Blog

Menene NPU?

2024-08-13 16:29:49

A cikin saurin haɓakawa a yau AI-kore duniya, Buƙatar ƙididdiga mai sauri, mafi inganci ya haifar da na'urori na musamman kamar su Sashin sarrafa Jijiya (NPU). Sabanin gargajiya CPUs ko GPUs, NPU shine manufa-gina don haɓakawa inji koyan inference, musamman don zurfin hanyoyin sadarwa na jijiyoyi amfani a hangen nesa na kwamfuta, sarrafa harshe na halitta, kuma gane murya.

An NPU nau'i ne na AI accelerator wanda ke ba da babban aiki, aiki mara ƙarancin aiki don ayyukan da ke buƙata layi daya lissafi kuma ayyukan tensor. Waɗannan na'urori masu sarrafawa yanzu an haɗa su a cikin dandamali da yawa-daga wayoyin komai da ruwanka kuma na'urorin IoT ku tsarin motoci kuma gefen kwamfuta kayayyakin more rayuwa- kunnawa kan na'urar AI tare da ƙarancin wutar lantarki.

Me yasa NPUs ke da mahimmanci:

Ingantattun ayyuka na lokaci-lokaci don aikin AI
Ƙananan amfani da makamashi idan aka kwatanta da mafita na tushen GPU
Rage dogara ga gajimare, inganta bayanan sirri da latency
Ƙimar ƙarfi don ƙananan na'urori da tsarin da aka haɗa

Tashi na gaba AI da yawaitar Aikace-aikace masu ƙarfin AI sanya mahimmanci don fahimtar menene NPU da yadda ya bambanta da sauran sassan sarrafawa. A cikin wannan labarin, mun bincika da gine-gine, amfani lokuta, da ƙimar dabarun NPUs a cikin faffadan mahallin kwamfuta mai hankali.

Menene NPU? Bayanin Fasaha

A Sashin sarrafa Jijiya (NPU) keɓaɓɓen microprocessor ne wanda aka ƙera musamman don ɗauka ilimin artificial (AI) kayan aiki na musamman, wanda ke da alaƙa zurfafa ilmantarwa kuma Neural cibiyar sadarwa inference. Sabanin manufa ta gaba ɗaya CPUs ko graphics-daidaitacce GPUs, NPUs an inganta su don ayyukan lissafin da ke iko samfurin koyon inji, kamar matrix multiplications, rikice-rikice, kuma sarrafa tensor.

Babban Halayen NPU:

Manufar-gina don ƙaddamar da AI, ba horo
Yana kashewa a layi daya ayyuka yadda ya kamata
An inganta don low-latency, high-throughput ayyukan aiki
Yana aiki da mahimmanci ƙananan amfani da wutar lantarki fiye da GPUs

Yayin CPUs su ne manufa domin jerin dabaru da kuma GPUs ƙware wajen yin ayyuka da ayyukan caca, NPUs an keɓance su don haɓaka ayyuka kamar:

Gano abu
Gane fuska
Canjin murya-zuwa-rubutu
Smart kamara nazari

Nau'in Mai sarrafawa	Matsayin Farko	Ƙarfi	Ƙarfin Ƙarfi	AI Dace
CPU	Babban ayyuka	M	Ƙananan	Iyakance
GPU	Graphics, AI	Babban kayan aiki	Matsakaici	Mai ƙarfi don horo
NPU	Bayanin AI	AI- ingantacce	Babban	Mafi kyawun zaɓi na AI

Yayin da AI ke shiga cikin ƙarin na'urori-daga wayoyin komai da ruwanka ku masana'antu IoT-da NPU tsaya a matsayin mafi inganci kuma mai iya daidaitawa na'ura mai sarrafawa don kunna fasalulluka kai tsaye a cikin baki, ba tare da dogaro da samun damar gajimare akai-akai ba.

Ta yaya NPU ke aiki?

A Sashin sarrafa Jijiya (NPU) an ƙera shi don aiwatarwa Ayyukan inference AI tare da na kwarai gudun da inganci. Ba kamar na'urori masu sarrafawa na gargajiya ba, NPUs suna aiki Tsarin bayanai na tushen tensor, aiwatar da ɗimbin yawa matrix ayyuka a layi daya-mahimmanci don gudanar da zamani samfurin ilmantarwa mai zurfi kamar CNNs (Convolutional Neural Networks) kuma RNNs (Cibiyoyin Cibiyoyin Jijiya na Maimaituwa).

