IGoogle ithi i-AI yayo iyashesha ekwakhiweni kwe-chip

I-Google ithi ithuthukile isoftware ye- ubuhlakani bokufakelwa obukwazi ukuklama ama-chip chip ngokushesha kunabantu. Ku-athikili eshicilelwe ezinsukwini ezimbalwa ezedlule, abakwaGoogle bathi i-chip engathatha abantu izinyanga ukuyiklama ingacatshangwa yi-AI yayo entsha ngaphansi kwamahora ayisithupha.

Ukuhlakanipha okungekhona okwangempela isivele isetshenziselwe ukuthuthukisa ukubuyiselwa kwamuva kwama-chips Iyunithi Yokucubungula Tensioner (TPU) yi-Google, ezisetshenziselwa ukwenza imisebenzi ehlobene nobuhlakani bokufakelwa, kusho iGoogle. Onjiniyela bakwaGoogle bathi ukuthuthuka kungaba "nomthelela omkhulu" embonini yama-semiconductor.

Empeleni, kumayelana nokuthola ukuthi izinto ezinjenge-CPU ne-GPU cores nememori zibekwa kanjani zodwa ku-chip. Ukutholakala kwawo kula mabhodi amancane kubalulekile ngoba kuthinta ukusetshenziswa kwamandla nokusheshisa kokusebenza kwe-chip; umzila wiring nowesiginali odingekayo ukuxhuma yonke into kubaluleke kakhulu.

Onjiniyela bakwaGoogle u-Azalia Mirhoseini no-Anna Goldie, kanye nozakwabo, bachaza encwadini yabo uhlelo lokufunda oluqinisa kakhulu olukwazi ukudala "amaphethini ayisisekelo" ngaphansi kwamahora ayisithupha, kanti kwesinye isikhathi kuthatha izinyanga.

Ngamanye amagama, I-Google isebenzisa ubuhlakani bokuzenzela ukuklama ama-chip angasetshenziswa ukudala amasistimu obuhlakani bokuzenzela ayinkimbinkimbi ngokwengeziwe.

Amasistimu afanayo nawo angashaya abantu emidlalweni eyinkimbinkimbi efana ne-go and chess. Kulezi zimo, ama-algorithms aqeqeshelwa ukuhambisa izingcezu ezandisa amathuba akho wokuwina umdlalo, kepha esimweni se-tile, i-AI iqeqeshelwe ukuthola inhlanganisela engcono kakhulu yezakhi ukuyenza iphumelele ngangokunokwenzeka emdlalweni.

Inethiwekhi ye-neural nayo isebenzisa amasu athile lezo ezake zacatshangelwa imboni yama-semiconductor, kepha zashiywa njengezinjongo zokufa. Ngokusho kwendatshana, uhlelo lwezobunhloli lokufakelwa luthole amapulani ayi-10.000 wamachips ukuze "afunde" okusebenzayo nokungasebenzi.

"Indlela yethu isetshenziselwe ukuklama isizukulwane esilandelayo se-Google AI accelerators futhi inamandla okusindisa izinkulungwane zamahora omzamo womuntu esizukulwaneni ngasinye esisha," kubhala onjiniyela. "Ekugcineni, sikholelwa ukuthi i-hardware enamandla kakhulu eyenzelwe i-AI izoshayela ukuthuthuka kwe-AI, idale ubudlelwano bokubambisana phakathi kwale mikhakha yomibili."

Ngokwe-athikili, lapho kuklanywa i-microprocessor noma i-accelerator yokulayisha umsebenzi, imvamisa kuyadingeka ukuchaza ukuthi ama-subsystems awo asebenza kanjani ngolimi oluphakeme, njenge-VHDL, iSystemVerilog, noma mhlawumbe neChisel.

Le khodi ekugcineni izoguqulela kulokho okubizwa ngokuthi yi-netlist, okuchaza ukuthi isethi yama-macroblocks namaseli ajwayelekile kufanele axhumeke kanjani ngezintambo ukwenza imisebenzi ye-chip.

Amaseli ajwayelekile aqukethe izinto eziyisisekelo ezifana namasango we-NAND ne-NOR logic, kanti ama-macroblocks aqukethe isethi yamaseli ajwayelekile noma ezinye izinto ze-elekthronikhi ezihloselwe ukwenza umsebenzi okhethekile, njengokuhlinzeka ngememori ye-chip noma ingqikithi yeprosesa. Ngakho-ke, ama-macroblocks makhulu kakhulu kunamaseli ajwayelekile.

Ngemuva kwalokho kufanele ukhethe ukuthi uluhlela kanjani lolu hlu lwamaseli nama-macroblocks ku-chip. Ngokusho kwabasebenzi bakwaGoogle, kungathatha onjiniyela babantu amasonto noma izinyanga ukusebenzisana namathuluzi akhethekile we-chip design futhi kuphindwe kaningi ukuthola uhlelo olwenziwe ngokuya ngezidingo zokusetshenziswa kwamandla, isikhathi, isivinini, njll.

Okuvame ukwenzeka kule nqubo ukuthi indawo yama-macroblocks amakhulu kufanele iguqulwe lapho ukwakhiwa kukhula. Futhi-ke kufanele uvumele amathuluzi azenzakalelayo, asebenzisa ama-algorithms angenangqondo, wehle ngobuningi bamaseli amancane ajwayelekile, bese uhlanza uphinde uphinde uze uqede, kusho i-doc.

Ukusheshisa lesi sinyathelo sokuklama kwe-chip, ochwepheshe bezobuchwepheshe bokuzenzela bakaGoogle benze uhlelo lwe-convolutional neural network olwenza ukubekwa kwe-macro-block lodwa emahoreni ambalwa ukufeza ukwakheka okuhle.

Amaseli ajwayelekile abekwa ngokuzenzakalela ezindaweni ezingenalutho ngenye isoftware, ngokusho kwendatshana. Lolu hlelo lokufunda ngomshini kufanele sikwazi ukukhiqiza umdwebo ofanele ngokushesha okukhulu nangcono kunendlela yonjiniyela babantu besebenzisa amathuluzi wendabuko azenzakalelayo embonini, abasebenzi bakwaGoogle bachaze ku-athikili yabo.

Umthombo: https://www.theregister.com/


Yiba ngowokuqala ukuphawula

Shiya umbono wakho

Ikheli lakho le ngeke ishicilelwe. Ezidingekayo ibhalwe nge *

*

*

  1. Ubhekele imininingwane: Miguel Ángel Gatón
  2. Inhloso yedatha: Lawula Ugaxekile, ukuphathwa kwamazwana.
  3. Ukusemthethweni: Imvume yakho
  4. Ukuxhumana kwemininingwane: Imininingwane ngeke idluliselwe kubantu besithathu ngaphandle kwesibopho esisemthethweni.
  5. Isitoreji sedatha: Idatabase ebanjwe yi-Occentus Networks (EU)
  6. Amalungelo: Nganoma yisiphi isikhathi ungakhawulela, uthole futhi ususe imininingwane yakho.