Nge-BMS, ibhasi, yezimboni, ikhebula lokungenisa izimboni.
Njengoba umkhosi wasentwasahlobo usondele, injabulo ezungeze i-Deepseed ihlala iqinile. Iholide lakamuva liqokomise umuzwa obalulekile wokuncintisana ngaphakathi komkhakha wezobuchwepheshe, ngokuxoxa okuningi futhi uhlaziya le "catfish." ISilicon Valley ibhekene nomuzwa ongakaze abulawe ngobunzima: Abameli bomthombo ovulekile bathola imibono yabo futhi, futhi ngisho ne-Opeequai iphinde iphinde isulwe isu lokuvalwa komthombo. I-paradigm entsha yezindleko eziphansi ze-computational iveze ukusabela kwe-chain phakathi kwama-giants we-chip afana ne-nvidia, okuholele ekuqopheni ukulahleka kwenani lemakethe yosuku olulodwa emlandweni we-US stock, kanti ama-ejensi kahulumeni aphenya ngokuhambisana kwama-chip asetshenziswa yi-Deepseek. Phakathi kokubuyekezwa okuxubile kwama-Deepseeek phesheya kwezilwandle, ikakhulukazi, kuthola ukukhula okungavamile. Ngemuva kokwethulwa kwemodeli ye-R1, uhlelo lokusebenza oluhambisana nalo lubonile ukuhlinzwa ku-traffic, okukhombisa ukuthi ukukhula kwemikhakha yohlelo lokusebenza kuzoshayela i-autoscem yohlelo lokusebenza jikelele. Isici esihle ukuthi i-Deepseed izokwandisa amathuba okufaka isicelo, iphakamisa ukuthi ukuthembela ku-chatgpt ngeke kubize kakhulu ngokuzayo. Lokhu kuguqulwa kuboniswe emisebenzini yakamuva ka-Opelai, okubandakanya ukuhlinzekwa kwemodeli yokubonisana ebizwa nge-O3-Mini ukuthi ikhululeke abasebenzisi ukuphendula i-Deepseek R1, kanye nokuthuthuka okulandelayo okwenze umfutho womphakathi we-O3-Mini. Abasebenzisi abaningi baphesheya kwezilwandle bazwakalise ukubonga ku-Deepseek yale ntuthuko, yize leli ntekene lokucabanga lisebenza njengesifinyezo.
Ngokuzenzakalelayo, kusobala ukuthi ukujula kubandakanya abadlali basekhaya. Ngokugxila kwayo ekunciphiseni izindleko zokuqeqeshwa, abakhiqizi abahlukahlukene be-Phic Chip, abahlinzeki be-CHCEP abaphakathi nendawo, kanye nokubhalwa okuningi kujoyina ngenkuthalo i-ecosystem, ukuthuthukisa izindleko zokusebenzisa imodeli ejulile yokusebenzisa imodeli ejulile. Ngokusho kwamaphepha e-Deentieseeeek, ukuqeqeshwa okuphelele kwemodeli ye-V3 kudinga amahora angama-2.788 wezigidi ezingama-H800 GPU kuphela, futhi inqubo yokuqeqesha iqinile kakhulu. I-MOE (ingxube yezakhiwo) Izakhiwo zibalulekile ekunciphiseni izindleko zangaphambi kokuqeqeshwa ngesici esingu-10 ngokuqhathaniswa ne-LLAMA 3 enamapharamitha ayizigidi ezingama-405. Njengamanje, i-V3 iyimodeli yokuqala ebonwa obala esidlangalaleni ekhombisa ama-sparsity aphezulu anjalo eMoe. Ngaphezu kwalokho, i-MLA (i-Multi ungqimba ukunakwa) isebenza ngokuhlanganyela, ikakhulukazi ezicini zokubonisana. "I-sparser i-moe, ubukhulu be-batch edingekayo ngesikhathi sokubonisana ukusebenzisa ngokugcwele amandla okukhawulelwa, ngosayizi we-KVcache kakhulu kube yinto ekhawulelwe," kuphawula umcwaningi ovela kubuchwepheshe be-chuanjon. Sekukonke, impumelelo kaSunseeek ilele ekuhlanganiseni kobuchwepheshe obuhlukahlukene, hhayi eyodwa nje. Abangena ngaphakathi kwemboni badumisa amakhono wobunjiniyela beqembu elijulile, baphawula ubuhle babo ekuqeqesheni okuhambisanayo kanye nokwenza kahle kwe-opharetha, kufinyelela imiphumela ebabazekayo ngokucophelela yonke imininingwane. Indlela evulekile yomthombo ophumelelayo iqhubeka nokuthuthuka okuphelele kwamamodeli amakhulu, futhi kulindeleke ukuthi uma amamodeli afanayo anwebekile abe yizithombe, amavidiyo, nokuningi, lokhu kuzothuthukisa kakhulu embonini embonini.
