Ukufakela iCUDA kwiWindows akufuneki kube yingxaki ukuba uyazi ukuba ujike phi kwaye zeziphi iinguqulelo ezihambelana nazo. Kwesi sikhokelo sisebenzayo, ndiza kukuthatha inyathelo ngenyathelo kuyo yonke into oyifunayo ukuze ufumane isixhobo, umqhubi ochanekileyo, kunye nezixhobo zokuqinisekisa ezisebenzayo, zombini kwiWindows kunye nokusebenzisa iWSL. Umbono kukubulala iGPU yakho. ngenene isantya imithwalo yakho yomsebenzi, ngaphandle kokumangalisa kokuhambelana.
Ukongeza kufakelo olusisiseko, ndiza kugubungela iimeko zokusetyenziswa zehlabathi lokwenyani kunye neemfuno ezikhethekileyo (ezifana ne-ML.NET Model Builder enokuxhomekeka kwilifa), ezinye iindlela zofakelo kunye ne-Conda kunye ne-pip, ukuhlanganiswa kunye ne-Visual Studio, kunye ne-classic deviceQuery kunye novavanyo lwe-bandwidthTest. Uya kubona kwakhona indlela yokuxilonga iingxaki zomqhubi eziqhelekileyo, ukufunyanwa kwe-GPU kunye ne-CUDA yenguqulo engafanelekanga edla ngokucotha abantu okokuqala.
Yintoni iCUDA kwaye kutheni ibalulekile kwiWindows?
I-CUDA yiqonga kunye nomzekelo we inkqubo yeNVIDIA kwicomputing ehambelanayo ekuvumela ukuba ukhuphele ezona ndawo zinzulu zezicelo zakho kwiGPU. I-CPU iphethe umsebenzi olandelelanayo kwaye i-GPU iphethe ihambelana kakhulu, ngalinye linenkumbulo yalo, liphepha imiqobo kunye nokuvumela umsebenzi ngaxeshanye.
I-CUDA-enabled GPUs ibonisa amakhulu okanye amawaka eecores aqhuba amashumi amawaka eentambo. Amathala eencwadi anje cuBLAS, cuDNN kunye neNVCC Gqibezela ikhithi ukuze ukwazi ukuqulunqa, ukulungisa ingxaki, kunye neprofayile yakho yesoftware yeGPU ekhawulezileyo kwiWindows.
Iimfuno zenkqubo kunye nokuhambelana kwiWindows
Okokuqala, kufuneka ujonge ukuba isixokelelwano sakho siyahambelana nesixhobo sezixhobo ozifunayo. Kwiinkqubo zangoku zeWindows, CUDA 13.0 U1 inkxaso: Windows 11 24H2, 23H2, 22H2-SV2; Windows 10 22H2; Windows Server 2022 kunye 2025.
Abaqulunqi abaxhaswayo kunye nee-IDE: I-Visual Studio 2022 17.x (MSVC 193x) ene-C ++ 11/14/17/20 kunye ne-Visual Studio 2019 16.x (MSVC 192x) kunye ne-C ++ 11/14/17. I inkxaso ye-VS2017 Yasuswa kwi-CUDA 13.0 kunye ne-VS2015 ihoxisiwe ukususela kwi-CUDA 11.1, ngoko lungisa iiprojekthi zakho.
Inqaku elibalulekileyo le-32-bit: Ukususela kwi-CUDA 12.0, akusekho i-toolchain ye-32-bit (yendabuko okanye i-cross-platform). Abaqhubi baya kuqhubeka nokuqhuba iibhinari ze-32-bit kwiGeForce. Uyilo lweAda; IHopper ayisaxhasi iibits ezingama-32.
Izixhobo ezicetyiswayo: I-NVIDIA GPU enoKusebenza kweKhompyutha efanelekileyo kuguqulelo lwakho (kwimeko ngokubanzi, 3.0 okanye ngaphezulu; bona uluhlu olusemthethweni), inkumbulo eyaneleyo, kwaye kwiimeko ezithile ezifuna kakhulu. KuMakhi oMzekelo (ukuhlelwa komfanekiso), ubuncinci 6 GB yeVRAM ezinikezelweyo, ngelixa kusetyenziso olusisiseko ubuncinci be-4 GB bukhankanyiwe.
