Lokaci Mai Dorewa-Aware Neuromorphic Computing tare da NVM

Teburin Abubuwan Ciki

1. Gabatarwa

Lissafin Neuromorphic tare da ƙwaƙwalwar ajiya mara-sauƙi (NVM) yana wakiltar sauyin tsari a cikin kayan aikin koyon na'ura, yana ba da gagarumin ci gaba a cikin aiki da ingantaccen kuzari don lissafin tushen spike. Duk da haka, manyan ƙarfin lantarki da ake buƙata don sarrafa NVM kamar ƙwaƙwalwar ajiya mai canzawa (PCM) suna haɓaka tsufa a cikin da'irar neuron na CMOS, suna barazana ga dogon lokacin amincin kayan aikin neuromorphic.

Wannan aikin yana magance matsalar matuƙar tsawon rai a cikin tsarin neuromorphic, yana mai da hankali kan hanyoyin gazawa kamar rashin kwanciyar hankali na yanayin zafi mara kyau (NBTI) da rushewar dielectric na lokaci (TDDB). Muna nuna yadda yanke shawara na tsarin-tsarin, musamman dabarun shimfiɗa na lokaci-lokaci, zasu iya haifar da muhimman ciniki-tsakanin dorewa da aiki a cikin aikace-aikacen koyon na'ura na zamani.

2. Tsarin Dorewar Crossbars

2.1 Matsalolin NBTI a cikin Lissafin Neuromorphic

Rashin Kwanciyar Hankali na Yanayin Zafi mara Kyau (NBTI) yana faruwa lokacin da cajin inganci ya zama kamar a kulle a iyakar oxide-semiconductor a ƙarƙashin ƙofar na'urorin CMOS a cikin da'irar neuron. Wannan al'amari yana bayyana a matsayin raguwar magudanar ruwa da transconductance, tare da ƙara kashe wutar lantarki da ƙarfin lantarki na bakin kofa.

Tsawon rai na na'urar CMOS saboda NBTI ana ƙididdige shi ta amfani da Matsakaicin Lokacin Gasa (MTTF):

$MTTF_{NBTI} = A \cdot V^{\gamma} \cdot e^{\frac{E_a}{KT}}$

Inda $A$ da $\gamma$ su ne masu alaƙa da kayan aiki, $E_a$ shine kuzarin kunnawa, $K$ shine akai na Boltzmann, $T$ shine zafin jiki, kuma $V$ shine ƙarfin lantarki na ƙofar gaba.

2.2 Hanyoyin Gazawar TDDB

Rushewar Dielectric na Lokaci (TDDB) tana wakiltar wata muhimmiyar matsalar amincin inda oxide ɗin ƙofa ya rushe bayan lokaci saboda matsin lantarki. A cikin crossbars na neuromorphic, TDDB yana haɓaka ta manyan filayen lantarki da ake buƙata don aikin NVM.

Tsarin tsawon rai na TDDB yana biye da:

$MTTF_{TDDB} = \tau_0 \cdot e^{\frac{G}{E_{ox}}}$

Inda $\tau_0$ akai ne na kayan aiki, $G$ shine ma'aunin haɓaka filin, kuma $E_{ox}$ shine filin lantarki a fadin oxide.

2.3 Haɗaɗɗen Tsarin Dorewa

Gabaɗayan amincin kayan aikin neuromorphic yana la'akari da hanyoyin gazawar NBTI da TDDB duka. Jimlar adadin gazawar yana biye da:

$\lambda_{total} = \lambda_{NBTI} + \lambda_{TDDB} = \frac{1}{MTTF_{NBTI}} + \frac{1}{MTTF_{TDDB}}$

3. Hanyar Gwaji

Tsarin gwajin mu yana kimanta dorewar tsawon rai ta amfani da ingantaccen tsarin neuromorphic na DYNAP-SE tare da sandunan guringuntsi na tushen PCM. Mun aiwatar da ma'auni da yawa na koyon na'ura ciki har da rarrabuwar lambobi MNIST da gane lambobi da aka faɗa don tantance tasirin amincin a ƙarƙashin ayyukan aiki na gaske.

Saitin gwaji ya haɗa da:

• 28nm fasahar CMOS don da'irar neuron
• Na'urorin guringuntsi na PCM tare da karatun karatu na 1.8V
• Duba zafin jiki daga 25°C zuwa 85°C
• Zamewa-magani tare da masu canjin aiki

4. Sakamako da Bincike

4.1 Cinikin Dorewa-Aiki

Sakamakonmu yana nuna ainihin ciniki-tsakanin amincin tsarin da aikin lissafi. Ci gaba da aiki a manyan ƙarfin lantarki yana ba da matsakaicin kayan aiki amma yana lalata dorewar tsawon rai. Gabatar da lokutan shimfiɗa na lokaci-lokaci yana haɓaka MTTF sosai yayin kiyaye matakan aiki masu karɓuwa.

4.2 Tasirin Shimfiɗa na Lokaci-Lokaci

Aiwatar da hanyar lissafi ta tsayawa-da-tafi tare da kashi 30% na aiki ya nuna haɓaka MTTF sau 3.2 idan aka kwatanta da ci gaba da aiki, tare da raguwar daidaiton rarrabuwa kawai 15% don ayyukan MNIST. Wannan hanyar tana daidaita matsalolin amincin da buƙatun lissafi yadda ya kamata.

