Ukuhlela kwe-GPU Okuyinhloko nge-NVIDIA CUDA Rust

Isibuyekezo sokugcina: Novemba 17 we-2026
  • I-NVIDIA iqalise uhlelo lokuhlanganisa i-Rust ekuthuthukisweni kwe-GPU kernel, isuse ukuncika kwezinye izilimi.
  • I-ecosystem ihlukaniswe ngemizila emibili eyinhloko: i-cuda-oxide yokulawula okuphelele nge-SIMT kanye ne-cutile-rs ukuze kube lula ukusebenzisa ama-Tiles.
  • Womabili amathuluzi aletha iziqinisekiso zokuphepha kanye nobunikazi beRust kuhadiwe, ethola amaphutha ememori kanye nokusebenza kwedatha ngesikhathi sokuhlanganiswa.

Amakhadi amathathu ehluzo e-NVIDIA GeForce RTX abekwe ndawonye, ​​​​amele ihadiwe esebenza kahle esetshenziselwa ukubala i-CUDA kanye ne-AI.

Ingqalasizinda esekela ubuhlakani bokwenziwa besimanje ishintsha ngesivinini esikhulu, futhi ukuze ihambisane nesivinini, i-NVIDIA inqume ukutshala imali eningi ezinhlelweni zomdabu ku-Rust . Ngenkathi i-CUDA C++ kanye ne-CUDA Python sezivele zisunguliwe futhi zingamathuluzi aqinile emkhakheni webhizinisi, imboni ishintshela ku-Rust ngenxa yekhono layo lokususa amaphutha enkumbulo abalulekile ngaphambi kokuba ikhodi isebenze, ngaphandle kokulahlekelwa ukusebenza okuluhlaza.

Kuze kube muva nje, i-Rust yayisetshenziswa ezingqimbeni zangaphandle njengezinjini zokucabanga noma abashayeli, kodwa ama-GPU kernels ahlala "yindawo engaboni" lapho abathuthukisi kwadingeka basebenzise ezinye izilimi. Ngokufika kwe-CUDA Rust, i-NVIDIA ifuna ukuvala lesi sikhala sobuchwepheshe , ivumela abathuthukisi ukuthi babhale i-logic yedivayisi ngqo ku-Rust futhi bayihlanganise ku-PTX, ngaleyo ndlela bagweme izimbozo ezinzima ezaziyinkimbinkimbi endleleni yokusebenza.

Izindlela ezimbili zokunqoba i-GPU

Ukuchazwa okuningiliziwe kwekhadi lehluzo le-GeForce elibonisa izingxenye zehadiwe eziyinkimbinkimbi kanye namasekethe.

Ukuze ivumelane nezidingo ezahlukahlukene zabahleli bezinhlelo, i-NVIDIA isebenzise izindlela ezimbili ezahlukene ezilingisa ukuguquguquka kwe-CUDA. Ngakolunye uhlangothi, kukhona imodeli ye-SIMT, kanti ngakolunye uhlangothi, imodeli ye-Tile. Isincomo esijwayelekile ukuqala ngendlela ye-Tile , njengoba umqoqi ethuthukisa indlela idatha ehlelwe ngayo kuyo yonke imiklamo, evimbela umhleli ukuthi enze izinqumo ezenziwe ngesandla nezikhathazayo mayelana nehadiwe ethile.

Ukulawula okugcwele nge-cuda-oxide (SIMT)

Isikrini sekhompyutha esibonisa ikhodi yomthombo enezinketho zokuthuthukiswa ezisizwa yi-AI, esibonisa ubunjiniyela besofthiwe yesimanje.

Kulabo abangafuni ukuyeka ukulawula futhi abadinga ukunemba okuqondile phezu kwezintambo kanye nenkumbulo eyabiwe, kukhona i-cuda-oxide . Leli thuluzi lisebenza njenge-backend yokukhiqiza ikhodi yangokwezifiso ye-rustc. Inqubo iyinkimbinkimbi: imisebenzi ephawulwe njengoba ama-kernel edlula ku-MIR kaRust, edlula kuhlaka lwe-Pliron kanye ne-LLVM IR, futhi ekugcineni aguqulwa abe yikhodi ye-PTX esebenzisekayo ku-GPU.

