta bayyana cewa matsakaicin hanzari yana iyakance ta ɓangaren bi da bi na shiri. Haɓakar saurin 15 don ayyukan matrix yana nuna yuwuwar lissafin GPU don ayyukan algebra na layi, kama da ribar aikin da aka ruwaito a cikin takardun cuBLAS na NVIDIA. Duk da haka, rashin aikin aiki akan ayyukan ma'ana ya nuna ainihin iyakar gine-gine - GPUs suna yi wa ayyukan bayanai masu kama da juna amma suna fama da ayyukan sarrafawa masu yawa. An rubuta wannan rarrabuwar kawuna a cikin aikin farko "Demystifying GPU Microarchitecture Through Microbenchmarking" na Wong et al. (IEEE Micro 2010). Binciken zai amfana daga kwatanta tare da ƙarin ci gaba na baya-bayan nan kamar ROCm na AMD da ƙaddamarwar oneAPI na Intel waɗanda ke ba da mafita na lissafin GPU na dandamali. Aikin gaba yakamata a bincika lissafin daidaitawa gauraye da ayyukan tensor waɗanda suka mamaye ayyukan AI na zamani, gina akan tsare-tsare kamar dlarray na MATLAB don aikace-aikacen zurfin koyo.

Misalin Tsarin Bincike

Harka: Bututun Sarrafa Hotuna
Don aikace-aikacen hoto na likita yana sarrafa yankunan MRI 1000:
• Ayyuka na gaba da juna (tace FFT): Ana ba da shawarar hanzarin GPU
• Ayyukan ma'ana (gano fasali): An fi son sarrafa CPU
• Hanyar haɗin gwiwa: Rarraba 70% GPU + 30% CPU mafi kyau

6. Ayyuka Na Gaba

Aikace-aikacen da ke tasowa sun haɗa da sarrafa siginar lokaci-lokaci, horar da samfurin AI, da manyan simintin gwargwado. Haɗin kai tare da sabis na GPU na girgije da kwantena zai ba da damar yin amfani da albarkatun lissafi masu inganci.

7. Nassoshi

1. Jagorar Shirye-shiryen NVIDIA CUDA, 2022
2. Wong, H. da saur. "Demystifying GPU Microarchitecture Through Microbenchmarking" IEEE Micro, 2010
3. Takardun Akwatin Kayan Aikin Lissafi Na Gaba Da Juna na MathWorks
4. Dandamalin Buɗe Lissafi na AMD ROCm
5. Ci gaban Tsarin Gine-gine Guda ɗaya na Intel