5. Aiwar Fasaha

5.1 Tsarin Lissafi

Algorithm na tsarin lokaci mai sane da aminci yana haɓaka ciniki-tsakanin kayan aikin lissafi da tsufar da'ira. Matsalar haɓakawa za a iya tsara ta kamar haka:

$\max_{D} \quad \alpha \cdot Kayan aiki(D) + \beta \cdot MTTF(D)$

$batun \ zuwa: \quad D \in [0,1]$

Inda $D$ shine zagayen aiki, $\alpha$ da $\beta$ sune abubuwan auna maƙasudan aiki da aminci.

5.2 Aiwar Lamba

A ƙasa akwai sauƙaƙan pseudocode na aiwar mai tsara lokaci mai sane da aminci:

class MaiTsaraLokaciMaiSaneDaAminci:
    def __init__(self, matsakaicin_karfin_lantarki=1.8, mafi_karamin_karfin_lantarki=1.2):
        self.matsakaicin_v = matsakaicin_karfin_lantarki
        self.mafi_karamin_v = mafi_karamin_karfin_lantarki
        self.lokacin_damuwa = 0
        
    def tsara_aiki(self, aikin_lissafi, manufar_aminci):
        """Tsara lissafi tare da takurawa na aminci"""
        
        # Ƙididdige mafi kyawun zagaye na aiki bisa manufar aminci
        zagaye_na_aiki = self.lissafta_mafi_kyawun_zagaye_na_aiki(manufar_aminci)
        
        # Aiwar lissafi na tsayawa-da-tafi
        yayin da aikin_lissafi yana da_aiki():
            # Matakin lissafi mai ƙarfin lantarki
            self.aiwatar_karfin_lantarki(self.matsakaicin_v)
            lokacin_lissafi = zagaye_na_aiki * wannan.lokacin_quantum
            self.aiwatar_lissafi(aikin_lissafi, lokacin_lissafi)
            wannan.lokacin_damuwa += lokacin_lissafi
            
            # Matakin farfadowa mai ƙaramin lantarki
            self.aiwatar_karfin_lantarki(self.mafi_karamin_v)
            lokacin_farfadowa = (1 - zagaye_na_aiki) * wannan.lokacin_quantum
            lokaci.barci(lokacin_farfadowa)
            
    def lissafta_mafi_kyawun_zagaye_na_aiki(self, manufar_aminci):
        """Ƙididdige zagaye na aiki don cika buƙatun aminci"""
        # Aiwar algorithm na haɓakawa
        # la'akari da tsarin NBTI da TDDB
        dawo da ingantaccen_zagaye_na_aiki

6. Aikace-aikace da Jagorori na Gaba

Hanyar lissafin neuromorphic mai sane da aminci tana da muhimman tasiri ga tsarin AI na gefe, motocin cin gashin kansu, da na'urorin IoT inda amincin aiki na dogon lokaci yake da mahimmanci. Jagororin bincike na gaba sun haɗa da:

• Gudanar da Aminci na Karbuwa: Daidaita sigogin aiki na ainihi bisa duban tsufa na ainihi-lokaci
• Tsarin Sikelin Guda da Yawa: Haɗa tsarin amincin matakin na'ura tare da haɓaka aikin matakin tsarin
• Fasahohin NVM Masu Tasowa: Binciken halayen aminci a cikin sabbin fasahohin ƙwaƙwalwar ajiya kamar ReRAM da MRAM
• Koyon Na'ura don Aminci: Yin amfani da dabarun AI don hasashen da rage tasirin tsufa

Yayin da lissafin neuromorphic ke tafiya zuwa ga ɗaukar kowa a cikin aikace-aikacen mahimman aminci, hanyoyin ƙira masu sane da aminci za su zama mafi mahimmanci. Haɗa waɗannan dabarun tare da sabbin tsarin lissafi kamar lissafin cikin ƙwaƙwalwar ajiya da kusan lissafi yana ba da dama masu ban sha'awa don bincike na gaba.

7. Nassoshi

1. M. Davies et al., "Loihi: Na'urar Lissafi ta Neuromorphic da Yawa tare da Koyo akan Guntu," IEEE Micro, 2018
2. P. A. Merolla et al., "Da'irar haɗakar neuron miliyan ɗaya tare da cibiyar sadarwa mai iya faɗaɗawa da mu'amala," Kimiyya, 2014
3. S. K. Esser et al., "Cibiyoyin sadarwa don sauri, lissafin neuromorphic mai ingantaccen kuzari," PNAS, 2016
4. G. W. Burr et al., "Lissafin Neuromorphic ta amfani da ƙwaƙwalwar ajiya mara-sauƙi," Ci gaba a cikin Physics: X, 2017
5. J. Zhu et al., "Kimantawa da Tsarin Amincin Tsarin Lissafin Neuromorphic," IEEE Transactions akan Kwamfutoci, 2020
6. Dabarar Fasaha ta Duniya don Masu Rarraba Lantarki (ITRS), "Na'urorin Bincike masu Tasowa," 2015
7. Y. LeCun, Y. Bengio, da G. Hinton, "Koyo mai zurfi," Yanayi, 2015