  Indlela yokusebenzisa iMarkdown ngokushesha ku-GitHub naku-Reddit

Ngokuphathelene nezidingo zobuchwepheshe, akuyona into elula ukuyifinyelela; udinga i-Linux, i-GPU enamandla okusebenzisa ikhompyutha angu-8.0 noma ngaphezulu, kanye ne- isethi yansuku zonke ye-toolchainEnye yezinto eziyigugu zalolu hlelo ukusetshenziswa DisjointSliceuhlobo lwedatha oluqinisekisa ukuthi intambo ngayinye inayo ukufinyelela okukhethekile kwento yakhoukuvimbela i-Rust compiler ukuthi ivimbele ikhodi yokuzama ukwenza imalimboleko engaguquki engahambisani.

Ubulula nokuphepha nge-cutile-rs (ithayela)

Ukusondelana okufushane nokugqamile kwemigqa yekhodi yokuhlela enemibala eminingi esikrinini, okufanekisela ukusebenza kahle kolimi lwe-Rust.

Uma ukhetha okuthile okufana nokusebenza ne-NumPy noma i-PyTorch, i-cutile-rs iyindlela engcono kakhulu. Lapha, awuhleli imicu ngayinye, kodwa usebenza "kuma-tile" noma amabhlogo edatha. Umhlanganisi unquma ukuthi zingaki imicu yangempela edingekayo ukuhambisa lelo bhloko, okwenza intuthuko ibe lula kakhulu futhi ingathambekeli emaphutheni okukhomba.

Ngokungafani ne-SIMT yakubo, le phrojekthi ifinyeleleka kalula ngoba isebenza ku -Rust ezinzile futhi ayidingi ukufakwa kwe-LLVM ngesandla. Isebenzisa i-macro ehlanganisa i-kernel isebenzisa ama-CUDA Tile IR ngesikhathi sokuqaliswa kokuqala (JIT). Isici esihle kakhulu ukuphathwa kwayo kokusebenza okuvilaphayo : uhlelo lubhalisa umucu womyalo futhi lusebenzisa kuphela isethi yonke lapho indlela yokuvumelanisa isetshenziswa, okuthuthukisa ukuxhumana phakathi kwe-CPU ne-GPU.

Umlingo womhlanganisi: Sala kahle kwi-Data Races

Ukusondela kwe-GPU esebenza kahle kakhulu enabalandeli abathathu, okugqamisa ukupholisa namandla adingekayo emisebenzini ye-AI.

Okwenza i-CUDA Rust ibe yinto eshintsha umdlalo ngempela yindlela ephatha ngayo inkumbulo. Kuhlelo lwendabuko lwe-GPU, amaphutha okuhlanganisa (lapho imicu emibili ibhalela endaweni efanayo) ayizinhlupho ezimbi ezivela kuphela ekukhiqizweni. I-Rust ixazulula lokhu ngokuphoqelela imithetho yobunikazi bayo emngceleni wokukhululwa kwe-kernel.

  • Kumodeli ye-SIMT, uma uzama ukusebenzisa i-output buffer njengokufaka, i-compiler izofaka iphutha. iphutha lemali mboleko eguquguqukayo (E0502).
  • Kumodeli yethayili, impahla ye-tensor iyadluliselwa, futhi noma yimuphi umzamo wokusebenzisa inani elihanjisiwe uzobangela iphutha E0382.
  Konke mayelana ne-metaprogramming: izibonelo ezingokoqobo kanye nokusetshenziswa

Lokhu kusho ukuthi isofthiwe umshwalense wokwakhaNgenkathi kumemori eyabiwe ye-SIMT kusadinga amabhlogo unsafeImodeli yethayela isusa isidingo sokuphatha izinkomba zemicu ngesandla, isusa ithuba lokwenza lawo maphutha ajwayelekile ku-equation.

Isimo samanje kanye nekusasa lephrojekthi

Kubalulekile ukugcina izinto zingokoqobo: womabili amaphrojekthi asesigabeni sokuhlola . I-cuda-oxide isesigabeni sokuqala se-alpha, kuyilapho i-cutile-rs ithuthuke kakhulu futhi isivele isetshenziswa kumaphrojekthi afana ne-HuggingFace's Grout inference engine. I-NVIDIA ikwenze kwacaca ukuthi ama-API angashintsha nokuthi kusenendlela okufanele ilandelwe ngaphambi kokuba la mathuluzi afanelekele izindawo zokukhiqiza ezinkulu.

