- Ukuthuthukiswa kwezisheshisi ze-Instinct MI350 kanye ne-MI400 ukuze zincintisane ezikhungweni zedatha ezine-bandwidth ephezulu yememori.
- I-Ryzen AI Max PRO yethulwa ngama-NPU ahlanganisiwe ukuze kuqhutshwe amamodeli ezilimi ayinkimbinkimbi endaweni.
- Betha ku-ROCm njengendlela ehlukile yomthombo ovulekile esikhundleni sokubusa kwe-CUDA ye-NVIDIA ecosystem.
- Ukuhlanganiswa kwamaprosesa e-EPYC Venice anama-cores afinyelela ku-256 ukuze kuthuthukiswe ingqalasizinda yeseva ye-AI.
Umncintiswano wokubusa ubuhlakani bokwenziwa ususuke esithembisweni uye empini yehadiwe ephelele. Kule nkundla, i-AMD isheshisa imizamo yayo yokunikeza ezinye izindlela ezithembekile ezivumela amabhizinisi nabasebenzisi bokugcina ukuthi basebenzise amamodeli e-AI ngaphandle kwezindleko ezinkulu noma ingqalasizinda iphelelwe yisikhathi zingakapheli izinsuku.
Umgomo ucacile: ukusuka kumaphrojekthi okuhlola alula ukuya ekukhiqizeni kwangempela, okungandiswa . Ukuze kufezwe lokhu, inkampani ayigxili nje kuphela emandleni aluhlaza kodwa futhi nasekuxazululeni izinkinga ezivamile, njengokuphathwa kwamandla, ukuphathwa kwedatha, kanye nomzabalazo oqhubekayo wokugwema ukuthembela kumthengisi oyedwa emakethe.
Ama-Accelerators e-Instinct: Inselelo Eqondile ku-NVIDIA
Uma kukhulunywa ngezikhungo zedatha, ama-GPU avamile awanele; ama-accelerator abalulekile. Yilapho i- Instinct MI350 ingena khona , ithembisa igxathu elikhulu phambili, ishesha izikhathi ezingu-35 emisebenzini yokuphetha kunezandulelayo. Lezi zilwane zinememori engu-288 GB ye-HBM3E, ezivumela ukuthi ziphathe imisebenzi emikhulu kalula.
Kodwa i-AMD ayigcini lapho futhi isivele ihlose i- Instinct MI400 ngo-2026. Lawa ma-chip kulindeleke ukuthi afinyelele ku-40 PFLOPS ngokunemba kwe-FP4 futhi abe nememori ye-HBM4 efika ku-432 GB. Konke lokhu kuzofakwa kuma-racks e-Helios, isakhiwo esenzelwe ukuncintisana ngqo nezixazululo ze-NVIDIA, enikeza umthamo we-bandwidth omangalisayo ngempela .
Esikhathini eside, imephu yomgwaqo ihlanganisa i-EPYC Verano kanye ne-Instinct MI500X, ezingasebenzisa i-node yokukhiqiza ye-TSMC engu-1,6nm. Isu licacile: lidlula umncintiswano ngomthamo wememori kanye nesivinini sokudlulisa, okuyilapho inzuzo yangempela ikhona lapho kuqeqeshwa amamodeli olimi amakhulu.
I-ecosystem yesofthiwe kanye nodonga lwe-CUDA
Akukhona nje kuphela ngehadiwe enamandla. Ubuthakathaka obukhulu be-AMD bube yisofthiwe ngokomlando. Ngenkathi i-NVIDIA ibusa nge-CUDA, indinganiso esetshenziswa yiwo wonke umuntu, i-AMD ikhethe i-ROCm . Inguqulo yesi-7 yale pulatifomu yomthombo ovulekile ihlose ukuvala igebe, ize isebenze kahle kakhulu kune-CUDA ezimweni ezithile njengemodeli ye-DeepSeek R1.
Noma kunjalo, indlela ebheke phambili iyinselele ngoba abathuthukisi abaningi banomuzwa wokuthi isofthiwe ye-AMD iseneziphazamiso ezenza kube nzima ukuqeqesha . Isihluthulelo sempumelelo yesikhathi esizayo asikho ekwengezeni ama-core amaningi, kodwa ekuqinisekiseni ulwazi oluhle lokuhlela olungeyona inkinga konjiniyela.
I-AI kuma-desktops nama-laptops: i-Ryzen AI kanye ne-Radeon
Kubasebenzisi abangenayo isikhungo sedatha esingaphansi, i-AMD yethule amaprosesa e -Ryzen AI Max PRO . Lawa ma-chip ahlanganisa i-NPU ezinikele efinyelela ku-50 TOPS, evumela imisebenzi efana ne-Llama 3.1 70B ukuthi isebenze endaweni kanye nokungaxhunyiwe ku-inthanethi—into eyayibonakala njengenganekwane yesayensi kudivayisi encane esikhathini esingeside esidlule—okwenza kube lula ukusebenzisa amamodeli e-AI ku-PC yakho.
Ngokuphathelene nemidwebo, i-Radeon RX 9000 Series, ikakhulukazi i-RX 9070 ene-VRAM engu-16 GB, iyindawo yokungena efinyelelekayo kulabo abafuna ukwenza isiphetho se-AI ngaphandle kokuchitha imali eningi. Kubasebenzisi abadinga kakhulu, i-Radeon AI PRO R9700 inyusa izinga layo libe ngu-32 GB, isusa noma yiziphi izinkinga kudivayisi.
Ingqalasizinda yeseva: i-EPYC Venice
Lawa ma-GPU athembele kuma-processor eseva. Ama-processor e-EPYC Venice azayo , asekelwe ku-Zen 6, athembisa izinhlobo ezinama-cores afinyelela ku-256. Lokhu kwanda kokusebenza okungu-70% esizukulwaneni esidlule kubalulekile ukuvimbela i-CPU ukuthi ingabi yisithiyo salezi zisheshisi ze-AI ezinamandla.
Kuyathakazelisa ukubona ukuthi imakethe isiqala kanjani ukusabela. Abantu abanjengoSam Altman we-OpenAI sebevele bavumile ukuthi izixazululo ze-AMD zizoba yizingxenye ezimangalisayo ezikhungweni zabo zedatha. Lokhu kukhombisa ukuthi, yize i-NVIDIA inesabelo esikhulu semakethe, imboni ifuna ngokulangazela enye indlela esebenza kahle futhi, ngaphezu kwakho konke, eshibhile ukuyisebenzisa.
Isu le-AMD lihlanganisa ihadiwe ephazamisayo kwimemori kanye namandla okucubungula ngomzamo onamandla wokwenza isofthiwe ye-AI ibe yentando yeningi. Ngokuhlanganisa izinjini ze-AI ezizinikele kusukela kuma-laptop kuya kuma-server racks ayinkimbinkimbi kakhulu, inkampani ihlose ukwenza i-AI ingabi yinto yokunethezeka ebizayo kodwa ithuluzi elifinyelelekayo nelisebenzayo lanoma yiluphi uhlobo lochwepheshe noma ibhizinisi.
