Andrzej Janik, onye a ma ama maka ọrụ ya na-eke mmejuputa teknụzụ CUDA maka Intel na AMD GPU, ka a manyere, na arịrịọ AMD, iwepu akụkụ ọha nke koodu nke ọrụ ZLUDA mepere emepe, nke na-enye ohere ịme ngwa CUDA na-agbanweghị agbanwe. na sistemu nwere AMD GPUs na-eji ROCm stack na oge HIP (Interface-computing Interface maka Portability). Achọpụtara ya iche na onye nrụpụta enwetabeghị arịrịọ ọ bụla metụtara ọrụ ZLUDA sitere na NVIDIA, nke nwere ikike nke teknụzụ CUDA.
Na mbu, emebere ZLUDA oru ngo maka Intel GPUs, mana na 2022 Andrzej hapụrụ Intel wee binye aka na AMD iji wuo oyi akwa maka ndakọrịta CUDA. Ná mmalite nke 2024, AMD kpebiri na ịgba ọsọ CUDA ngwa na AMD GPUs adịghị amasị azụmahịa, nke, dị ka usoro nke nkwekọrịta ahụ, kwere ka onye mmepụta mepee mmepe ya. N'oge ozi email ahụ, onye nrụpụta nwetakwara ikike ibipụta n'aka onye nnọchi anya AMD.
Ọnwa 6 ka e bipụtachara mgbanwe ndị e kere mgbe ha na-arụ ọrụ AMD, ndị ọka iwu maka ụlọ ọrụ a kpọtụrụ Andrzej ma mee ka o doo anya na ikike enyere na mbụ n'oge akwụkwọ ozi enweghị ikike iwu. A naghị enye nkọwa gbasara akwụkwọ ozi na ndị ọka iwu AMD n'ihi nkwekọrịta enweghị nkwupụta. Otú ọ dị, ihe si na ya pụta bụ na a manyere onye mmepụta ka ọ wepụ koodu niile emepụtara maka AMD site na ngalaba ọha.
Andrzej achọghị ịkwụsị ịrụ ọrụ na ọrụ ahụ ma na-ezube imegharị ya dabere na koodu ahụ dị tupu mmalite nke imekọ ihe ọnụ na AMD (E mere koodu ochie iji rụọ ọrụ na Intel GPUs). A ga-ewepụta nkọwa gbasara mmepe nke ọrụ ahụ n'ime izu ole na ole. Ka ọ dị ugbu a, a na-akọ na ụfọdụ ọrụ ga-efunahụ ya, nakwa na ụfọdụ ọrụ ndị a na-emepe emepe agaghị emepe. Dịka ọmụmaatụ, e mere atụmatụ ibipụta nsonaazụ nke ọrụ na nkwado maka teknụzụ NVIDIA GameWorks, bụ nke e wetara na o nwere ike iji mmetụta dabeere na GameWorks na egwuregwu "Batman: Arkham Knight".
Na mgbakwunye, anyị nwere ike ịhụ mmepe nke SCALE Toolkit, nke ezubere inye ikike iji ngwa CUDA rụọ ọrụ na sistemụ nwere AMD GPUs. Akụkụ bụ isi nke ngwa ngwa emepụtara bụ ihe nchịkọta nhọrọ-ọkwa kwekọrọ na nvcc, dabere na koodu ntọala nke ọrụ LLVM ma na-enye gị ohere ikpokọta koodu CUDA ka ọ bụrụ ihe nnọchianya maka ịgba ọsọ na AMD GPUs. Ihe oru ngo a na-enyekwa ọba akwụkwọ ejikọtara na-emejuputa CUDA-X API n'elu nchịkọta AMD ROCm, Oge ọsọ na API ọkwọ ụgbọala maka AMD GPUs. Ngwá ọrụ ahụ na-eme ka ọnụnọ nke NVIDIA CUDA Toolkit maka ngwa ngwa, nke na-enye gị ohere ịmepụta koodu CUDA maka AMD GPU na-emeghị mgbanwe na koodu isi mmalite ma wuo scripts.
Ugbu a, ngwa SCALE dị na nnwale beta (ngwugwu dị maka nkesa dị iche iche Linux). A nwalela SCALE site na sistemụ dịka Blender, hashcat, llama-cpp, NVIDIA Thrust, GOMC, stdgpu, xgboost, faiss, na AMGX na sistemụ nwere AMD gfx1030 (Navi 21, RDNA 2.0) na AMD gfx1100 (Navi 31, RDNA 3.0) GPU. Ọrụ SCALE bụ nkeonwe ma kesaa ya na-enweghị koodu isi mmalite. Akwụkwọ ikike a na-enye ohere maka ojiji azụmaahịa, nkesa, na mgbanwe, mana ọ na-egbochi injinịa azụ. Ndị mmepe anaghị ewepụ ohere nke imepe koodu isi mmalite n'ọdịnihu, mana n'oge mmepe a, ha ekpebiela na ha agaghị ebipụta koodu ahụ.
isi: opennet.ru
