Uma ukhuluma ngamaphrosesa akhethekile, ungase uzwe i-NPU, TPU, DSP, ne-VPU. Ngamunye unomsebenzi wakhe. I-NPU isiza ukusheshisa imisebenzi ye-AI, njengasemafonini. I-TPU, eyenziwe i-Google, isebenza ngezibalo zokufunda ngomshini. I-DSP iphatha amasignali esikhathi sangempela, njengomsindo noma ividiyo, ngokushesha okukhulu. I-VPU igxile emisebenzini yokubona, isiza amadivayisi aqonde izithombe namavidiyo. Lawa maphrosesa awafani nama-GPU avamile. Zenzelwe imisebenzi ethile, zisebenzisa amandla amancane futhi zisebenza kangcono emisebenzini yazo.
Izitoreji Eziyinhloko
Ama-NPU akhelwe Imisebenzi ye-AI, okubenza basheshe futhi babe ngcono kunamaphrosesa avamile. Basebenza kahle ezintweni ezifana nokuqashelwa ubuso nezimoto ezizishayelayo.
Ama-TPU, enziwe i-Google, asebenza kahle kuzibalo ze-tensor ukufunda okujulile. Basiza ukuqeqesha amamodeli amakhulu e-AI ngokushesha futhi basebenze kahle emafini.
Ama-DSP aphatha amasiginali wesikhathi sangempela njengomsindo nevidiyo. Zilungele imisebenzi edinga isivinini nokunemba, njengokunciphisa umsindo kumasistimu omsindo.
Ama-VPU enzelwe imisebenzi yokubona ngekhompyutha, esiza amadivaysi ukuphatha izithombe namavidiyo ngokushesha. Zibalulekile ezintweni ezifana nokuqashelwa kobuso kanye neqiniso elingeziwe.
Ukukhetha iphrosesa elungile kuncike kulokho okudingayo: sebenzisa ama-NPU noma ama-TPU e-AI, ama-DSP amasiginali, nama-VPU ngemisebenzi yokubona ukuze uthole imiphumela engcono kakhulu.
I-Neural Processing Unit (NPU)
Yini i-NPU?
I-NPU, noma i-Neural Processing Unit, iyi-chip ekhethekile. Yenzelwe ukuphatha imisebenzi ye-artificial intelligence (AI). Ngokungafani nama-processor ajwayelekile, ama-NPU ayashesha ngomsebenzi we-AI. Basiza ngokuqeqesha amamodeli e-AI nokusebenzisa amanethiwekhi e-neural. Ama-NPU angcono ezibalweni ezidingekayo ku-AI, okuwenza asheshe futhi asebenze kahle kunama-GPU.
Isebenza kanjani i-NPU?
Ama-NPU asebenzisa ihadiwe ekhethekile ukucubungula imisebenzi ye-AI ngokushesha. Bangenza imisebenzi eminingi yenethiwekhi ye-neural ngesikhathi esisodwa. Lokhu kubenza baziqeqeshe futhi basebenzise amamodeli amakhulu e-AI ngokushesha okukhulu.
Ukuze wonge amandla, ama-NPU asebenzisa izibalo ezilula, njengezinombolo ezingu-8-bit noma ezincane. Lokhu kugcina imisebenzi inemba kodwa ibe nzima kangako. Futhi banenkumbulo esheshayo yokuphatha idatha enkulu ngaphandle kokwehlisa ijubane. Amathuluzi athuthukile, njengama-systolic arrays, awenza abe ngcono nakakhulu kumisebenzi ye-AI.
Ama-NPU ashesha izikhathi ezingaphezu kwe-100 kunama-GPU kodwa asebenzisa amandla afanayo. Lokhu kubenza balungele umsebenzi we-AI.
Sebenzisa izimo zama-NPU
Ama-NPU asetshenziswa kumadivayisi adinga i-AI. Amafoni awasebenzisela ukuvula ngobuso, izisizi zezwi, nokuhlela isithombe. Izimoto ezizishayelayo zisebenzisa ama-NPU ukwenza izinqumo ezisheshayo kusuka kudatha yezinzwa. Kwezokunakekelwa kwempilo, ama-NPU asiza ukutadisha izithombe zezokwelapha futhi abikezele impilo yesiguli.
Zibalulekile futhi ekuqeqesheni amamodeli amakhulu e-AI ezimbonini ezifana nezamabhange, ukuthenga nokuzijabulisa. Noma ngabe okokusebenzisa amanethiwekhi e-neural noma ukusheshisa i-AI, ama-NPU ayashesha futhi asebenza kahle emisebenzini enzima.