Jonga ukuba unayo iGPU ehambelanayo kunye nokuba yeyiphi inguqulelo onayo
Ukujonga imodeli yakho yeGPU kwiWindows: Useto > Inkqubo > Idispleyi > Useto olukwinqanaba eliphezulu. Apho uya kubona ukwenziwa kunye nemodeli phantsi koLwazi lokubonisa, iluncedo kakhulu ukuqinisekisa ukuhambelana.
Ungasebenzisa kwakhona uMphathi woMsebenzi kwithebhu yokuSebenza kwaye ukhethe iphaneli ye-GPU ukubona iinkcukacha. Ukuba ayiveli, vula i Umphathi wesixhobo kwaye ujonge kwi-Display adapters; Ukuba umqhubi we-NVIDIA ulahlekile, yifake.
Ukujonga ukufakela kweCUDA, vula iPowerShell okanye iCMD kwaye uqhube nvcc --version o nvcc -V. Oku kubuyisela uguqulelo lomqokeleli NVCC ifakiwe; ukuba ayiphenduli, i-Toolkit PATH ayimiselwanga okanye ayifakelwanga ngokuchanekileyo.
Ukuba ufuna ukuhlaziya abaqhubi, amava eGeForce acebisa iinguqulelo zamva nje, okanye uzikhuphele kwiwebhusayithi esemthethweni yomqhubi we-NVIDIA. Umqhubi ohlaziyiweyo udla ngokuwusombulula umcimbi. Ukungahambelani kwezixhobo kunye nokusilela ekubhaqweni.
Khuphela i-NVIDIA CUDA Toolkit: Iifomathi kunye neMfezeko
Yiya kwi-NVIDIA yekhasi lokukhuphela elisemthethweni le-CUDA kwaye ukhethe iWindows. Unokukhetha nokuba yiNethiwekhi yesiFakelo (ukhuphelo oluncinci, emva koko olufunwayo) okanye uMfaki ogcweleyo (yonke into ibandakanyiwe). Isifakeli esipheleleyo silungile kwiikhompyuter ezingaxhunyiwe kwi-intanethi okanye ukusasazwa kwamashishini.
Emva kokukhuphela, yinto efanelekileyo yokuqinisekisa i-MD5 checksum epapashiweyo ukuqinisekisa ukuba ifayile ayonakaliswa. Ukuba ihashi ayihambelani, khuphela kwakhona umfakeli kwaye uphephe iintloko emva kwexesha.
Ukusukela kwi-CUDA 13, umqhubi we-NVIDIA akasafakwa kwi-Toolkit. Faka umqhubi ngokwahlukileyo kwi iphepha labaqhubi beNVIDIA kwaye emva koko i-Toolkit; kuphephe ukuxuba iinguqulelo ezingahambelaniyo.
Ufakelo kwiWindows: imizobo, cwaka, kunye neendlela zokukhupha
Ufakelo lomzobo: Sebenzisa isifakeli kwaye ulandele amanyathelo. Khetha indawo, yamkele ilayisenisi, kwaye ukhethe zixhobo njenge Toolkit, Nsight Compute/Systems kunye nemizekelo.
Ukufakela okuthe cwaka: Ungaqalisa isifakeli nge -s kwimo ethe cwaka kwaye yongeza iparameters zesubpackage ezithile. Ukuba awufuni ukuqalisa kwakhona ngokuzenzekelayo, yongeza -n kwaye ulawula ukuqalisa kwakhona ekugqibeleni.
Ukutsalwa ngesandla: Nge-7-Zip/WinZip, unokukhupha yonke ipakethe ukujonga imixholo yayo. I-Toolkit ibekwe ngaphakathi kwisiqulathi seefayili. CUDA Toolkit kunye nokudityaniswa kweVisual Studio kuluhlu lwayo lwamagama; i-.dll kunye neefayile ze-nvi ozibona kwingcambu yesifakeli ayizofayile ezifakekayo ngolo hlobo.
