I-DeepSeek ne-Huawei baqinisa i-Ascend ukuze bancintisane ne-Nvidia ngesofthiwe

  • I-DeepSeek ne-Huawei bakhipha ingxenye yengqalasizinda yama-processor e-Ascend njengomthombo ovulekile.
  • Ukubambisana kuhlanganisa imitapo yolwazi yokubala kanye nezokuxhumana, kanye ne-supernode ene-128 Ascend 950s.
  • I-DeepGEMM-Ascend namanye amathuluzi afuna ukwenza lula imisebenzi ye-AI ngaphandle kokuthembela kuhlelo lwe-Nvidia.
  • I-CUDA ayisuswa futhi i-Ascend ayikafiki ezingeni elifanayo ne-Nvidia: umgomo ukunciphisa izindleko zokushintsha nokwandisa uhlelo lwe-ecosystem.
Alberto NavarroKubuyekezwe ngomhla ka-

Imizuzu ye-4

I-DeepSeek ne-Huawei uma kuqhathaniswa ne-Nvidia

I-DeepSeek ne-Huawei bathathe isinyathelo esishintsha impi ne-Nvidia kusukela kuma-chip kuya kwi-software. Inkampani ye-AI yaseShayina kanye nomkhiqizi wezokuxhumana bahlanganyele ekwakheni amathuluzi okuhlela aklanyelwe ama-processor e-Ascend. Umgomo uwukuba abathuthukisi bakwazi ukuqeqesha nokusebenzisa amamodeli obuhlakani bokwenziwa ngaphandle kokudinga ukudlulela ohlelweni lwe-Nvidia. Isihluthulelo ukunciphisa ukuncika kumapulatifomu ezinkampani zangaphandle kanye nokwenza kube lula ukushintshela kwihadiwe yaseShayina.

Lesi simemezelo senzeke ngoLwesithathu futhi asigcini nje ngokuboniswa okulula kobuchwepheshe. I-DeepSeek ikhiphe njengengxenye yengqalasizinda eyilungiselele i-Ascend, enamalabhulali ekhompyutha kanye nokuxhumana phakathi kwama-chip. Sekuphinde kwenziwa umsebenzi ekwakhiweni kwe-supernode okusekelwe kuma-processor angu-128 e-Ascend 950, aklanyelwe ukwenza ngcono kokubili ukubala kanye nokushintshana kwedatha phakathi kwama-accelerator.

IHuawei AI
I-athikili ehlobene:
I-Huawei isheshisa imizamo yayo ye-AI: ama-chip amasha, amaqoqo amakhulu, kanye nekusasa eligcwele ama-ejenti azimele

Yini i-DeepSeek eyifuna nge-Huawei emkhakheni we-AI?

Esinye sezithiyo ezinkulu zeHuawei ngokumelene neNvidia akusikho nje ukukhiqiza isisheshisi sokuncintisana. Inkinga iphakama lapho abathuthukisi kufanele bakhe kabusha amathuluzi, imitapo yolwazi, kanye nezinhlelo zokusebenza ukuze basebenzise ithuba laso. I-DeepSeek ikhiphe izinguqulo ze-Ascend zamathuluzi esevele iwasebenzisa ezinhlelweni zayo. Phakathi kwawo kukhona i-DeepGEMM-Ascend, ukusetshenziswa okuhambisana nama-DeepGEMM APIs amaprosesa e-Huawei futhi okulungiselelwe ukusebenza kwe-matrix okuvamile emisebenzini ye-AI.

Indawo evulekile, kodwa ngaphandle kokususa i-CUDA esihlalweni sobukhosi

Kubalulekile ukungadukiswa. Isinyathelo se-DeepSeek ne-Huawei asisho ukuthi i-CUDA isisusiwe, noma ukuthi abaprosesa be-Ascend sebefinyelele kuma-accelerator e-Nvidia. Ukugxila okuyinhloko, okwamanje, ukwakha ingqalasizinda evumela abathuthukisi abaningi ukuthi basebenze ngehadiwe yaseShayina futhi kunciphisa izindleko zokushintsha amapulatifomu . Yilapho ingxenye yempi ilwiwa khona.