Ƙa'idodin Aiki na Musamman:

Daidaituwa: NPUs suna aiki ayyuka da yawa a lokaci guda, sosai rage aiki lokaci don ayyuka kamar rarraba hoto kuma gane magana.
Ƙididdigar ƙididdige ƙididdiga na tensor-hardware: Gina-ciki tensor muryoyin yadda ya kamata rike manyan-sikelin matrix multiplications.
Ƙarƙashin ƙididdiga: Yawancin NPUs suna amfani da su INT8 ko FP16 daidaitattun tsari don daidaita daidaito tsakanin daidaito kuma karfin wutar lantarki.
Bututun gine-gine: Ayyuka kamar samun data, sarrafa nauyi, kuma lissafin kunnawa ana aiwatar da su a cikin jerin bututun mai don haɓaka kayan aiki.

Sauƙaƙe Tafiya:

Bayanan shigarwa (misali, hoto ko siginar sauti) an riga an sarrafa shi.
Data shiga cikin Injin Tensor na NPU don lissafin jijiya.
Samfurin ta nauyi da son zuciya shiryar da inference.
Ana isar da sakamako a cikin ainihin lokaci, galibi ba tare da taimakon girgije ba.

Ta hanyar haɗawa low-latency kisa, ƙira mai inganci, kuma sadaukar AI dabaru, NPUs suna ba da damar na'urori suyi aiki hadaddun AI model na gida. Wannan ya sa su zama makawa gefen kwamfuta, na'urori masu wayo, kuma masana'antu AI aikace-aikace ina ainihin yanke shawara yana da mahimmanci.

Mabuɗin Amfanin NPUs

The tallafi na Rukunin sarrafa Jijiya (NPUs) ya kawo sauyi da turawa Bayanin AI a duka gefen na'urorin kuma yanayin girgije. NPUs suna ba da a na musamman sarrafa gine-gine wanda ke ba da damar yin aiki mai mahimmanci da samun ingantaccen makamashi akan na gargajiya CPU kuma GPU mafita lokacin aiwatarwa samfurin ilmantarwa mai zurfi.

Babban fa'idodin Amfani da NPUs:

Babban Ayyuka don Ƙaddamar AI
NPUs an gina su don dalilai AI kayan aiki, aiwatarwa ayyukan sadarwa na jijiyoyi tare da mafi girma gudu da ƙananan latency. Suna aiwatarwa daidaitattun ƙididdigar tensor yadda ya kamata, isar da ainihin-lokaci fahimta a aikace-aikace kamar gano abu, rabuwar hoto, kuma gane magana.
Ingantaccen Makamashi
NPUs suna cinye ƙarancin ƙarfi fiye da GPUs, godiya ga su ƙananan madaidaicin lissafi (misali, INT8, FP16) da ingantattun gine-gine. Wannan ya sa su manufa domin mobile AI kuma na'urorin IoT ina rayuwar baturi kuma thermal iyaka suna da mahimmanci.
Kan-Na'urar AI Capabilities
Ta hanyar kunnawa la'akari na gida, NPUs sun rage buƙatar sarrafa girgije, ingantawa sirrin bayanai, lokacin amsawa, kuma ingancin bandwidth cibiyar sadarwa.
Ƙirƙirar ƙira mai ƙima
Ana iya haɗa NPUs a ciki Tsarin-on-Chip (SoC) kayayyaki, kyale masana'antun su saka AI accelerators cikin m na'urori kamar wayoyin komai da ruwanka, masu sawa, bakin kofa, kuma m tsarin.

Teburin Kwatanta: NPU vs. GPU da CPU

Siffar	NPU	GPU	CPU
Manufar	Bayanin AI	Graphics, general AI	Ƙididdigar manufa ta gaba ɗaya
Ƙarfin Ƙarfi	⭐⭐⭐⭐⭐	⭐⭐	⭐
Latency (ainihin lokaci)	Ƙananan	Matsakaici	Babban
Parallel AI Ayyukan Aiki	Madalla	Yayi kyau sosai	Iyakance
Ideal Case Amfani	Edge AI, Mobile AI	AI horo, graphics	Sarrafa dabaru, OS ayyuka

Ta hanyar haɗa NPUs cikin mafitacin lissafin AI gefen, kasuwanci na iya buɗe ƙwarewar koyon injin mai sauri tare da ƙananan farashin aiki, yin NPUs tushen tushen kayan aikin AI na gaba.