Indlela yeSixhobo esiMiselweyo: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.0. Xa ufaka i-subpackages Unokukhetha kuphela into oyifunayo (umzekelo, nvcc_13.0 kumqokeleli, cudart_13.0 ixesha lokusebenza, cublas_13.0/cublas_dev_13.0, cufft_13.0, curand_13.0, cusolver_13.0, cusparse_13.0, nsight_compute_13.0, nsight_systems_13.0, nvrtc_13.0, nvdisasm_13.0, nvprune_13.0, opencl_13.0, sanitizer_13.0, thrust_13.0, njl). Ngale ndlela unciphisa unyawo kunye namaxesha ofakelo.
Khipha: Zonke iipakeji ezincinci zinokususwa kwiPhaneli yoLawulo> Iinkqubo kunye neMiba. Iluncedo ukuba ufuna ukuphinda ufake inguqulelo. ucocekile okanye utshintshe amasebe ngaphandle kokushiyeka.
Iimodeli zomqhubi kwiWindows: WDDM vs TCC
On Windows 10 kwaye kamva, umqhubi we-NVIDIA unokusebenza ngeendlela ezimbini: i-WDDM (yezixhobo zokubonisa) kunye ne-TCC (yokubala ii-GPU ezifana neTesla okanye iiTitans ezithile). Ungajonga kwaye utshintshe indlela nge nvidia-smi kumakhadi axhaswayo; Uninzi lwamakhadi eGeForce anamhlanje asebenzisa i-WDDM ngokungagqibekanga.
Qinisekisa ufakelo: nvcc, iisampuli, kunye neemvavanyo
Jonga inguqulelo yeNVCC nge nvcc -V. Ukuba umyalelo uyasebenza, i PATH kunye nofakelo olusisiseko lukhona.
Cola iisampulu zeCUDA ukusuka kwiGitHub e nvidia/cuda-isampulu, ziqokelele kwaye uziqhube deviceQuery. Kufuneka ibone isixhobo sakho kwaye ibuyisele a uvavanyo lupasisiwe; ukuba akunjalo, khangela umqhubi kunye ne-hardware.
Kananjalo iyabaleka bandwidthTest ukuqinisekisa i-bandwidth yesixhobo somkhosi. Ukuba iimvavanyo zidlula, unxibelelwano CPU-GPU ulungile. Naziphi na iimpazamo apha zikholisa ukukhomba abaqhubi okanye ufakelo lwe Toolkit.
Ukudityaniswa neVisual Studio kunye neziseko zeprojekthi
Iisampulu kunye neeprojekthi zakho zinokudityaniswa kunye nezisombululo ze-VS 2019/2022. Iitemplates zeNVIDIA ziseta iprojekthi yeC++ nge Yakha Customizations yeCUDA yoguqulelo lwakho (umzekelo, CUDA 13.0 Runtime).
Iifayile ze-Toolkit .props: VS 2019 ineeprops kwi C:\Program Files (x86)\Microsoft Visual Studio\2019\Professional\MSBuild\Microsoft\VC\v160\BuildCustomizations kunye neVS 2022 kwi C:\Program Files\Microsoft Visual Studio\2022\Professional\MSBuild\Microsoft\VC\v170\BuildCustomizations. Ezi ndlela zivumela i-VS ukuba ifumane imithetho kunye neempawu ukusuka eCUDA.
Kwiiprojekthi ezintsha, dala ukusuka kwiZifanekiso zeNVIDIA> CUDA kwinguqulelo efakiweyo. Kwiiprojekthi ezikhoyo, yiya ku Yakha iziXhomekeko > Yakha uLungiselelo kwaye uphawule uguqulelo lwakho lweCUDA okanye walathe kuyo $(CUDA_PATH) ukuba ufuna ukusebenzisa esanda kufakelwe enye.
Emva kokufaka okanye ukukhupha ii-Toolkits, qinisekisa oko $(CUDA_PATH) ikhomba kwisiqulathi seefayili esichanekileyo. Ukufikelela Izinto eziNgqongileyo ukusuka kwiiPropati zeNkqubo ukujonga kwaye uhlengahlengise ixabiso ukuba kuyimfuneko.
Iifayile ezinekhowudi yeCUDA mazibhalwe njengohlobo lweCUDA C/C++. Ungazongeza ukusuka Yongeza Into Entsha > NVIDIA CUDA 13.0 > CUDA C/C++ Code kwiVisual Studio.