Isici | Incazelo |
|---|---|
Ukucubungula okufanayo | Ama-NPU angenza imisebenzi eminingi yenethiwekhi ye-neural ngesikhathi esisodwa. |
Izibalo ezinemba ephansi | Isebenzisa izibalo ezilula, njengezinombolo ze-8-bit, ukuze wonge amandla. |
Inkumbulo yomkhawulokudonsa ophezulu | Inememori esheshayo yokusingatha idatha enkulu yemisebenzi ye-AI. |
Ukusheshiswa kwekhwalithi | Isebenzisa amathuluzi athuthukile, njengama-systolic arrays, ekusebenzeni okungcono kwe-AI. |
Iyunithi Yokucubungula I-Tensor (TPU)
Yini i-TPU?
I-TPU, noma i-Tensor Processing Unit, iyi-chip ekhethekile. Yenzelwe ukuphatha imisebenzi enzima ye-AI. I-Google yakhe ama-TPU ukwenza ukufunda okujulile kusheshe. Zinhle ngezibalo ze-tensor, okuwukhiye ku-AI. Ngokungafani namaphrosesa ajwayelekile, ama-TPU angcono ekuqeqesheni nasekusebenziseni amamodeli e-AI. Zakhelwe ukwenza izibalo ezinkulu ngokushesha. Lokhu kubenza bafanelekele imisebenzi ye-AI edinga isivinini nokusebenzisa amandla aphansi.
Ama-TPU asebenza kahle kakhulu ngama-tensors. Ama-tensors angamathuluzi ezibalo asetshenziswa ekufundeni okujulile. Zilungele imisebenzi efana nokuthola izinto ezithombeni noma ukuqonda ulimi. Ama-TPU ayashesha kunama-GPU ngoba agxila kuzibalo ze-tensor.
Isebenza kanjani i-TPU?
Ama-TPU asebenzisa ihadiwe ekhethekile ukwenza izibalo ze-tensor. Basebenzisa ukuphindaphinda kwe-matrix ukuphatha idatha enkulu ngokushesha. Ama-TPU angenza imisebenzi eminingi ngesikhathi esisodwa, awenze asheshe kunama-GPU. Banenkumbulo esheshayo namathuluzi angokwezifiso ukusiza ngomsebenzi we-AI.
Nasi ukuqhathanisa ukusebenza kwe-TPU ne-GPU:
I-Metric | Ukusebenza kwe-TPU | Ukusebenza kwe-GPU |
|---|---|---|
Isivinini Sokuqeqeshwa Kwenethiwekhi Yezinzwa | 15–30x ngokushesha | Isisekelo (GPU) |
Ukusebenza nge-Watt ngayinye | 25–50x kangcono | Isisekelo (GPU) |
I-TPU v3 vs NVIDIA V100 | 8x ngokushesha kwe-BERT | Isisekelo (NVIDIA V100) |
I-TPU v4 vs NVIDIA A100 | 1.2–1.7x kangcono | Isisekelo (NVIDIA A100) |
Leli thebula libonisa ukuthi ama-TPU ashesha kanjani futhi asebenzisa amandla amancane. Zilungele imisebenzi ye-AI edinga isivinini esikhulu.
Sebenzisa izimo zama-TPU
Ama-TPU asetshenziselwa imisebenzi edinga umsebenzi we-AI osheshayo. Zivamile kumapulatifomu amafu ukuqeqesha amamodeli amakhulu e-AI. I-Google isebenzisa ama-TPU ngamathuluzi ayo e-AI, njengenkulumo nokuhumusha.
Ama-TPU futhi asiza amamodeli aqeqeshiwe ukwenza izibikezelo. Kwezokunakekelwa kwempilo, bafunda izithombe zezokwelapha. Izitolo eziku-inthanethi zizisebenzisa ukuphakamisa imikhiqizo. Ama-TPU mahle ngezibalo ze-tensor, okuwenza abe ukhiye we-AI yesimanje.