NPU vs. CPU, GPU, da TPU: Kwatancen Kwatancen

Kamar yadda ake bukata ainihin lokacin AI inference yana girma a cikin na'urori - daga wayoyin hannu da motoci masu cin gashin kansu zuwa injiniyoyin masana'antu - muhawara kan wanne na'ura mai sarrafa kayan aiki yana ƙaruwa. Yayin CPUs, GPUs, TPUs, kuma NPUs duk iya aiwatarwa koyon inji ayyuka, an inganta su don dalilai daban-daban. Fahimtar waɗannan bambance-bambance yana da mahimmanci yayin zabar kayan aiki AI aikace-aikace, musamman a wurin baki.

CPU (Sashin sarrafawa ta tsakiya)

The CPU ya kasance mafi m processor. Yana gudanar da ayyuka masu yawa na ƙididdiga na gaba ɗaya, gami da ayyukan tsarin aiki, Gudanar da I/O, kuma dabara dabaru. Duk da yake CPUs na iya yin aikin AI ta hanyar fasaha, ba a inganta su ba a layi daya aiki cewa hanyoyin sadarwa na jijiyoyi suna buƙata.

GPU (Sashin Gudanar da Zane-zane)

Da farko an haɓaka don yin hotuna da bidiyo, da GPU ya tabbatar da tasiri sosai ga a layi daya ayyuka, yin shi dace da duka biyu AI horo kuma ra'ayi. GPUs sun yi fice a ciki muhallin cibiyar bayanai, inda ake amfani da wutar lantarki da sarari.

TPU (Tsarin sarrafa Tensor)

Google ne ya haɓaka, da TPU an keɓe don lissafin tensor a cikin aikin AI. Yana ba da horo mai ƙarfi da aikin tantancewa amma ana samunsa da farko a ciki girgije kayayyakin more rayuwa ko hardware-sa kayan aiki.

NPU (Sashin sarrafa Jijiya)

An tsara shi daga ƙasa har zuwa Bayanin AI, da NPU yana ba da aikin da bai dace da kowace watt ba. Yanzu an haɗa shi sosai a ciki Tsarin-on-Chip (SoC) ƙira a cikin wayoyin hannu, na'urorin gefen AI, da injuna masu zaman kansu.

Teburin Kwatancen Mai sarrafawa

Siffar	CPU	GPU	TPU	NPU
An tsara don	Babban manufa	Graphics & AI	Tensor Ops	AI Inference
AI horo	Talakawa	Madalla	Yayi kyau sosai	Iyakance
AI Inference	Iyakance	Yayi kyau	Madalla	Madalla
Ƙarfin Ƙarfi	Ƙananan	Matsakaici	Ƙananan	Babban
Latency (Ayyukan Edge)	Babban	Matsakaici	Ƙananan	Ƙarƙashin Ƙasa
Dacewar Wayar hannu/IoT	Talakawa	Talakawa	Iyakance	Madalla
Tsarin Tallafi	Fadi	Fadi	kunkuntar (Google)	Matsakaici (ya bambanta)

Key Takeaways:

GPUs an fi son AI model horo kuma girgije kwamfuta.
TPUs su ne girgije-tsakiyar kuma mafi kyau a cikin Google AI muhalli.
CPUs su ne gama gari, ba su dace da zurfin koyo ba.
NPUs su ne tafi-zuwa zabi ga kan na'urar AI, miƙa da mafi kyawun ma'auni na sauri, inganci, da ƙananan ƙarfi don ƙaddamar da kayan aiki.

Kamar yadda AI ke ci gaba da motsawa zuwa baki, NPU ya fito waje a matsayin mafi kyawun mafita na dogon lokaci don ainihin lokacin, koyo na inji.

Ina Ana Amfani da NPUs? Maɓallin Aikace-aikace

Rukunin sarrafa Jijiya (NPUs) yanzu sun kasance bangaren tushe a cikin kewayon da yawa Fasahar AI-kore, musamman wadanda ake bukata nuni akan na'urar, karancin wutar lantarki, kuma ainihin lokacin amsawa. Kamar yadda masana'antu ke ci gaba da hadewa basirar wucin gadi A cikin ayyukan su, ana karɓar NPUs a duk faɗin biyun masu amfani da lantarki kuma tsarin masana'antu.

