Don BMS, BUS, Masana'antu, Kebul na Kayan aiki.
Yayin da bikin bazara ya zo kusa, jin daɗin da ke kewaye da DeepSeek ya kasance mai ƙarfi. Biki na baya-bayan nan ya nuna muhimmiyar ma'anar gasa a cikin masana'antar fasaha, tare da tattaunawa da yawa da kuma nazarin wannan "catfish." Silicon Valley yana fuskantar ma'anar rikicin da ba a taɓa ganin irinsa ba: masu fafutukar buɗe ido suna sake bayyana ra'ayoyinsu, har ma OpenAI yana sake kimanta ko dabarun tushen sa shine mafi kyawun zaɓi. Sabuwar yanayin ƙarancin ƙididdiga ta haifar da hatsar sarkar tsakanin tarihin kasuwar ta biyu, yayin da hukumomin gwamnati suna binciken yarda da kwakwalwan kwamfuta amfani da shi. A tsakiyar sake dubawa na DeepSeek a ƙasashen waje, cikin gida, yana samun ci gaba mai ban mamaki. Bayan ƙaddamar da samfurin R1, ƙa'idar da ke da alaƙa ta ga karuwar zirga-zirgar ababen hawa, wanda ke nuna cewa haɓakar sassan aikace-aikacen zai fitar da yanayin yanayin AI gaba ɗaya. Kyakkyawan yanayin shine DeepSeek zai faɗaɗa damar aikace-aikacen, yana ba da shawarar cewa dogaro da ChatGPT ba zai yi tsada ba a nan gaba. An bayyana wannan motsi a cikin ayyukan OpenAI na kwanan nan, gami da samar da samfurin tunani da ake kira o3-mini don masu amfani da 'yanci don amsawa ga DeepSeek R1, da kuma haɓakawa na gaba wanda ya sanya sarkar tunani na o3-mini jama'a. Yawancin masu amfani da ke ketare sun nuna godiya ga DeepSeek don waɗannan ci gaban, kodayake wannan sarkar tunani ta zama taƙaice.
A cikin kyakkyawan fata, a bayyane yake cewa DeepSeek yana haɗa 'yan wasan cikin gida. Tare da mayar da hankali kan rage farashin horarwa, masana'antun keɓaɓɓiyar guntu daban-daban, masu samar da girgije mai tsaka-tsaki, da farawar yawa suna shiga cikin yanayin yanayin, suna haɓaka ƙimar farashi don amfani da ƙirar DeepSeek. Dangane da takaddun DeepSeek, cikakken horon samfurin V3 yana buƙatar sa'o'i miliyan 2.788 H800 GPU kawai, kuma tsarin horon yana da ƙarfi sosai. Gine-ginen MoE (Gaurayawan Kwararru) yana da mahimmanci don rage farashin horarwa da kashi goma idan aka kwatanta da Llama 3 tare da sigogi biliyan 405. A halin yanzu, V3 ita ce samfurin farko da aka gane a bainar jama'a wanda ke nuna irin wannan babban rashi a cikin MoE. Bugu da ƙari, MLA (Multi Layer Attention) yana aiki tare, musamman a fannin tunani. "Mafi girman girman MoE, mafi girman girman girman da ake buƙata yayin tunani don cikakken amfani da ikon ƙididdigewa, tare da girman KVCache shine maɓalli mai iyakancewa; MLA yana rage girman KVCache sosai, "in ji wani mai bincike daga Chuanjing Technology a cikin bincike don Binciken Fasaha na AI. Gabaɗaya, nasarar DeepSeek ta ta'allaka ne a cikin haɗin fasaha daban-daban, ba guda ɗaya kawai ba. Masu cikin masana'antu suna yaba iyawar aikin injiniya na ƙungiyar DeepSeek, tare da lura da ƙwararrunsu a cikin horo iri ɗaya da haɓaka aikin ma'aikata, suna samun sakamako mai ban mamaki ta hanyar tace kowane daki-daki. Hanyar buɗe hanyar DeepSeek tana ƙara haɓaka haɓakar manyan samfura gaba ɗaya, kuma ana tsammanin idan irin wannan ƙirar ta faɗaɗa zuwa hotuna, bidiyo, da ƙari, wannan zai ƙara haɓaka buƙatu a cikin masana'antar.