Isici/Metric | Incazelo |
|---|---|
Ingxenye Eqinile (HS) / Ingxenye Ethambile (SS) | Isilinganiso sishintsha amandla e-TPU nokuvumelana nezimo. |
Amandla Amandla | Ama-TPU asuselwa ku-Bio asukela ku-30.69 kuya ku-11.07 MPa. |
Ukunwebeka ngesikhathi sekhefu | Ama-TPU asuselwa ku-Bio asuka ku-977% aya ku-489%. |
Ukwehla Kwezinkinga | Ukuqina kwe-TPU kusuka ku-Shore 60A kuya ku-85A. |
Ama-TPU ashintsha izimboni ngokwenza i-AI isheshe futhi ibe ngcono. Noma ngabe ukuqeqeshwa noma ukusebenzisa amamodeli e-AI, ama-TPU anikeza isivinini namandla adingekayo emisebenzini enzima.
I-Digital Signal Processor (DSP)
Iyini i-DSP?
I-DSP, noma iphrosesa yesiginali yedijithali, iyi-chip ekhethekile. Icubungula amasignali njengomsindo, ividiyo, nedatha yenzwa. Lawa masignali aguqulwa abe amafomethi edijithali ukuze asetshenziswe kalula. Ngokungafani namaphrosesa avamile, ama-DSP enzelwe imisebenzi yezibalo esheshayo, ephindaphindiwe. Lokhu kubenza balungele imisebenzi yesikhathi sangempela njengokuthuthukisa umsindo, ukucindezela amavidiyo, noma ukusiza ngokuxhumana.
Ama-DSP akhelwe imisebenzi ethile. Baphatha izinto ezinjengokuhlunga kanye nezinguquko zesignali ngokunemba. Isivinini sabo nokunemba kwabo kubenza babaluleke ekunakekelweni kwezempilo, ezimotweni nakuma-electronics.
Isebenza kanjani i-DSP?
I-DSP ishintsha amasignali edijithali ukuze ithole ulwazi oluwusizo. Isebenzisa izibalo ezisheshayo ukuze iguqule noma ihlaziye amasignali. Iziqondiso ezikhethekile nesivinini esiphezulu zisiza ukuthi isebenze ngokushesha. Ama-DSP angakwazi ukuphatha izinombolo eziningi ngezibalo zamaphuzu antantayo ukuze kube nokunemba okungcono.
Ukuthi i-DSP isebenza ngokushesha kangakanani kuncike ezintweni ezintathu: imisebenzi yezibalo, imisebenzi yokulawula, nedatha ehambayo. Ukuhlela okuhlakaniphile kunciphisa ukubambezeleka, kwenza ama-DSP asheshe. Baphinde basebenze namanye ama-chips ukwenza imisebenzi eyengeziwe, ukugcina ukucutshungulwa kwesignali kubushelelezi.
Nali ithebula elibonisa izici ezibalulekile ze-DSP:
Uhlobo lokucaciswa | Incazelo |
|---|---|
Icubungula Amandla | Ama-DSP asezingeni eliphezulu anamandla aqinile anama-cores amaningi. |
Imiyalo Yokusetha | Iziqondiso ezikhethekile zisiza ngezinkinga zezibalo ezinzima. |
isivinini sewashi | Ijubane lewashi elisheshayo livumela izibalo ezisheshayo. |
Amakhono Ephuzu Elintantayo | Ama-DSP asendaweni entantayo asebenzisa amabhithi angu-32 ukuze anembe kangcono. |
Ama-IC ashumekiwe | Ama-chips engeziwe athuthukisa inkumbulo futhi axhumeke kwamanye amadivaysi. |
Sebenzisa izimo ze-DSPs
Ama-DSP asetshenziswa ezindaweni eziningi ezidinga umsebenzi wesignali osheshayo. Ezinhlelweni zomsindo, zenza umsindo ucace futhi zinciphise umsindo. Kumavidiyo, asiza ngokuminyanisa nezithombe eziqinile. Amafoni namanethiwekhi asebenzisa ama-DSP ukuthumela nokwamukela amasiginali ngokushelela.
Ama-DSP aphinde asize nge-AI ngokucubungula idatha yezinzwa kumadivayisi ahlakaniphile. Ekunakekelweni kwezempilo, bafunda amasignali afana nokushaya kwenhliziyo noma amaskena ukuze bathole imiphumela engcono. Ikhono labo lokusingatha imisebenzi yesignali enzima libenza babe usizo ezimbonini eziningi.
I-Vision Processing Unit (VPU)
Iyini i-VPU?