1. Wayoyin hannu da na'urorin hannu

Na zamani wayoyin komai da ruwanka an sanye su da NPUs don ci gaba da wutar lantarki AI fasali kamar:

Gane fuska (misali, tantancewar biometric)
Haɓaka hoto kuma gano abu a cikin kyamara apps
Mataimakan murya kuma fassarar harshe
Ƙarfafa gaskiya (AR) bayarwa a ainihin lokacin

Manyan kwakwalwan kwamfuta na wayar hannu kamar Injin Neural na Apple's A-Series, Qualcomm Snapdragon, kuma Huawei's Kirin SoC Haɗa NPUs don inganci gefen AI aiki.

2. Edge Computing da IoT

A cikin duniyar gaba AI, NPUs kunna kyamarori masu wayo, na'urori masu sawa, kuma cibiyoyi masu sarrafa kansa na gida don gudanar da hadaddun samfurin koyon inji na gida. Wannan yana rage dogara ga sabobin girgije kuma yana tabbatar da sauri yanke shawara, har ma a wuraren da ke da iyakacin haɗin kai.

3. Motoci da Robotics

Ana ƙara tura NPUs a ciki motoci masu zaman kansu kuma tsarin robotics don aiwatarwa:

Fuskar Sensor
Gano layi
Sanin masu tafiya a ƙasa
Tsarin umarnin murya

Ta hanyar gudanar da bincike kai tsaye akan abin hawa ko robot, NPUs sun tabbatar aiki na ainihi tare da ƙarancin latti.

4. Masana'antu da Lafiya AI

Hakanan ana amfani da NPUs a ciki masana'anta sarrafa kansa, kula da tsinkaya, kuma hoto na likita. A ciki mai kaifin masana'antu, NPUs kunna AI hangen nesa tsarin don gano lahani, rarraba kayan aiki, da saka idanu akan ayyukan ci gaba ba tare da buƙatar haɗin haɗin girgije mai girma ba.

Manyan Dillalan Hardware na NPU da Ecosystem

Girman AI a gefe ya haifar da zuba jari mai yawa a ciki NPU hardware ci gaban ta manyan kamfanonin fasaha. Waɗannan dillalai suna tsara yanayin yanayin shigar AI ta hanyar bayar da ingantattun mafita waɗanda aka inganta don saurin fahimta, karfin wutar lantarki, kuma sassaucin haɗin kai. Kowane iri yana da nasa NPU gine, kayan aiki, da goyon bayan tsarin muhalli.

1. Apple - Injin Jijiya

Apple's Injin Jijiya, hadedde cikinsa A-jerin da kuma M-jerin kwakwalwan kwamfuta, yana iko da ayyukan AI masu ci gaba a cikin iPhones, iPads, da Macs. Yana bayarwa:

Real-lokaci gane fuska
sarrafa Siri akan na'urar
Haɓaka hoto mai wayo da bidiyo

2. Huawei - Da Vinci Architecture

Huawei ta Farashin NPUs kuma Da Vinci architecture su ne tsakiyar ta Kirin SoCs kuma AI kayayyakin more rayuwa:

Haɓaka ƙaddamar AI a ciki na'urorin hannu
turawa cikin cibiyoyin bayanai don kasuwancin AI
Yana goyan bayan MindSpore AI tsarin

3. Qualcomm - Hexagon DSP tare da AI Engine

Qualcomm's Hexagon DSP + AI Engine An saka shi a cikin dandamali na Snapdragon, yana ba da:

AI hanzari don Android apps
Ingantacciyar sarrafa na'urar
Dace da TensorFlow Lite kuma ONNX

4. Google - Edge TPU

The Farashin TPU, wani bangare na Google's Dandalin murjani, an keɓe don IoT da ƙaddamarwa:

An inganta don TensorFlow Lite
Ƙarfafa ƙarancin ikon AI nuni
Karamin nau'i na nau'i don na'urori masu auna firikwensin kuma ƙofofin shiga

5. NVIDIA - NVDLA da Jetson Platform

NVIDIA ta NVDLA (Deep Learning Accelerator) da Jetson jerin isar da babban aikin AI kwamfuta:

Amfani a robotics, jirage marasa matuka, kuma injuna masu cin gashin kansu
Yana goyan bayan DABAN, TensorRT, kuma DeepStream SDK

Kalubale da iyakancewar NPUs

Yayin Rukunin sarrafa Jijiya (NPUs) bayar da ban mamaki abũbuwan amfãni a Ayyukan inference AI, karfin wutar lantarki, kuma gefen kwamfuta, ba su da iyaka. Masu haɓakawa da masu haɗa tsarin dole ne suyi la'akari da ƙalubalen fasaha da aiki da yawa yayin ɗaukar NPUs a ciki Na'urori masu kunna AI.