Faka iCUDA ngeConda kunye nepip
Conda: I-NVIDIA ipapasha iipakethi kwijelo layo le-Anaconda (anaconda.org/nvidia) ukufakela iKit epheleleyo okanye iinguqulelo ezithile zangaphambili. Ukuqhobosha inguqulelo yangaphambili, yongeza i ukukhulula ileyibhile kumyalelo wokufakela kwaye iConda iya kusombulula ukuxhomekeka.
umbhobho: I-NVIDIA ibonelela ngamavili okufakela amacandelo e-CUDA ngokuyintloko ajolise kwixesha lePython. Faka kuqala nvidia-pyindex kwaye uqiniseke ukuba unepip kunye ne-setuptools ezihlaziyiweyo. Ungabandakanya umgca ohambelanayo kweyakho requirements.txt Ukuba ukhetha.
IiMetapackages kwiWindows yeCUDA 12 (cu12) ezitsalela kuhlobo lwamva nje olukhoyo: phakathi kwabanye, nvidia-cublas-cu12, nvidia-cuda-cccl-cu12, nvidia-cuda-cupti-cu12, nvidia-cuda-nvcc-cu12, nvidia-cuda-nvrtc-cu12, nvidia-cuda-opencl-cu12, nvidia-cuda-runtime-cu12, nvidia-cuda-sanitizer-api-cu12, nvidia-cufft-cu12, nvidia-curand-cu12, nvidia-cusolver-cu12, nvidia-cusparse-cu12, nvidia-npp-cu12, nvidia-nvfatbin-cu12, nvidia-nvjitlink-cu12, nvidia-nvjpeg-cu12, nvidia-nvml-dev-cu12, nvidia-nvtx-cu12.
Ezi metapackages zifakela iipakethe ezithile eziguqulelweyo, umzekelo ngesimamva cu129: nvidia-cublas-cu129, nvidia-cuda-cccl-cu129, nvidia-cuda-cupti-cu129, nvidia-cuda-nvcc-cu129, nvidia-cuda-nvrtc-cu129, nvidia-cuda-opencl-cu129, nvidia-cuda-runtime-cu129, nvidia-cuda-sanitizer-api-cu129, nvidia-cufft-cu129, nvidia-curand-cu129, nvidia-cusolver-cu129, nvidia-cusparse-cu129, nvidia-npp-cu129, nvidia-nvfatbin-cu129, nvidia-nvjitlink-cu129, nvidia-nvjpeg-cu129, nvidia-nvml-dev-cu129, nvidia-nvtx-cu129. Khumbula ukuba i-CUDA Wheels ayibandakanyi izixhobo zomphuhlisi, ixesha lokusebenza kuphela.
Ukusebenzisa umbhobho kuthetha ukuba indawo yeCUDA ibotshelelwe kwindawo yakho yePython. Ukuba uceba ukuqokelela okanye ukuqhuba ngaphandle kwaloo ndawo, faka kwakhona i Isixhobo somgangatho wenkqubo ukuphepha ukungangqinelani kwendlela.
I-CUDA kwi-WSL: Windows 11 kunye ne-Windows 10 21H2
Windows 11 kunye Windows 10 I-21H2 (kwaye kamva) ikuvumela ukuba usebenzise i-CUDA-izakhelo ezikhawulezayo zeML kunye namathala eencwadi ngaphakathi kwe-WSL. Oku kubandakanya I-PyTorch, iTensorFlow, iDocker kunye ne-NVIDIA Container Toolkit, njengokuba ubuya kwenza kwiLinux yasekhaya.
Inyathelo 1: Faka i-CUDA-envidia umqhubi we-WSL kwiwebhusayithi esemthethweni. Lo mqhubi wenzelwe ngokukodwa ukuveza i-GPU ngaphakathi kosasazo. Linux kwiWSL.
Inyathelo 2: Yenza i-WSL isebenze kwaye wongeze unikezelo olusekwe kwi-glibc (Ubuntu/Debian). Hlaziya i-kernel kuHlaziyo lweWindows kwaye uqinisekise ukuba injalo I-5.10.43.3 okanye ngaphezulu ukubaleka wsl cat /proc/version kwiPowerShell.
