Tsarin TREE: Ingantaccen Amfani da Makamashi Mai Amfani da Token don Cibiyoyin Sadarwa na 6G Masu Haɗa AI

1. Gabatarwa & Bayyani

Haɗaɗɗun Fasahar Hankali (AI) cikin cibiyoyin sadarwa na zamani na shida (6G) yana wakiltar sauyi zuwa ga hankali ko'ina da haɗin kai mai zurfi. Kamar yadda aka zayyana a cikin hangen nesa na IMT-2030, 6G yana nufin tallafawa aikace-aikacen da ke buƙatar babban bandeji kamar ƙarin gaskiya, tsarin kai da kai, da manyan ayyukan IoT, tare da AI a matsayin mai ba da dama. Duk da haka, wannan haɗuwa yana gabatar da ƙalubale mai mahimmanci: ma'auni na yau da kullun na ingantaccen amfani da makamashi (EE), wanda aka saba bayyana shi azaman ƙimar aiki na cibiyar sadarwa a kowace raka'a makamashi ($EE = \frac{Ƙimar Aiki}{Makamashi}$), ya kasa ɗaukar amfani da ƙimar ayyukan AI na musamman, kamar waɗanda Manyan Samfuran Harshe (LLMs) ke yi. Wannan takarda ta gabatar da Tsarin Ingantaccen Amfani da Makamashi Mai Amfani da Token (TREE), sabon ma'auni da aka ƙera don rufe wannan gibi ta hanyar haɗa ƙimar aiki na manyan samfuran AI cikin lissafin amfanin tsarin, don haka yana ba da madaidaicin ma'auni na dorewar makamashi don cibiyoyin sadarwa na 6G masu haɗa AI.

2. Tsarin TREE

Tsarin TREE ya sake bayyana ingantaccen amfani da makamashi don zamanin AI. Ya wuce kawai bayanan bit zuwa la'akari da "tokens" na lissafi da samfuran AI ke sarrafa su a matsayin manyan masu ɗaukar amfani a cikin cibiyar sadarwa mai hankali.

2.1 Ma'anar Ma'auni na Asali

Ma'auni na asali na TREE an tsara shi azaman rabo na amfanin aikin AI mai inganci (wanda aka auna a cikin tokens) zuwa jimillar amfani da makamashi na tsarin. Ya yarda cewa ba duk zirga-zirgar cibiyar sadarwa ke ɗaukar ƙima daidai ba; sarrafa tokens don sabis na fassara harshe na ainihin lokaci yana da amfani da tasirin makamashi daban-daban da bayanan bidiyo mai yawo.

2.2 Ka'idojin Ƙira

Tsarin yana nazarin ƙirar cibiyar sadarwa ta hanyar la'akari da abubuwa uku masu mahimmanci na AI:

• Ƙarfin Lissafi: Rarraba albarkatun lissafi a cikin girgije, gefe, da na'urori na ƙarshe.
• Samfuran AI: Tsarin, girma, da ingancin samfuran da aka tura (misali, LLMs, samfuran gani).
• Bayanan: Ƙarar, nau'in, da kwararar bayanan da ake buƙata don horar da AI da fahimta.

Haɗin kai tsakanin waɗannan abubuwan yana ƙayyadaddun jimillar TREE na tsarin.

3. Binciken Fasaha

3.1 Tsarin Lissafi

Ma'aunin TREE da aka gabatar ana iya bayyana shi kamar haka: $$\text{TREE} = \frac{\sum_{i \in \mathcal{A}} w_i \cdot U_i(T_i) + \sum_{j \in \mathcal{D}} w_j \cdot R_j}{P_{\text{total}}}$$ Ina:

• $\mathcal{A}$ shine saitin ayyukan AI kuma $\mathcal{D}$ shine saitin ayyukan bayanai na al'ada.
• $U_i(T_i)$ shine aikin amfani don sabis na AI $i$, ya dogara da ƙimar aikin sa na token $T_i$.
• $R_j$ shine ƙimar bayanai don sabis na al'ada $j$.
• $w_i, w_j$ sune abubuwan da ke nuna fifikon sabis.
• $P_{\text{total}}$ shine jimillar amfani da wutar lantarki na tsarin.

Wannan tsari ya haɗa amfanin aikin AI a fili, yana wucewa daga tsarin bit-per-joule na al'ada.

3.2 Tsarin Tsarin Tsarin

An ƙera TREE don tsarin gine-ginen girgije-ƙarshe. Muhimman abubuwan da aka yi la'akari sun haɗa da:

• Rarraba & Juyar da Samfura: Rarraba aiwatar da samfurin AI a tsakanin gefe da girgije bisa ga ƙuntatawa na makamashi da jinkiri don haɓaka TREE.
• Koyo na Tarayya: Ba da damar horar da AI mai rarrabawa yayin rage makamashin watsa bayanai, yana tasiri kai tsaye ga ma'auni na TREE.
• Matsawa Samfura Mai Daidaitawa: Amfani da dabaru kamar Low-Rank Adaptation (LoRA) don rage farashin makamashi na lissafi na gyara samfura a gefe.