1. Daidaituwar Tsarin tsari

Ɗaya daga cikin mabuɗin shinge shine iyakataccen tallafi don AI frameworks. Ba kamar GPUs ba, waɗanda ke goyan bayan faɗuwar dandamali kamar TensorFlow, PyTorch, kuma ONNX, NPUs sau da yawa suna buƙata kayan aiki na al'ada ko SDKs na mallaka. Wannan na iya haifar da:

Dogayen hawan ci gaba
Wahala a ciki jujjuya samfuri da ingantawa
Ayyukan da ba su dace ba a tsakanin masu siyar da kayan aiki

2. Ƙarfin Ƙarfi-Kawai

Yawancin NPUs an tsara su na musamman don ra'ayi, ba horo. Nufin wannan:

Dole ne a horar da ƙirar AI akan GPUs ko TPUs
Duk wani sabuntawa ko gyarawa yana buƙatar kayan aikin waje
Rage sassauci a ilmantarwa akan na'urar

3. Matsalolin Haɗin Hardware

Haɗa NPUs cikin tsarin da aka haɗa ko Tsarin-on-Chip (SoC) ƙira ya haɗa da ciniki-offs:

Iyakokin yankin Silicon
Ƙuntataccen ƙirar thermal
Rikici mai yiwuwa tare da I/O na yanzu da bandwidth na ƙwaƙwalwar ajiya

Makomar NPUs da AI Hardware

Kamar yadda ake bukata ainihin lokacin AI inference ya ci gaba da girma a fadin masana'antu, makomar gaba Rukunin sarrafa Jijiya (NPUs) ana sa ran zama tsakiya ga juyin halitta na AI hardware. Waɗannan ƙwararrun na'urori masu sarrafawa na musamman suna ƙara ƙarfi, ƙanƙanta, da ingantaccen kuzari- sanya su a matsayin ɓangarorin ɓangarorin na gaba na gaba. gefen AI na'urorin, na'urori masu auna firikwensin, kuma m tsarin.

Abubuwan da ke tasowa a cikin Ci gaban NPU

Haɗin SoC mai tsauri
NPUs suna ƙara haɗawa cikin Tsarin-on-Chip (SoC) ƙira tare da CPUs, GPUs, da ISPs. Wannan haɗin gine-gine yana ba da damar sauri bututun sarrafa bayanai, yana rage jinkiri, kuma yana rage wahalar tsarin.
Taimako don Kwamfuta iri-iri
NPUs na gaba za su yi aiki tare da sauran masu haɓaka AI, suna rarraba nauyin aiki a duk faɗin CPUs, GPUs, da TPUs don mafi kyawun aiki.
Ingantattun Daidaituwar Tsarin Mulki
Masu siyarwa suna gina goyan baya ga tsarin AI na yau da kullun kamar ONNX, TensorFlow Lite, kuma PyTorch Mobile, ragewa lankwasa koyo da kuma ƙarfafa ƙaddamar da dandamali.
Fadada Bayan Na'urorin Mabukaci
NPUs za su taka rawar gani a ciki masana'antu sarrafa kansa, lafiya mai hankali, gefen robotics, noma AI, kuma tsarin sa ido, ku low-power, high-gudu inference yana da mahimmanci.

A cikin duniyar da take tafiya hankali a ko'ina, NPUs ba za su hanzarta AI kawai ba amma har ma su kara sa shi m, mai dorewa, kuma amintacce. Matsayin su wajen tsarawa na gaba-gen saka tsarin, daga kyamarori masu ƙarfin AI ku dandamali kewayawa masu cin gashin kansu, ya jaddada muhimmancin girma na ƙayyadaddun gine-ginen ƙididdiga na yanki a zamanin AI.