Amathuba ezinsizakalo zokucabanga komuntu wesithathu
Idatha ikhombisa ukuthi selokhu kwakhululwa, i-Deepseek itholile abasebenzisi abasebenza ngezigidi ezingama-22.15) kungakapheli izinsuku ezingama-50, ngaleyo ndlela ibe yi-ChatGPT I-APPER SPEANS, ngaleyo ndlela iba uhlelo lokusebenza lwe-ChatGPT, i-Topping the apple app emazweni angu-157. Kodwa-ke, ngenkathi abasebenzisi bahlasela ama-Droves, abaduni be-cyber bebelokhu behlasela ngokungapheliyo uhlelo lokusebenza lwe-deefieseed, kubangele ubunzima obukhulu kumaseva alo. Abahlaziyi bezimboni bakholelwa ukuthi ngokwengxenye ngenxa yamakhadi okususela ama-Deepseety ukuqeqeshwa ngenkathi entula amandla okwanele wokuhlangana ukuze acabange. Ukubuyekezwa kwe-Insided Insided Insided Ai Technology, "Izinkinga ezivame ukwenziwa zingaxazululwa kalula ngokushaja imali noma imali yokuthenga imishini eminingi; ekugcineni, kuya ngezinqumo zikaSeveSeek." Lokhu kuveza ukuhweba okuthengiswayo ekugxileni ekukhiqizweni kobuchwepheshe. I-Deepseek ithembele kakhulu kubungako bezinkomba zokuzinakekela, lapho ithole imali encane yangaphandle, okuholela ekucindezelweni okuphansi kwemali okuphansi kanye nemvelo yezobuchwepheshe ende. Njengamanje, ngokukhanya kwezinkinga ezingenhla, abanye abasebenzisi bakhuthaza ukujula kwezokuxhumana ukuze baphakamise imikhawulo yokusebenzisa noma ukwethula izici ezikhokhelwayo ukuthuthukisa induduzo yomsebenzisi. Ngokwengeziwe, abathuthukisi sebeqalile ukusebenzisa i-API esemthethweni noma i-API yesithathu yokwenza kahle. Kodwa-ke, ipulatifomu evulekile kaSeppseeed isanda kumenyezelwa, "Izinsizakusebenza ze-Server zamanje ziyindlala, kanti kumiswe kabusha ama-API Service Rekher."
Akungabazeki ukuthi lokhu kuvula amathuba amaningi kubathengisi abavela eceleni emkhakheni wengqalasizinda ye-AI. Muva nje, iziqhwaga zamafu ezifuywayo kanye namazwe aphesheya azokwethula amamodeli we-deepseeety phesheya kwezilwandle ama-midost microsoft ne-Amazon kwakuphakathi kokuqala ukujoyina ngasekupheleni kukaJanuwari. Umholi wasekhaya, ifu laseHuawei, wenza ukunyakaza kokuqala, ukukhulula izinsizakalo ezijulile ezingama-R1 kanye ne-V3 kanye ne-V3 Ukuqubuka kwezobuchwepheshe kwe-SILICON ngoFebhuwari osuselwa ku-Silicon kuboniswe i-Flowx yabasebenzisi, ngempumelelo "i-Cropking" yesikhulumi. Izinkampani ezinkulu ze-Tech Three Tech, i-baidu, i-Alibaba, iTencent) kanye nokunikezwa okuthe xaxa, okuyisikhumbuzo se-Cloud Stender Price Star, lapho kujule khona i-Cloud's V2 Launch, lapho kujule khona i-Debseek. " Izenzo ze-Frantic zabathengisi befu ama-ties wangaphambilini aqinile phakathi kwe-Microsoft Azure kanye ne-Openai, lapho ubudlelwane be-OpenGPT buthole i-acleach evulekile ngo-2023. Noma kunjalo, lobu budlelwano obukhulu baqala ukukhipha i-Microsoft Azure ecosystem. Kulesi simo, i-Deepseek ayikadediwa nje kuphela engxoxtpt mayelana nokushisa komkhiqizo kepha yethule amamodeli omthombo avulekile alandela ukukhishwa kwe-O1, okufana nentokozo ezungeze ukuvuselelwa kwe-llama ye-GPT-3.