I-Vision Processing Unit (VPU) iyi-chip ekhethekile. Iphatha imisebenzi ehlobene nombono wekhompyutha. Ama-VPU acubungula izithombe namavidiyo ngokushesha nangempumelelo. Ngokungafani namaphrosesa avamile, enzelwe imisebenzi yombono wesikhathi sangempela. Zilungele ukutholwa kwento, ukubonwa kobuso, nokuhlaziya ividiyo. Ama-VPU asusa le misebenzi kusiprosesa esikhulu. Lokhu kwenza amadivayisi asheshe futhi konga amandla.
Isebenza kanjani i-VPU?
Ama-VPU asebenzisa ihadiwe ekhethekile ukucubungula idatha ebonakalayo. Bangakwazi ukuphatha izithombe eziningi noma ukusakazwa kwamavidiyo ngesikhathi esisodwa. Lokhu kubasiza ukuthi benze imisebenzi eyinkimbinkimbi efana ne-AI inference ngokushesha. Ama-VPU aphinde asebenzise ama-algorithms ahlakaniphile ukuze afunde idatha ebonakalayo. Lokhu kubenza bafanelekele imisebenzi ye-AI yesikhathi sangempela.
Nazi izindlela ezintathu ezibalulekile zokulinganisa ukusebenza kwe-VPU:
Imethrikhi yokusebenza | Incazelo | Ukubaluleka |
|---|---|---|
Speed | Ukuthi imisebenzi iqedwa ngokushesha kangakanani | Ibonisa ukuthi i-VPU isebenza kahle kangakanani |
Isikhathi sempendulo | Isabela ngokushesha kangakanani kokufakwayo | Ibalulekile kulwazi lomsebenzisi |
Ukusebenzisa | Ingakanani idatha engayiphatha | Ibonisa ukuthi isebenza kahle kangakanani ngaphansi kokusetshenziswa kanzima |
Lawa mamethrikhi abonisa ukuthi ama-VPU ashesha futhi athembekile kanjani. Zilungele imisebenzi edinga imiphumela yesikhathi sangempela.
Sebenzisa amakesi ama-VPU
Ama-VPU asetshenziswa ezindaweni ezidinga ukucutshungulwa kwedatha ebonakalayo. Kumakhamera ahlakaniphile, asiza ukubona izinto noma ubuso ngokushesha. Ezimotweni ezizishayelayo, ama-VPU acubungula ividiyo ukuze agweme izithiyo futhi ashayele ngokuphepha. Ama-drones awasebenzisa ukwenza imephu nokubuka izindawo.
Kwezokunakekelwa kwempilo, ama-VPU afunda izithombe zezokwelapha ezifana nama-X-ray ukuze athole izinkinga. Kumagajethi, athuthukisa okwangempela okungathandwa kwabathelisi esikubona kanye nokusakaza ividiyo. Amandla abo okusebenzisa i-AI kumadivayisi abenza babe ukhiye wemisebenzi esheshayo nephumelelayo.
Umehluko Obalulekile Phakathi kwe-NPU, i-TPU, i-DSP, ne-VPU
Isifinyezo sokuqhathanisa sokusebenza nokusebenza
Iphrosesa ngayinye inamandla akhethekile emisebenzi ethile. Ukwazi lokhu kusiza ukukhetha okulungile:
I-NPU: Yenzelwe imisebenzi ye-AI, ama-NPU enza imisebenzi eminingi ngesikhathi esisodwa. Zinhle kakhulu ekufundeni okujulile, njengokuqeqeshwa namamodeli asebenzayo. I-TOPS (Tera Operations Per Second) ikhombisa ukuthi bayiphatha kahle kangakanani imisebenzi enzima. Isibonelo, i-NPU enamayunithi angu-1000 MAC ku-1 GHz ingafinyelela i-TOPS ephezulu, ibonise amandla ayo.
I-TPU: Ama-TPU akhelwe imisebenzi emikhulu yokufunda ngomshini. Ziyashesha kakhulu futhi zonga amandla. Lawa ma-chips angcono kakhulu ekuqeqesheni amamodeli e-AI futhi uwasebenzise emafini.
DSP: Ama-DSP asebenza kumasiginali wesikhathi sangempela njengomsindo, ividiyo, noma izinzwa. Ahamba kancane kunama-NPU noma ama-TPU e-AI kodwa alungile emisebenzini yezibalo ephindaphindiwe.
I-VPU: Ama-VPU enzelwe imisebenzi yokubona. Baphatha kahle izithombe namavidiyo. Lokhu kubenza bafanelekele izinto ezifana nokubonwa kobuso, ukutholwa kwento, kanye neqiniso eli-augmented. Ama-VPU ayashesha futhi onga amandla emisebenzi ebonakalayo.