Inyathelo 3: Landela i-NVIDIA CUDA kwisikhokelo se-WSL ukufaka i-Toolkit ngaphakathi kwe-distro kunye/okanye uqwalasele i-NVIDIA Docker. Uya kukwazi ukusebenzisa iLinux eqhelekileyo ihamba nayo ukukhawulezisa kwaye yahlule ngokugqibeleleyo imekobume kumamkeli weWindows.
Imeko yokwenyani: i-AMD exutywe kunye ne-NVIDIA laptop kunye neempazamo nge-12.8 / 11.8
Makhe sicinge ngeLenovo Ideapad 5 Pro ene-AMD Ryzen, AMD Radeon iGPU kunye neNVIDIA GeForce GTX dGPU. Xa ubaleka nvidia-smi i-aparece Umqhubi 526.56 kunye neCUDA Version 12.0, kodwa ukuzama ukufaka CUDA 12.8 okanye 11.8 akubonakali kusebenza.
Izitshixo zokuyiqonda: i-CUDA Version umgca we nvidia-smi ibonisa ubuninzi bexesha le CUDA loguqulelo oluxhaswa ngumqhubi wakho, hayi i Toolkit efakiweyo. Umqhubi 526.56 Ixhasa i-CUDA 12.0 kunye ne-11.8, kodwa kungekhona i-12.8, efuna umqhubi omtsha kakhulu (55x / 56x yesebe). Hlaziya kumqhubi wamva nje (iSitudiyo okanye i-Game Ready DCH) kwiwebhusayithi ye-NVIDIA kwaye uphinde ufakelo.
Kwiikhompyuter ezixubileyo, jonga kwiPaneli yoLawulo ye-NVIDIA okanye iiSetingi zeMizobo yeWindows ukuze uqiniseke ukuba usetyenziso lwakho lwekhompyuter lisebenzisa i-NVIDIA GPU. Ukuba idGPU ayisetyenziswa, nvidia-smi Oku kunokubangela iingxaki, okanye izikhokelo zinokulayisha kwi-AMD iGPU, engaxhasi i-CUDA. Cinga ngokufaka umqhubi ococekileyo kwaye uqinisekise ukuba amava eGeForce afumanisa i-hardware.
IiMfuno eziKhethekileyo: ML.NET Model Builder (ukuhlelwa komfanekiso kuphela)
Kwimeko yokuhlelwa kwemifanekiso kunye noMqulu woMakhi, iMicrosoft ifuna isitaki esithile: CUDA 10.1 kunye cuDNN 7.6.4. Kulumkele oku kuba iinguqulelo ezintsha azikuxhasi oko kuqukuqela okuthile.
Amanyathelo aphambili nge-cuDNN 7.6.4: khuphela i-ZIP ye-CUDA 10.1, yikhuphe kwaye uyikopishe cudnn64_7.dll kusuka cuda\bin a C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\bin. Awukwazi ukuba neenguqulelo ezininzi ze cuDNN ngaxeshanye; isusa iintsalela zezinye iinguqulelo ukunqanda ungquzulwano ngexesha lokuphunyezwa.
Iimfuno ze-Hardware kule meko: ubuncinane i-CUDA-ekwaziyo i-GPU kunye ne-6 GB yememori ezinikeleyo. Ukuba awunayo i-GPU yalapha ekhaya, iModeli yoMakhi inokuqeqesha koomatshini benyani beGPU. I-Azure, okanye ekugqibeleni kwi-CPU, nangona kukho ixesha elide loqeqesho.
Ukuxilongwa ngokukhawuleza kunye nesisombululo seengxaki eziqhelekileyo
I-GPU ayibonakali kwiiSetingi okanye kuMphathi woMsebenzi: Vula uMphathi weSixhobo, jonga phantsi kweDisplay Adapter, kwaye ufake umqhubi ofanelekileyo. Akukho mqhubi. NVIDIA, IiWindows aziyi kuveza i-dGPU kwii-apps.
Jonga inguqulelo yeCUDA yesixokelelwano: kwi-PowerShell/CMD run nvcc --version. Ukuba ayiphumelelanga, khangela ukuba i Toolkit ifakiwe kwaye i PATH quka incwadi eneenkcukacha bin ukusuka kwi-Toolkit.