Dama don Sabis na Tunani na ɓangare na uku
Bayanai sun nuna cewa tun lokacin da aka saki shi, DeepSeek ya tara masu amfani da aiki na yau da kullun (DAU) miliyan 22.15 a cikin kwanaki 21 kacal, wanda ya kai kashi 41.6% na tushen mai amfani da ChatGPT kuma ya zarce miliyan 16.95 masu amfani da Doubao na yau da kullun, don haka ya zama aikace-aikacen da ya fi saurin girma a duniya, wanda ke kan gaba da Apple App Store a cikin kasashe 15.7. Duk da haka, yayin da masu amfani suka yi ta tururuwa, masu satar yanar gizo sun ci gaba da kai hare-hare a kan DeepSeek app, suna haifar da matsala a kan sabobin sa. Manazarta masana'antu sun yi imanin wannan wani bangare ne saboda DeepSeek tura katunan don horarwa yayin da ba shi da isasshen ikon lissafi don tunani. Wani masanin masana'antu ya sanar da AI Technology Review, "Ana iya magance matsalolin uwar garken akai-akai cikin sauƙi ta hanyar cajin kudade ko samar da kudade don siyan ƙarin inji; a ƙarshe, ya dogara da shawarar DeepSeek." Wannan yana gabatar da ciniki a cikin mayar da hankali kan fasaha tare da samarwa. DeepSeek ya dogara da yawa akan ƙididdige ƙididdigewa don dogaro da kai, bayan samun kuɗi kaɗan daga waje, wanda ya haifar da ƙarancin matsi na tsabar kuɗi da ingantaccen muhallin fasaha. A halin yanzu, dangane da matsalolin da aka ambata, wasu masu amfani suna roƙon DeepSeek akan kafofin watsa labarun don haɓaka ƙofofin amfani ko gabatar da fasalin da aka biya don haɓaka ta'aziyyar mai amfani. Bugu da ƙari, masu haɓakawa sun fara amfani da API na hukuma ko APIs na ɓangare na uku don ingantawa. Koyaya, sabon dandalin DeepSeek ya sanar kwanan nan, "Kayan aikin sabar na yanzu ba su da yawa, kuma an dakatar da cajin sabis na API."
Wannan babu shakka yana buɗe ƙarin dama ga masu siyarwa na ɓangare na uku a cikin ɓangaren abubuwan more rayuwa na AI. Kwanan nan, ɗimbin gizagizai na cikin gida da na ƙasa da ƙasa sun ƙaddamar da samfurin APIs na DeepSeek—Gwayoyin ƙetare Microsoft da Amazon na cikin waɗanda suka fara shiga a ƙarshen Janairu. Jagoran cikin gida, Huawei Cloud, ya yi motsi na farko, yana sakin ayyukan tunani na DeepSeek R1 da V3 tare da haɗin gwiwar Flow na tushen Silicon a ranar 1 ga Fabrairu. Rahotanni daga AI Technology Review sun nuna cewa ayyukan Flow na tushen Silicon sun ga yawan masu amfani, da kyau "raguwa" dandamali. Manyan kamfanonin fasaha guda uku-BAT (Baidu, Alibaba, Tencent) da ByteDance - suma sun ba da tayin farashi mai rahusa, ƙayyadaddun lokaci wanda zai fara daga ranar 3 ga Fabrairu, wanda ke tunawa da yaƙe-yaƙe na farashin gajimare na bara wanda DeepSeek's V2 samfurin ƙaddamar da ƙaddamarwa, inda DeepSeek ya fara zama lakabin "farashin mahauta." Ayyukan dillalai na girgije sun yi daidai da dangantakar da ta gabata tsakanin Microsoft Azure da OpenAI, inda a cikin 2019, Microsoft ya kashe dala biliyan 1 a cikin OpenAI kuma ya sami fa'ida bayan ƙaddamar da ChatGPT a cikin 2023. Duk da haka, wannan kusancin ya fara lalacewa bayan Meta ya buɗe Llama, yana ba da damar sauran masu siyar da kayayyaki a waje da tsarin su na Microsoft. A cikin wannan misalin, DeepSeek ba wai kawai ya zarce ChatGPT dangane da zafin samfur ba amma kuma ya gabatar da samfuran buɗe ido biyo bayan sakin o1, kama da jin daɗin da ke tattare da farfado da Llama na GPT-3.