Eqinisweni, abahlinzeki ngamafu nabo bazibeka njengezesango zomgwaqo wezinhlelo zokusebenza ze-AI, okusho ukuthi ukujulisa ubudlelwane nabathuthukisi kuhumusha ngezinzuzo ezivakashile. Imibiko ikhombisa ukuthi ifu le-baidu smart lalinamakhasimende angaphezu kuka-15,000 asebenzisa imodeli ejulile ngepulatifomu ye-qianfan ngosuku lokuqalisa lwemodeli. Ngokwengeziwe, amafemu amancane amancane anikela ngezixazululo, kufaka phakathi ukugeleza okususelwa ku-silicon, ubuchwepheshe beLuchen, ubuchwepheshe be-Chuaning, kanye nabahlinzeki abahlukahlukene be-AI Infra eyethule ukwesekwa kwamamodeli we-Deepseek. Ukubuyekezwa kwezobuchwepheshe kwe-AI kufundile ukuthi amathuba okusebenzisa imali yamanje yokuphakwa kwe-deepseeek ikakhulukazi akhona ezindaweni ezimbili zemodeli ye-moe esebenzisa imodeli ye-moe exubile esebenzisa imodeli ye-moe engu-671 billion Moe / CPU. Ngaphezu kwalokho, ukusebenza kwe-MLA kubalulekile. Kodwa-ke, amamodeli amabili ajulile asabhekene nezinselelo ezithile ekuthumelweni kokwenza kahle. "Ngenxa yosayizi wemodeli namapharamitha amaningi, ukusebenza kahle kuyinkimbinkimbi ngempela, ikakhulukazi ekusetshenzisweni kwasendaweni lapho kufinyelela khona ukulingana okuphezulu phakathi kokusebenza nezindleko kuzoba inselele," kusho umcwaningi kubuchwepheshe beChuaning. Isithiyo esibaluleke kakhulu silele ekunqobeni umkhawulo wememori. "Samukela indlela yokusebenzisana enobuhlakani yokusebenzisa i-CPUs ngokugcwele neminye imithombo ye-computational, ukubeka kuphela izingxenye ezingezona ezabiwe ze-matrix ye-CPU / DRAM yokucubungula usebenzisa i-CPU Performance, ngenkathi izingxenye ezibunjiwe zihlala kwi-GPU," kusho futhi. Imibiko ikhombisa ukuthi abahlelekile bomugqa ovulekile be-chuaning bafaka amasu ahlukahlukene ekusebenzeni kwama-transformers okuqala ngethempulethi, ukuthuthukisa kakhulu isivinini sokuhlobisa usebenzisa izindlela ezinjengezindlela ezinjenge-cudlate. I-Deepseek idale amathuba walezi zibalo, njengoba izinzuzo zokukhula ziba sobala; Amafemu amaningi abike ukukhula kwamakhasimende okubonakalayo ngemuva kokwethula i-Deepseek API, ethola imibuzo evela kumakhasimende angaphambilini afuna ukusebenza. Abangena ngaphakathi emkhakheni baphawulile, "Esikhathini esedlule, amaqembu amaklayenti asunguliwe avame ukuvalelwa ezinkonzweni ezijwayelekile zezinkampani ezinkulu, kepha ngokuzumayo aqede izicelo zokubambisana ezivela kumakhasimende ambalwa owaziwayo, futhi ngisho nalapho amakhasimende akhona ngaphambili asungule oxhumana naye ukwethula izinsizakalo zethu ezijulile." Njengamanje, kuvela ukuthi i-Deepseed yenza ukuthi ukusebenza kwemodeli kusebenza ngokugxeka kakhulu, nangokutholwa okubanzi kwamamodeli amakhulu, lokhu kuzoqhubeka nokuthonya intuthuko embonini ye-AI Infra kakhulu. Uma imodeli ye-Deepseek-Level ingathunyelwa endaweni yangakini ngezindleko eziphansi, kungamsiza kakhulu imizamo kahulumeni kanye ne-Enterprital Transformation imizamo. Kodwa-ke, izinselelo ziyaqhubeka, njengoba amanye amaklayenti angalindela okuthe xaxa maqondana namandla amakhulu amamodeli, okwenza kubonakale ukuthi ukulinganisela kokulinganisa nezindleko kuba balulekile ukuthunyelwa okusebenzayo.