Naphezu kokuba kusha, le ndlela yokusebenza iyahambisana futhi isebenzisana nemizamo yangaphambilini efana ne-rust-gpu kanye ne-CubeCL. Umgomo ukwenza i-Rust ibe ulimi olusezingeni eliphezulu lokubala okusheshayo, okuvumela ikhodi ye-GPU ukuthi ibe namandla futhi igcinwe njengekhodi yesistimu. Amandla okusebenzisana phakathi kwama-kernel e-Tile kanye ne-SIMT ekusakazweni okufanayo kuqinisekisa ukuthi abathuthukisi abavalelwe endleleni eyodwa.

Ukuzibophezela kwe-NVIDIA ekuhlanganiseni i-Rust enkabeni ye-CUDA kumelela ukuvela okudingekayo kwengqalasizinda ye-AI, ukushintsha ukuphepha kusukela esikhathini sokusebenza kuya esikhathini sokuqoqa. Ngokunikeza izindlela zokulawula intambo ye-granular kanye nokuphathwa kwe-tensor okulula, kuvumela ukufinyelela ekusebenzeni okweqile kwama-GPU ngokuqeda ubuthakathaka bememori obujwayelekile ku-C++, ngaleyo ndlela kuqalise inkathi lapho ukusebenza kahle kanye nokuphepha okuphelele kuhlangana khona ngaphakathi kwe-binary efanayo.

  I-Python IDE Ukuqhathanisa: Thola Okungcono Kakhulu Kwakho

Imibuzo ebuzwa njalo

Uyini umehluko phakathi kwe-cuda-oxide ne-cutile-rs?

I-Cuda-oxide igxila kumodeli ye-SIMT kulabo abadinga ukulawula okuphelele imicu nememori, kuyilapho i-cutile-rs isebenzisa imodeli ye-Tile ukusebenzisa ama-tensioner kalula nangokuphephile.

Kungenzeka yini ukusebenzisa i-CUDA Rust kumaphrojekthi okukhiqiza namuhla?

Akukhuthazwa ukukhiqizwa ngoba womabili amaphrojekthi ayahlolwa, asesigabeni sokuqala futhi ama-API awo angase ashintshe kakhulu.

Yiziphi izidingo zehadiwe engizidingayo ukuze ngisebenzise la ma-kernels?

Kudingeka i-NVIDIA GPU enekhono lokubala elingu-8.0 noma ngaphezulu kanye nokusetshenziswa kwezinhlelo zokusebenza ze-Linux.

I-Rust iwavimbela kanjani amaphutha ememori ye-GPU?

Isebenzisa uhlelo lobunikazi nokuboleka ukuthola ukungqubuzana okungaziwa kanye nokuqhutshwa kwedatha ngesikhathi sokuhlanganiswa, ukuvimbela ikhodi engaphephile ukuthi isebenze.

Ingabe ngidinga ukusebenzisa inguqulo yasebusuku ye-Rust kukho konke?

Indlela ye-cuda-oxide kuphela edinga uchungechunge lwamathuluzi lwasebusuku, kuyilapho ama-cutile-rs ehambisana nezinguqulo ezizinzile ze-Rust 1.89 noma ngaphezulu.

Ngingakwazi yini ukuxuba ama-kernel amamodeli ahlukene kuhlelo lokusebenza olufanayo?

Yebo, ama-kernel e-Tile ne-SIMT ayasebenzisana futhi angasebenza ekusakazweni okufanayo kusetshenziswa ama-buffer amadivayisi abiwe.

Iyini ikhodi ye-PTX kumongo we-CUDA Rust?

Lulimi lwe-intermediate assembly lapho kuhlanganiswa khona ama-Rust kernels ukuze umshayeli we-GPU akwazi ukuwasebenzisa kahle.

I-TecnoDigital

TecnoDigital

Nginogqozi ngobuchwepheshe nokuthuthukiswa kwesofthiwe, ngingena ezweni lezinhlelo nekhompuyutha ngenhloso yokukhuthaza ukuqamba okusha nokuxazulula izinselele eziyinkimbinkimbi.