I-toolkit ayiboni izixhobo ze-CUDA: Qinisekisa ukuba umqhubi uhlaziywe kwaye i-GPU yakho ikuluhlu lwe-CUDA olunamandla. isixhoboQuery y bandwidthTest Ziyithemometha yakho yokuqinisekisa ukuba isitaki se-hardware-driver-toolkit siyaphendula.
Imowudi yomqhubi engalunganga yekhompyutha: Kwi-GPU yekhompyutha, sebenzisa nvidia-smi ukujonga/ukuguqula i-TCC ukuba ikhona. KwiiGeForces ezininzi zeelaptop, WDDM Yimo ecetywayo kwaye awukwazi ukuyitshintsha.
Ukuxhomekeka kungquzulwano kunye nesakhelo sePython: Ukuba usebenzisa amavili eCUDA ngombhobho, khumbula ukuba agubungela kuphela ixesha lokubaleka kwaye abotshelelwe kwindawo enenyani. Ukuqokelela izandiso okanye ukusebenzisa izixhobo, faka kwakhona i System Toolkit, okanye sebenzisa iConda ukulawula yonke into ngobuninzi.
Imizekelo, iisampulu kunye nezenzo ezilungileyo
Ukuqinisekisa ukusebenza kunye nokusebenza, qulunqa kwaye uqhube deviceQuery y bandwidthTest ukusuka kwizisombululo ze-VS kwindawo yokugcina iisampuli. Ulwakhiwo ludla ngokushiya iibhinari kwiindlela ezinje Iisampulu zeCUDA\v13.0\bin\win64\Release ukuba ugcine amaxabiso angagqibekanga.
Ukuze ubone ngokubonakalayo amandla eCUDA, sebenzisa iisampulu zegraphical njenge particles. Ngaphaya kwedemo, baya kukunceda ukuba uthathe iireferensi zokusetyenziswa imemori ekwabelwana ngayo kunye neepateni zegridi/zebhloko kwiiprojekthi zakho.
Ezinye iindlela zamafu ngeeGPUs
Ukuba awufuni ukulwa nabaqhubi bendawo okanye ufuna iincopho zamandla, ungakhetha iimeko zamafu ngeeGPU zangoku (A100, RTX 4090, A6000, njl.). Ezi nkonzo zibonelela ukusasazwa kwangoko, iitemplates zePyTorch/TensorFlow kunye nokuhlawula-njengoko-uhamba, luncedo kuqeqesho olunzulu okanye uvavanyo olukhawulezayo.
Amanqaku kunye namanqaku
Amaxwebhu obugcisa be-NVIDIA kunye neepakethi zinokutshintshwa ngaphandle kwesaziso. Jonga rhoqo i khupha amanqaku kunye nokuhambelana okuhlaziyiweyo phambi kokufaka abaqhubi okanye ii-Toolkits kwimveliso.
I-OpenCL luphawu lwentengiso lwe-Apple Inc. olusetyenziswa phantsi kwelayisensi yiKhronos Group. I-NVIDIA kunye nelogo yayo ziimpawu zorhwebo okanye ziimpawu zorhwebo ezibhalisiweyo ze NVIDIA Corporation e-US nakwamanye amazwe.
Ngomqhubi ofanelekileyo, i-toolkit echanekileyo kunye neemvavanyo ezimbalwa ezenziwe kakuhle, iWindows ngumhlaba oqinileyo we-CUDA: unokuphuhlisa nge-Visual Studio, uqinisekise kunye neesampuli ezisemthethweni, sebenzisa izikhokelo ze-AI kwi-WSL njenge-Linux kwaye, ukuba ubambekile, tsala ezinye izifakeli (i-Conda / pip) okanye ifu ukuze ugweme ukuxhoma; into ebalulekileyo kukulungelelanisa iinguqulelo zomqhubi kunye ne-toolkit, uqinisekise ukuba i-NVIDIA GPU yeyona isetyenziswayo, kwaye ikuxhase deviceQuery/bandwidthTest ukuqinisekisa ukuba yonke into iluhlaza. Yabelana ngesi sifundo kwaye abasebenzisi abaninzi baya kuyazi indlela yokufaka iCUDA kwiWindows.