Ukuhlola ukuthi i-Deepseek ingcono kune-chatgpt, kubalulekile ukuqonda ukungezwani kwabo okusemqoka, amandla, nokusebenzisa kanye nokusebenzisa amacala. Nakhu ukuqhathanisa okuningana:
| Isici / Isici | Ukujula | I-chatgt |
|---|---|---|
| Ubunini | Kuthuthukiswe yinkampani yaseChina | Kuthuthukiswe yi-Opena |
| Imodeli yomthombo | Umthombo ovulekile | Ophathelene nokuphathelene |
| Khokhisa | Mahhala ukuyisebenzisa; Izinketho Zokufinyelela Eshibhile API | Okubhaliselwe noma ukukhokha okusetshenziswayo kwamanani entengo |
| Ukwenza ngokwezifiso | Ngokwezifiso kakhulu, ukuvumela abasebenzisi ukuthi basebenze futhi bakha phezu kwalo | Ukwenza ngokwezifiso okulinganiselwe kuyatholakala |
| Ukusebenza emisebenzini ethile | Ama-exlels ezindaweni ezithile njenge-data analytics kanye nokubuyiselwa kolwazi | Ukuguquguquka ngokusebenza okunamandla emisebenzini yokuqamba nemisebenzi yokuxoxa |
| Ukusekelwa Kwezilimi | Gxila kakhulu ngolimi lwesiShayina namasiko | Ukuxhaswa kolimi olubanzi kodwa thina-centric |
| Izindleko zokuqeqesha | Izindleko zokuqeqeshwa eziphansi, ezenzelwe ukusebenza kahle | Izindleko eziphakeme zokuqeqeshwa, ezidinga izinsizakusebenza zezinqubo |
| Ukuhluka kokuphendula | Ingahlinzeka ngezimpendulo ezihlukile, okungenzeka ukuthi ithonywe umongo we-geopolitical | Izimpendulo ezingaguquki ezisuselwa kudatha yokuqeqesha |
| Izithameli ezihlosiwe | Kuhloswe abathuthukisi nabaphenyi bafuna ukuguquguquka | Kuhloswe ngabasebenzisi abajwayelekile abafuna amakhono okuxoxa |
| Sebenzisa amacala | Kusebenza kahle kakhulu ekuqinisekisweni kwamakhodi nemisebenzi esheshayo | Ilungele ukukhiqiza umbhalo, ukuphendula imibuzo, nokuzibandakanya engxoxweni |
Umbono obalulekile "ukuphazamisa Nvidia"
Njengamanje, eceleni kweHuawei, abakhiqizi abaningana basekhaya be-chip bathanda imicu ye-more, muxi, ubuchwepheshe be-biran, kanti uTianca zhixin nawo avumelanisa amamodeli amabili amabili. Umkhiqizi we-chip utshele i-AI Technology Ukubuyekezwa, "Isakhiwo sikaSenseeek sikhombisa ama -stiture, kepha sihlala siyi-LLM. Ukuzivumelanisa ne-Deepseek. Ukusebenza kwezobuchwepheshe kuqondile ngokuqondile futhi kusheshe." Kodwa-ke, indlela ye-MOE idinga izimfuno eziphakeme ngokuya ngokugcina nokusatshalaliswa, okuhambisana nokuqinisekisa ukuhambisana lapho kuthunyelwa ngama-chips ezifuywayo, kuveza izinselelo eziningi zobunjiniyela ezidinga ukulungiswa ngesikhathi sokuzivumelanisa nezimo. "Okwamanje, amandla okuhlangana ekhaya awahambelani ne-nvidia ekusebenzeni nasekusetshenzisweni, okudinga ukubamba iqhaza kwasefektri kwasekuqaleni ukuthola ukusetha kwemvelo