Kulesi simo, iphrojekthi ye-TileLang isekela ukwakheka okuhlukahlukene, okuhlanganisa i-Nvidia CUDA, i-AMD ROCm, kanye ne-Apple Metal. Ukusetshenziswa kwayo kwe-Ascend kwandisa lolu hlelo lokusebenza kumaprosesa e-Huawei. Kubathuthukisi baseYurophu nabaseSpain, lezi zindaba zibalulekile ngoba zingavula ezinye izindlela zehadiwe nesofthiwe emakethe ephethwe yi-Nvidia, yize ukwamukelwa kwangempela kuzoncika ekuvuthweni kwamathuluzi kanye nokusekelwa okutholakalayo.

hlonipha hubia deepseek-4
I-athikili ehlobene:
I-Honor, i-Nubia, kanye nokuthuthukiswa kwe-DeepSeek-R1 kubasizi beselula be-YOYO kanye nezinye izimiso zemvelo zaseShayina

Kungani ebuka iYurophu eduze?

ESpain nakwamanye amazwe aseYurophu, izinkampani eziningi zobuhlakani bokwenziwa kanye nezikhungo zedatha zihlala zixhumene ne-encyclopedia yeNvidia ngenxa yokulula kwayo, ukusebenza kwayo, kanye nemitapo yolwazi eseyikhona. Isiphakamiso esivela ku-DeepSeek kanye ne-Huawei asithembisi ushintsho olusheshayo, kodwa sidala ingcindezi yezinketho ezengeziwe. Uma isofthiwe ye-Ascend ithola ukuzinza, amaqembu obuchwepheshe angahlola amapulatifomu ahlukile ngaphandle kokubhala kabusha amaphrojekthi awo ngokuphelele.

Ukukhishwa kwezingxenye ezifana ne-DeepGEMM-Ascend kanye nemitapo yolwazi yokuxhumana phakathi kwama-chip kukhomba lowo mgomo ofanayo. Akukhona nje ukusebenza okuluhlaza, kodwa mayelana nokuhlinzeka abahleli bezinhlelo ngendlela ejwayelekile yokufuduka. Uma ukungezwani okuncane kukhona, kulapho kungenzeka khona ukuthi ihadiwe yaseShayina izothola indawo ezindaweni zobungcweti. Noma kunjalo, iNvidia igcina inzuzo enzima ukuyinqoba esikhathini esifushane ngenxa yomphakathi wayo, imibhalo, kanye neminyaka yokuba khona embonini.

Kusho ukuthini lokhu ngekusasa le-AI?

Impi phakathi kweShayina neNvidia ayiseyona nje indaba yokukhiqiza ama-chip kuphela. I-DeepSeek ne-Huawei bazama ukushintsha umncintiswano ube yisofthiwe, lapho kunqunywa khona amathuluzi abathuthukisi abawasebenzisa nsuku zonke. Ukwakhiwa kwe-128 Ascend 950 kanye nemitapo yolwazi ekhishwe kuyizingxenye zesu elibanzi: ukudala uhlelo lwabo lwe-ecosystem olungabizi kakhulu ukulusebenzisa.

Emakethe, isignali icacile: kuzoba nemizamo eminingi yokuphula ukuthembela ku-CUDA, yize kungebona bonke abazophumelela. I-TileLang kanye nokusekelwa kwayo kwezakhiwo ezahlukahlukene kubonisa ukuthi imboni nayo iphokophela ezingqimbeni eziguquguqukayo zokungabonakali. Kulesi simo, iYurophu ingathola amandla okuxoxisana kanye nokwehlukahlukana kobuchwepheshe uma izinkampani zayo zikwazi ukusebenzisa lezi zindlela ngaphandle kokubeka engcupheni ukuthembeka.

Ubudlelwano phakathi kwe-DeepSeek ne-Huawei, kanye nekhodi yayo yomthombo ovulekile ye-Ascend, amakhompyutha kanye nemitapo yolwazi yokuxhumana, i-supernode yamaprosesa angu-128 e-Ascend 950, kanye namathuluzi afana ne-DeepGEMM-Ascend, bumelela umzamo omkhulu wokuncintisana ne-Nvidia ngesofthiwe. Akukho ophumelelayo okwamanje, noma othatha indawo eqondile ye-CUDA, kodwa kukhona umzamo ocacile wokunciphisa izindleko zokushintsha nokwandisa uhlelo lwe-ecosystem. Umdlalo usavuliwe, futhi iYurophu ibukele ngentshisekelo.


Engeza njengomthombo okhethwayo ku-Google