4. Sakamakon Gwaji & Nazarin Lamura

Takardar ta gabatar da nazarin lamura don tabbatar da ikon TREE na musamman. A cikin yanayin zirga-zirgar haɗaɗɗiyar da ke haɗa ayyukan fahimtar AI (misali, nazarin bidiyo na ainihin lokaci) tare da kwararar bayanai na al'ada (misali, zazzage fayil), ma'auni na EE na al'ada sun kasance marasa isa. Sun kasa bayyana mahimman rashin daidaituwa na makamashi-sabisi—yanayin da ƙaramin adadin zirga-zirgar AI mai ƙima yana cinye makamashi maras daidaituwa idan aka kwatanta da babban ƙarar zirga-zirgar bayanai marasa ƙima. TREE ya yi nasara wajen ƙididdige wannan rashin daidaituwa, yana ba masu sarrafa cibiyoyin sadarwa cikakken hoto na inda ake kashe makamashi daidai da ƙimar da ake samarwa. Misali, wani yanayi na iya nuna cewa ba da 1000 tokens don mataimaki na tushen LLM yana cinye makamashi daidai da yawo 1GB na bidiyo, amma yana ba da amfani daban-daban, bambancin da TREE kawai zai iya ɗauka.

5. Misalin Tsarin Bincike

Yanayi: Tashar tushe ta 6G tana ba da sabis guda biyu a lokaci guda: (1) sabis na fahimtar LLM na tushen gefe don sarrafa tambayoyin birni mai hankali, da (2) loda bayanan firikwensin IoT na bango.

Matakan Binciken TREE:

1. Ayyana Amfanin: Sanya amfani $U_1 = \alpha \cdot T_1$ (tokens da aka sarrafa) don sabis na LLM da $U_2 = \beta \cdot R_2$ (bits da aka loda) don sabis na IoT. Ma'auni $\alpha > \beta$ yana nuna mafi girman ƙima a kowace raka'a na sabis na AI.
2. Auna Wutar Lantarki: Lura da jimillar wutar lantarki $P_{total}$ da lissafi (don LLM) da sadarwa (don duka biyun) suka cinye.
Yi Lissafi & Kwatanta: Lissafa TREE = $(\alpha T_1 + \beta R_2) / P_{total}$. Kwatanta wannan da EE na al'ada = $(R_1 + R_2)/P_{total}$. Binciken zai iya nuna cewa rabon ƙarin albarkatu ga sabis na LLM yana inganta TREE fiye da EE na al'ada, yana jagorantar tsarin albarkatu mai hankali.

Wannan tsarin yana ba masu sarrafa damar motsawa daga "rage makamashi a kowace bit" zuwa "haɓaka ƙima (tokens + bits) a kowace joule."

6. Bincike Mai Mahimmanci & Ra'ayoyin Ƙwararru

Bayanin Asali: Takardar TREE ba kawai tana gabatar da sabon ma'auni ba; tana ƙalubalantar lissafin tattalin arziki da injiniya na cibiyoyin sadarwa na gaba. Ta gano daidai cewa shawarar ƙimar 6G za ta kasance ta mamaye AI-a matsayin-Sabisi, ba kawai bututu masu sauri ba. Dogaro da inganci akan bit kamar auna ƙimar ɗakin karatu ta nauyin littattafansa—ya rasa batun gaba ɗaya. Sauyi zuwa tokens mataki ne mai mahimmanci, ko da yake sabo, zuwa ga cibiyar sadarwa mai sanin amfani.

Kwararar Hankali: Hujjar tana da inganci: 1) AI shine jigon ƙimar 6G. 2) Ƙimar AI tana cikin tokens/ayyuka, ba bit ba. 3) Don haka tsofaffin ma'auni (bits/Joule) sun tsufa. 4) Saboda haka, muna buƙatar sabon ma'auni (tokens/Joule). 5) Wannan sabon ma'auni (TREE) yana bayyana sabbin matsalolin ingantawa da ciniki. Hankali yana da ban sha'awa kuma yana magance babban makafi a cikin binciken 6G na yanzu, wanda sau da yawa yana ɗaukar AI a matsayin wani aiki kawai maimakon mai haɓaka ƙima.