yesoftware, ukuxazulula inkinga, kanye nokwenza umsebenzi wesisekelo," kusho ochwepheshe bezindawo abathi ngokuya ngesipiliyoni esisebenzayo. Ngasikhathi sinye, "ngenxa yesilinganiso esikhulu sepharamitha ye-Deepseek R1, amandla okuthuthukisa asekhaya adinga ama-node amaningi wokufana. Ngaphezu kwalokho, i-Huawei 910B okwamanje ayikwazi ukusekela ukutholwa kwe-FP8 eyethulwe yi-Deepseek." Enye yezinto ezivelele ze-Deepseek V3 Model ukwethulwa kohlaka lokuqeqeshwa oluxubile lwe-FP8 oluhlanganisiwe, oluqinisekisiwe ngempumelelo kwimodeli enkulu kakhulu, umaka impumelelo enkulu. Phambilini, abadlali abakhulu njengeMicrosoft neNvidia baphakamisa umsebenzi ohlobene, kepha ukungabaza ngaphakathi emkhakheni ophathelene nokwenzeka. Kuyaqondakala ukuthi uma kuqhathaniswa ne-Int8, inzuzo eyinhloko ye-FP8 yilokho ukwakhiwa kokuqeqeshwa kwangemva kokufinyelela kungafinyelela ukucaciswa okungenamkhawulo ngenkathi kuthuthukisa kakhulu isivinini sokuthathela. Uma uqhathanisa ne-FP16, i-FP8 ingabona kuze kube kabili ukusheshisa amahlandla amabili ku-NVIDIA H20 kanye nokusheshisa izikhathi ezingaphezu kuka-1.5 ku-H100. Ngokuphawulekile, njengoba izingxoxo zizungeze umkhuba wamandla we-Foundenin Computionational Plus athola amamodeli asekhaya athola ukufutheka, ukuqagela ukuthi kungaphazamiseka, nokuthi iCuda Mooat ingadlula yini, iya ngokuya idlula. Iqiniso elilodwa elingenakuqhathaniswa ukuthi ukujula ngempela kubangele ukwehla okukhulu enanini lemakethe yeNvidia, kepha lolu shintsho luphakamisa imibuzo ephathelene nobuqotho bamandla aphezulu e-Nvidia. Ukulandisa okwamukelwe ngaphambili maqondana nokuqongelelwa kwe-capital-eqhutshwa yinqwaba kuphonswa inselelo, kepha kuhlale kunzima ukuthi uNvidia athathelwe indawo ngokugcwele ezimweni zokuqeqesha. Ukuhlaziywa kokusetshenziswa okujulile kwe-Deepseeety ye-Cuda ye-Deepseeek kukhombisa ukuthi ukuguquguquka - njengokusebenzisa i-SM ngokuxhumana noma ukukhohlisa amakhadi wenethiwekhi - akunakwenzeka ukuthi kube ne-GPUS ejwayelekile ukuze kuhlalwe. Ukubuka Kwezimboni kugcizelela ukuthi i-moat yeNvidia ihlanganisa i-Cuda yonke icosystem kunokuba nje i-cuda uqobo, kanye ne-PTX (parallel text eccoution) imiyalo esetshenziswayo ye-deepseed iseyingxenye ye-Cuda Depseed Species iseyingxenye yemvelo ye-Cuda. "Esikhathini esifushane, amandla okuhlangana kweNvidia awakwazi ukwedlula - lokhu kucacile ikakhulukazi ekuqeqeshweni; Sekukonke, ngokuma kombono wokungathathili, izimo zikhuthaza ama-chip amakhulu amamodeli afuywayo. Amathuba abakhiqizi be-chip basekhaya ngaphakathi kwendawo yokuthola amandla abonakala ngokwengeziwe