A zahiri, masu samar da girgije kuma suna sanya kansu azaman ƙofofin zirga-zirga don aikace-aikacen AI, ma'ana cewa zurfafa alaƙa da masu haɓakawa suna fassara zuwa fa'idodi masu ƙima. Rahotanni sun nuna cewa Baidu Smart Cloud yana da abokan ciniki sama da 15,000 da ke amfani da samfurin DeepSeek ta dandalin Qianfan a ranar ƙaddamar da samfurin. Bugu da ƙari, ƙananan kamfanoni da yawa suna ba da mafita, gami da Silicon-tushen Flow, Fasahar Luchen, Fasahar Chuanjing, da masu samar da AI Infra iri-iri waɗanda suka ƙaddamar da tallafi ga ƙirar DeepSeek. Binciken Fasaha na AI ya koyi cewa damar ingantawa na yanzu don ƙaddamar da aikin DeepSeek da farko ya kasance a cikin yankuna biyu: ɗayan yana inganta halayen ƙima na ƙirar MoE ta amfani da hanyar haɗaɗɗiyar tunani don ƙaddamar da ƙirar MoE na biliyan 671 a cikin gida yayin amfani da haɗin GPU/CPU na gaba. Bugu da ƙari, haɓakar MLA yana da mahimmanci. Koyaya, samfuran DeepSeek guda biyu har yanzu suna fuskantar wasu ƙalubale wajen inganta turawa. "Saboda girman samfurin da ma'auni masu yawa, haɓakawa yana da wuyar gaske, musamman ga turawa cikin gida inda samun daidaito mafi kyau tsakanin aiki da farashi zai zama kalubale," in ji wani mai bincike daga Chuanjing Technology. Mafi mahimmancin matsala yana ta'allaka ne a cikin shawo kan iyakoki na ƙwaƙwalwar ajiya. "Muna amfani da tsarin haɗin gwiwa daban-daban don amfani da CPUs da sauran albarkatu na lissafi, muna sanya ɓangarorin da ba a raba su ba na matrix na MoE akan CPU / DRAM don aiki ta amfani da manyan ma'aikatan CPU, yayin da yawancin sassan ke kan GPU," in ji shi. Rahotanni sun nuna cewa tsarin bude tushen Chuanjing KTransformers da farko yana shigar da dabaru daban-daban da masu aiki a cikin ainihin aiwatar da Transformers ta hanyar samfuri, yana haɓaka saurin ƙima ta amfani da hanyoyi kamar CUDAGraph. DeepSeek ya haifar da dama ga waɗannan farawa, kamar yadda fa'idodin haɓaka ke bayyana; kamfanoni da yawa sun ba da rahoton ci gaban abokin ciniki bayan ƙaddamar da DeepSeek API, suna karɓar tambayoyi daga abokan cinikin da suka gabata suna neman ingantawa. Masana'antu insiders sun lura, "A baya, da ɗan kafa abokin ciniki kungiyoyin da aka sau da yawa kulle a cikin daidaitattun ayyuka na manyan kamfanoni, tam daure da su kudin abũbuwan amfãni saboda sikelin. Duk da haka, bayan kammala da tura DeepSeek-R1 / V3 a gaban Spring Festival, ba zato ba tsammani samu hadin gwiwa buƙatun daga dama sanannun abokan ciniki, kuma ko da a baya dormant abokan ciniki qaddamar da sabis na mu DeepSeek. " A halin yanzu, ya bayyana cewa DeepSeek yana yin aikin ƙaddamar da ƙima yana ƙara mahimmanci, kuma tare da ɗaukar manyan samfura, wannan zai ci gaba da yin tasiri ga ci gaba a masana'antar AI Infra sosai. Idan samfurin matakin DeepSeek za a iya tura shi cikin gida akan farashi mai rahusa, zai taimaka wa gwamnati da yunƙurin sauya dijital na kasuwanci. Koyaya, ƙalubalen suna ci gaba, yayin da wasu abokan ciniki na iya riƙe babban tsammanin game da manyan damar ƙirar ƙira, yana mai da shi ƙarara cewa daidaita aiki da farashi ya zama mahimmanci a aiwatar da aiki.