Ƙarfi & Kurakurai: Babban ƙarfin shine hangen nesa na ra'ayi. Marubutan suna kallon bayan ƙalubalen fasaha na 6G na nan take zuwa ga dalilinsa na ƙarshe. Kurakuri, kamar kowane ma'auni na majagaba, shine ma'auni mai yiwuwa. Ta yaya za mu daidaita aikin amfani $U_i(T_i)$? Token don GPT-4 ba daidai yake da token don sauƙaƙan mai canza gani ba. Ayyana da yarda da waɗannan ma'auni na amfani a cikin masu siyarwa da ayyuka zai zama rikici na siyasa da fasaha, mai kama da ƙalubalen ƙididdige Ingancin Kwarewa (QoE). Bugu da ƙari, tsarin a halin yanzu yana dogaro sosai akan fahimta; babban farashin makamashi na horar da AI mai rarrabawa a cikin cibiyoyin sadarwa, damuwa da bincike kamar na Ƙungiyar Tasirin CO2 na Koyon Injiniya ya bayyana, yana buƙatar zurfafa haɗawa cikin lissafin TREE.

Bayanai Masu Aiki: Ga masu sarrafa cibiyoyin sadarwa da masu siyar da kayan aiki, abin da ake buƙata yana da gaggawa: fara kafa cibiyoyin sadarwarku da dandamalin AI don auna ƙimar aikin token kuma a haɗa shi da amfani da makamashi a matakin ƙananan. Ayyukan gwaji ya kamata su gwada algorithms na tsarawa masu jagorar TREE. Ga ƙungiyoyin ƙa'idodi (3GPP, ITU), aikin ya kamata ya fara yanzu akan ayyana azuzuwan sabis na tushen token da bayanin amfani, kamar yadda aka ayyana azuzuwan QoS don 4G/5G. Yin watsi da wannan kuma tsayawa kan EE na al'ada hanya ce ta tabbatar da gina cibiyoyin sadarwa masu ingantaccen amfani da makamashi waɗanda ba su da inganci a tattalin arziki don zamanin AI.

7. Aikace-aikace na Gaba & Jagorori

Tsarin TREE yana buɗe hanyar don aikace-aikace da yawa masu ci gaba da jagororin bincike:

• Yankewa Cibiyar Sadarwa Mai Sauƙi: Ƙirƙirar yankunan cibiyar sadarwa da aka inganta AI tare da tabbataccen matakan TREE don manyan ayyukan AI, daban da yankunan bayanai mafi kyawun ƙoƙari.
• Kasuwannin AI na Kore: Ba da damar cinikin albarkatun lissafi da fahimta masu sanin makamashi a gefen cibiyar sadarwa, inda ayyuka ke yin tayin bisa buƙatunsu na amfani na tushen token.
• Haɗin Ƙira na Sadarwa da Lissafi: Haɗin ƙira na ka'idojin matakin zahiri, gine-ginen cibiyar sadarwa, da gine-ginen samfurin AI daga tushe don haɓaka TREE, wucewa daga tsarin yanzu na daidaita AI zuwa cibiyoyin sadarwa na yanzu.
• Ƙimar Tsawon Rayuwa: Tsawaita TREE don rufe cikakken tsarin rayuwar ayyukan AI a cikin cibiyar sadarwa, gami da farashin makamashi na horar da samfura, sabuntawa, da sarrafa bututun bayanai, haɗa ra'ayoyi daga binciken nazarin tsawon rayuwa.
• Daidaituwar Amfanin Token: Babban shiri na gaba shine haɓaka ƙa'idodin masana'antu gabaɗaya don daidaita "amfanin" ayyukan AI daban-daban, kamar yadda lambobin bidiyo ke ayyana ma'auni na inganci.

8. Nassoshi

1. ITU-R. “Tsarin da manufofi gabaɗaya na ci gaba na gaba na IMT don 2030 da bayansa.” ITU-R M.[IMT-2030.FRAMEWORK], 2023.
2. Zhou, Z., Chen, X., Li, E., Zeng, L., Luo, K., & Zhang, J. (2019). Hankali na gefe: Shinge mil na ƙarshe na fasahar hankali tare da lissafin gefe. Proceedings of the IEEE, 107(8), 1738-1762.
3. Hu, E. J., et al. (2021). LoRA: Ƙaramin Daraja Daidaitawa na Manyan Samfuran Harshe. arXiv preprint arXiv:2106.09685.
4. Lacoste, A., Luccioni, A., Schmidt, V., & Dandres, T. (2019) Ƙididdige Hayakin Carbon na Koyon Injiniya. arXiv preprint arXiv:1910.09700.
5. Wang, X., Han, Y., Leung, V. C., Niyato, D., Yan, X., & Chen, X. (2020). Haɗuwa na lissafin gefe da koyo mai zurfi: Cikakken bincike. IEEE Communications Surveys & Tutorials, 22(2), 869-904.
6. Zhu, J. Y., Park, T., Isola, P., & Efros, A. A. (2017). Fassarar hoto zuwa hoto mara haɗin gwiwa ta amfani da cibiyoyin adawa masu da'ira. A cikin Proceedings of the IEEE international conference on computer vision (shafi na 2223-2232). (An ambata a matsayin misalin aikin AI mai cike da lissafi wanda farashin makamashinsa a cikin mahallin cibiyar sadarwa zai fi dacewa da TREE ya kimanta).