ngenxa yezidingo eziphakeme kakhulu zokuqeqeshwa, ezivimbela ukungena. Abahlaziyi baphikisana nalokhu nje befaka amakhadi amakhadi okukhishwa asekhaya anele; Uma kunesidingo, ukuthola umshini owengeziwe kungenzeka, kanti amamodeli wokuqeqesha abeka izinselelo ezihlukile - ukuphatha inani elikhulayo lemishini kungaba yimiphumela yokuqeqeshwa. Ukuqeqeshwa kubuye kube nezidingo ezithile zesilinganiso seqembu, kuyilapho izimfuno ezimaqenjini zokutholwa zingeqi, ngaleyo ndlela zinciphise izidingo ze-GPU. Njengamanje, ukusebenza kwekhadi le-NVIDIA elithi H20 H20 alikupaki lokho kweHuawei noma i-cambrian; Amandla ayo alele ekuqondeni. Kususelwa kumthelela ophelele emakethe yamandla okuhlanganisa, umsunguli weLuchen Technology, wena wang, waphawula okwesikhashana ukuqeqeshwa okuhambisana nokuqeqeshwa okukhulu okuhambisana nalokhu, ngokunciphisa ukuguquguquka okuhlobene nokuqeqeshwa kwe-AI kususelwa njalo Shayela isidingo esinqunyelwe emakethe yamandla okuhlanganisa. " Ngaphezu kwalokho, "isidingo esikhuphukayo sikaSenseek sokubonisana kanye nezinsizakalo zokuhleleka okuhle sivumelana ngokwengeziwe ngesimo se-computational sendawo, lapho amakhono endawo abuthakathaka, lokhu kudala amathuba asebenzayo abakhiqizi kuwo wonke amazinga ahlukahlukene we-audicational ecosystem." Ubuchwepheshe beLuchen buhlanganyele ne-Huawei Cloud ukwethula i-Deepseek R1 Series Ukubonisana Ukubonisana Ama-API kanye nezinsizakalo zokucabanga ngamafu ezisuselwa emandleni okuhambisa ekhaya. Nina yang nitshengise ithemba ngekusasa: "I-Deepseek igcizelela ukuzethemba ngezixazululo ezikhiqizwe ngaphambi kwesikhathi, ukukhuthaza umdlandla omkhulu nokutshalwa kwezimali kumandla okukhokhwa ekhaya okuya phambili."
Ukugcina
Ukuthi ngabe i-Deentieseek "engcono" kune-chatgpt incike kwizidingo ezithile nezinhloso zomsebenzisi. Ngemisebenzi edinga ukuguquguquka, izindleko eziphansi, nokwenza ngokwezifiso, ukujula kungenzeka kube okuphezulu. Ngokubhala kokudala, uphenyo jikelele, kanye nokuxhumana okuguquguqukayo okuguquguqukayo, i-chatgpt ingahle ihole. Ithuluzi ngalinye lisebenza ngezinhloso ezahlukahlukene, ngakho-ke ukukhetha kuzoncika kakhulu kumongo lapho asetshenziswa khona.
Lawula izintambo
Uhlelo oluhlelekile lwe-cabling
Inethiwekhi nedatha, ikhebula le-fiber-optic, intambo ye-patch, amamojula, puleplate
I-APR.16th-18, 2024 Middle-East-Eader-Eader-East-Eader e Dubai
I-APR.16th-18, 2024 Securika eMoscow
Meyi.9th, 2024 Imikhiqizo emisha & Technologies Laulwa Umcimbi eShanghai
Oct.22ND-25, 2024 Ezokuphepha China eBeijing
Nov.19-20, 2024 I-World KSA Exhunyiwe
Isikhathi sePosi: Feb-10-2025