Ukukhetha iphrosesa efanele yezidingo zakho
Khetha iphrosesa efanele ngokusekelwe kulokho okudingayo:
Okwemisebenzi ye-AI: Sebenzisa i-NPU noma i-TPU ukuze ufunde ngokujulile. Ama-NPU mahle kumafoni, kuyilapho ama-TPU asebenza kahle kakhulu emafini.
Ngomsebenzi wesignali: Khetha i-DSP yomsindo, ividiyo, noma imisebenzi yezinzwa. Ama-DSP alungele imisebenzi yesikhathi sangempela edinga ukunemba.
Ngemisebenzi yombono: Khetha i-VPU yemisebenzi efana nombono wekhompyutha noma umsebenzi wevidiyo. Ama-VPU alungele ama-drones noma amakhamera ahlakaniphile.
Ukuze ube nezimo: Sebenzisa i-FPGA uma udinga iphrosesa yemisebenzi eminingi. Ama-FPGA angakhelwa imisebenzi eyahlukene, awenze abe wusizo kakhulu.
Ngokwazi ukuthi iphrosesa ngayinye yenza ini kangcono kakhulu, ungakhetha ngokuhlakanipha futhi uthole imiphumela engcono.
Ukuqonda umehluko phakathi kwe-NPU, TPU, DSP, ne-VPU kukusiza ukuthi ukhethe ithuluzi elifanele lemisebenzi yakho. Ama-NPU ahamba phambili emisebenzini ye-AI efana nokuqashelwa kobuso. Ama-TPU aphatha umshini wokufunda ngezinga elikhulu ngempumelelo. Ama-DSP acubungula amasiginali esikhathi sangempela anjengomsindo nevidiyo. Ama-VPU asebenza ngokukhethekile emisebenzini yokubona ikhompuyutha njengokutholwa kwento.
Ithiphu: Uma usebenza namamodeli e-AI, khetha i-NPU noma i-TPU. Ngokucubungula umsindo noma ividiyo, hamba ne-DSP. Ngemisebenzi ehlobene nombono, i-VPU iyisinqumo sakho esingcono kakhulu. Ukufanisa iphrosesa nezidingo zakho kuqinisekisa ukusebenza okungcono nokusebenza kahle.
Imibuzo Evame Ukubuzwa
Ahluke kanjani ama-NPU nama-TPU?
Ama-NPU asebenza emisebenzini ye-AI kumagajethi afana namafoni noma izimoto. Ama-TPU, enziwe i-Google, aphatha imisebenzi emikhulu yokufunda ngomshini, ikakhulukazi emafini. Ama-NPU angcono kakhulu kumadivayisi amancane, kuyilapho ama-TPU elungele ukusebenzisa ikhompyutha enamandla.
Ingabe ama-DSP angayenza imisebenzi ye-AI?
Yebo, kodwa ama-DSP awakhelwe i-AI. Baphatha amasignali esikhathi sangempela njengomsindo noma ividiyo. Ku-AI, ama-NPU noma ama-TPU angcono ngoba acubungula amanethiwekhi we-neural ngokushesha.
Kungani i-VPU ibalulekile ekuboneni ngekhompyutha?
Ama-VPU acubungula izithombe namavidiyo ngokushesha. Zilungele imisebenzi efana nokuthola izinto, ukubona ubuso, kanye neqiniso eli-augmented. Ijubane labo libenza bafanelekele imisebenzi yesikhathi sangempela kuma-drones, amakhamera, nokunakekelwa kwezempilo.
Iyiphi iprosesa engcono kakhulu yokuqeqesha amamodeli e-AI?
Ama-TPU angcono kakhulu ekuqeqesheni amamodeli e-AI. Zenzelwe imisebenzi emikhulu yokufunda komshini kanye nezibalo ze-tensor. Ama-TPU ayashesha futhi onga amandla, okuwenza alungele ukufunda okujulile.
Ingabe lawa maphrosesa angasebenza nama-CPU?
Yebo, bangasebenza nama-CPU. Lawa maphrosesa aphatha imisebenzi ekhethekile njenge-AI, amasiginali, noma umbono. Lokhu kuvumela i-CPU ukuthi igxile kweminye imisebenzi, ithuthukise ukusebenza kwesistimu.