Don kimanta ko DeepSeek ya fi ChatGPT kyau, yana da mahimmanci a fahimci mahimman bambance-bambancen su, ƙarfi, da shari'o'in amfani. Ga cikakken kwatance:
| Siffar/Hanyar | DeepSeek | Taɗi GPT |
|---|---|---|
| Mallaka | Wani kamfani na kasar Sin ne ya samar da shi | OpenAI ne ya haɓaka |
| Tushen Model | Bude tushen | Na mallaka |
| Farashin | Kyauta don amfani; zaɓuɓɓukan samun API mai rahusa | Biyan kuɗi ko farashin biyan-da-amfani |
| Keɓancewa | Ana iya daidaita shi sosai, yana bawa masu amfani damar tweak da ginawa akansa | Akwai iyakantaccen keɓancewa |
| Ayyuka a cikin Musamman Ayyuka | Excels a wasu wurare kamar nazarin bayanai da dawo da bayanai | M tare da aiki mai ƙarfi a cikin rubutun ƙirƙira da ayyukan tattaunawa |
| Taimakon Harshe | Mai da hankali sosai kan harshe da al'adun Sinanci | Faɗin tallafin harshe amma na tsakiya na Amurka |
| Kudin horo | Ƙananan farashin horo, an inganta shi don dacewa | Haɓaka farashin horo, yana buƙatar kayan aikin lissafi |
| Bambancin Amsa | Yana iya ba da amsoshi daban-daban, mai yuwuwa mahallin geopolitical ya rinjayi | Amsoshi masu daidaituwa bisa bayanan horo |
| Masu sauraro manufa | An yi nufin masu haɓakawa da masu bincike masu son sassauci | An yi niyya ga masu amfani gabaɗaya suna neman damar tattaunawa |
| Amfani da Cases | Mafi inganci don tsara lambar da ayyuka masu sauri | Mafi dacewa don ƙirƙirar rubutu, amsa tambayoyin, da shiga cikin tattaunawa |
Mahimman ra'ayi akan "Rikicin Nvidia"
A halin yanzu, baya ga Huawei, masana'antun guntu na cikin gida da yawa kamar Moore Threads, Muxi, Biran Technology, da Tianxu Zhixin suma suna daidaitawa da ƙirar DeepSeek guda biyu. Wani mai yin guntu ya gaya wa AI Technology Review, "Tsarin DeepSeek yana nuna ƙididdigewa, duk da haka ya kasance LLM. Daidaitawarmu ga DeepSeek ya fi mayar da hankali kan aikace-aikacen tunani, yin aiwatar da fasaha mai sauƙi da sauri." Koyaya, tsarin MoE yana buƙatar buƙatu masu girma dangane da ajiya da rarrabawa, haɗe tare da tabbatar da dacewa yayin turawa tare da kwakwalwan gida, suna gabatar da ƙalubalen injiniya da yawa waɗanda ke buƙatar ƙuduri yayin daidaitawa. "A halin yanzu, ikon lissafin gida bai dace da Nvidia a cikin amfani da kwanciyar hankali ba, yana buƙatar sa hannun masana'anta na asali don saitin yanayin software, magance matsala, da haɓaka aikin tushe," in ji wani ƙwararren masana'antu dangane da ƙwarewar aiki. A lokaci guda, "Saboda babban ma'auni na DeepSeek R1, ikon lissafin gida yana buƙatar ƙarin nodes don daidaitawa. Bugu da ƙari, ƙayyadaddun kayan aikin gida har yanzu suna da ɗan baya; misali, Huawei 910B a halin yanzu ba zai iya goyan bayan FP8 da DeepSeek ya gabatar ba." Ɗaya daga cikin mahimman bayanai na ƙirar DeepSeek V3 shine ƙaddamar da tsarin horarwa na FP8 gauraye, wanda aka inganta yadda ya kamata akan babban tsari mai girma, wanda ke nuna gagarumar nasara. A baya can, manyan 'yan wasa kamar Microsoft da Nvidia sun ba da shawarar aikin da ke da alaƙa, amma shakku na daɗe a cikin masana'antar dangane da yuwuwar. An fahimci cewa idan aka kwatanta da INT8, babban fa'idar FP8 shine cewa ƙididdige horo bayan horo na iya cimma daidaitaccen rashin asara tare da haɓaka saurin ƙima. Lokacin kwatanta da FP16, FP8 na iya gane haɓakawa har sau biyu akan Nvidia's H20 da sama da sau 1.5 akan H100. Musamman ma, yayin da tattaunawar da ke tattare da yanayin ikon lissafin cikin gida tare da samfuran gida ke samun ci gaba, hasashe game da ko Nvidia za a iya rushewa, da kuma ko za a iya wuce gona da iri na CUDA, yana ƙara yaɗuwa. Wata gaskiyar da ba za a iya musantawa ita ce DeepSeek da gaske ya haifar da raguwa mai yawa a darajar kasuwar Nvidia, amma wannan canjin yana haifar da tambayoyi game da amincin ikon lissafin ƙimar Nvidia. Ana ƙalubalantar labarun da aka karɓa a baya game da tara lissafin babban jari, duk da haka yana da wahala a sami cikakken maye gurbin Nvidia a yanayin horo. Binciken zurfin amfani da DeepSeek na CUDA ya nuna cewa sassauci-kamar amfani da SM don sadarwa ko sarrafa katunan cibiyar sadarwa kai tsaye-ba shi yiwuwa ga GPUs na yau da kullun don ɗauka. Ra'ayoyin masana'antu sun jaddada cewa ma'aunin Nvidia ya ƙunshi duk yanayin yanayin CUDA maimakon CUDA da kanta, da kuma PTX (Parallel Thread Execution) da umarnin da DeepSeek ke amfani da shi har yanzu wani ɓangare ne na yanayin CUDA. "A cikin ɗan gajeren lokaci, ba za a iya ƙetare ikon lissafin Nvidia ba - wannan ya bayyana a cikin horo; duk da haka, ƙaddamar da katunan gida don tunani zai zama mafi sauƙi, don haka ci gaba zai iya zama da sauri. Daidaitawar katunan cikin gida da farko ya fi mayar da hankali ga ƙaddamarwa; babu wanda ya riga ya sami nasarar horar da samfurin DeepSeek a kan katunan gida a ma'auni, "Masana nazarin masana'antu ya yi la'akari da AI Technology Review. Gabaɗaya, daga hangen nesa, yanayi yana ƙarfafa ga manyan guntu na ƙirar gida. Damar ga masana'antun guntu na cikin gida a cikin yanayin zance sun fi bayyana saboda manyan buƙatun horo, waɗanda ke hana shigowa. Manazarta sun yi iƙirarin cewa yin amfani da katunan ba da ƙima a cikin gida kawai ya wadatar; idan ya cancanta, samun ƙarin inji abu ne mai yuwuwa, yayin da ƙirar horarwa ke haifar da ƙalubale na musamman - sarrafa ƙarin adadin injuna na iya zama nauyi, kuma ƙimar kuskure mafi girma na iya yin mummunan tasiri ga sakamakon horo. Har ila yau horarwa yana da takamaiman buƙatun ma'aunin gungu, yayin da buƙatun kan gungu don tantancewa ba su da ƙarfi, don haka sauƙaƙe buƙatun GPU. A halin yanzu, aikin katin H20 na Nvidia guda ɗaya bai wuce na Huawei ko Cambrian ba; Ƙarfinsa yana cikin tari. Dangane da tasirin gabaɗaya akan kasuwar wutar lantarki, wanda ya kafa Luchen Technology, You Yang, ya lura a cikin wata hira da AI Technology Review, "DeepSeek na iya ɗan ɗan lokaci rushe kafa da hayar manyan tarurrukan ƙididdiga na horarwa. bukata a kasuwar wutar lantarki." Bugu da ƙari, "Ƙarin buƙatun DeepSeek na tunani da ayyukan daidaitawa ya fi dacewa da yanayin lissafin cikin gida, inda ƙarfin gida ke da rauni sosai, yana taimakawa wajen rage sharar gida daga albarkatun da ba su da aiki bayan kafa gungu; wannan yana haifar da damammaki ga masana'antun a cikin matakai daban-daban na tsarin lissafin gida." Fasahar Luchen ta haɗu tare da Huawei Cloud don ƙaddamar da DeepSeek R1 jerin dalilai APIs da sabis na hoto na girgije dangane da ikon lissafin gida. You Yang ya bayyana kyakkyawan fata game da nan gaba: "DeepSeek yana sanya kwarin gwiwa kan hanyoyin da aka samar a cikin gida, tare da karfafa himma da saka hannun jari a cikin ayyukan lissafin cikin gida da ke gaba."
Kammalawa
Ko DeepSeek ya "mafi kyau" fiye da ChatGPT ya dogara da takamaiman buƙatu da manufofin mai amfani. Don ayyukan da ke buƙatar sassauci, ƙarancin farashi, da keɓancewa, DeepSeek na iya zama mafi girma. Don ƙirƙira rubuce-rubuce, bincike na gabaɗaya, da mu'amalar abokantaka na mai amfani, ChatGPT na iya ɗaukar jagoranci. Kowane kayan aiki yana amfani da dalilai daban-daban, don haka zaɓin zai dogara sosai akan yanayin da ake amfani da su.
Sarrafa igiyoyi
Tsarin Caling System
Network&Bayani, Fiber-Optic Cable, Patch Cord, Modules, Faceplate
Afrilu 16th-18th, 2024 Gabas ta Tsakiya-Makamashi a Dubai
Afrilu 16-18, 2024 Securika a Moscow
Mayu.9th, 2024 SABBIN KYAKKYAWAN KYAUTATA DA FASAHA A KAN KADDAMAR DA BISA A Shanghai
Oct.22nd-25th, 2024 TSARO CHINA a birnin Beijing
Nov.19-20, 2024 HADAKAR DUNIYA KSA
Lokacin aikawa: Fabrairu-10-2025