
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 Abasiki: 
 
 
 
 
 
 Thomas Feltin Léo Marchó Juan-Antonio Cordero-Fuertes Frank Brockners Thomas H. Clausen Anoten: Ukubala kwe-Deep neural network (DNN) kudatha yokusakaza kufuna izinsiza zokubala ukuze kwanelise izidingo zokukhiqiza ukubala. Noma kunjalo, izicelo zokufunda ezibucayi nge-latency kanye nobumfihlo azikwazi ukubeka isibalo kwelifu elikude ngenxa yezindleko zokudlulisa kanye nokungathembi abahlinzeki belifu bangaphandle. Phakathi kwezixazululo zokukhulisa ukusebenza ngenkathi ugcina ukubalwa endaweni enemikhawulo, ukusheshisa kwezingcingo kungaba nzima, futhi ukwenza kahle kwezimo kudinga imizamo emikhulu yokuklama ngenkathi kuvimbela ukunemba. Ukuhlukaniswa kwe-DNN kuyindlela yesithathu ehambisanayo, futhi ihlanganisa ukusabalalisa umthwalo wokubala phakathi kwamadivayisi amaningi eseduze, kucatshangelwa izimfanelo zenethiwekhi eseduze kanye nesakhiwo se-DNN, ngenhloso yokwandisa ukukhiqiza ukubala (inani lokubala ngomzuzwana). Leli phepha lethula indlela yokubikezela ukubala kanye nokubala kokudlulisa kwezinhlelo zokubala ze-DNN ezisatshalaliswe ezindizeni eziningi, futhi lichaza inqubo yokwenza kahle yokwandisa ukukhiqiza ukubala. I-Branch and bound solver ibe isethulwa futhi ihlolwe ukuze kulinganiselwe ukusebenza okutholakele kanye nobunzima. Lokhu kuhlolwa kuholele ekuchazeni isifunda sokusheshisa, esichaza izimo eziqinisekile ezintweni ezithile ze-DNN kanye nenethiwekhi lapho ukuhlukaniswa kwe-DNN kungenza khona inzuzo. Ekugcineni, imiphumela yokuhlola iqinisekisa izibalo futhi ibonisa ukuthuthukiswa kokukhiqiza ukubala ezindaweni ezimbalwa ezisetshenzisiwe.  Ukuhlukaniswa kwe-DNN eseduze kuvumela ukusheshisa kokukhiqiza ukubala ngenkathi kunciphisa ukusetshenziswa kwebhandwidth kanye ne-latency kumaseva aphakathi nendawo, isb., efwini.  Isingeniso Amadivayisi axhunyiwe akhiqize cishe ama-byte ayizigidi eziyizinkulungwane ezingu-2.5 nsuku zonke ngo-2020.  Kulo mongo, i-Artificial Intelligence (AI) ingenye yezobuchwepheshe ezingcono kakhulu ezingabhekana nenani elikhulayo ledatha elikhiqizwayo, emikhakheni ehlukahlukene njengokuthola izinto ekuboneni ngekhompyutha, ukuqashelwa kobuso, ukuqashelwa kwenkulumo, ukucubungula ulimi lwemvelo, noma ukushayela okuzenzakalelayo. I-AI, futhi ikakhulukazi i-Deep Learning (DL), idinga, ngaphezu kwamavolumu amakhulu edatha, amakhono abalulekile ekubalweni kwamandla kanye nenkumbulo, okwenza ikhompyutha yelifu ibe isixazululo sokubamba samathuba e-AI. Ngenxa yalokho, ngo-2019, 96% yemisebenzi ye-AI yagijima efwini. 1 2 Izicelo ze-AI zingaba nemikhawulo eqinile ye-latency, isb., ekushayeleni okuzenzakalelayo, ukuqapha izimboni, noma noma yikuphi ukubala kwesikhathi sangempela okubandakanya ukuxhumana nezwe langempela. Isibonelo, izikhathi zokuphendula ezijwayelekile ekushayeleni okuzenzakalelayo zidingeka ukuba zibe ngaphansi kwamasentimitha angu-100  , okwenza izikhathi zokuhamba eziya efwini zibe mde kakhulu kulezi zicelo. Ngokufanayo, ezinye izicelo zingadala ukusetshenziswa okubalulekile kwesixhumanisi kunethiwekhi, isb., esimweni sokucubungula ividiyo esezingeni eliphezulu ekuboneni ngekhompyutha, okuyenza zingakwazi ukugijima efwini. Izinqubomgomo zobumfihlo bedatha noma imithetho yokuvikela idatha ingase futhi ivimbele idatha ukuthi ingahambi ezindaweni ezithile. Lokhu kukhuthaza ukusetshenziswa kwe-AI esikhungweni semingcele, esihlanganisa ukubuyisela imisebenzi yokubala esikhungweni sezimiso eziseduze namadivayisi eseduze, isb., eduze kwemithombo yedatha. [](javascript:void()) Uma kuqhathaniswa nezikhungo zamalungu omphakathi, umngcele uyindawo enemikhawulo yezinsiza enezinkinga ezibalulekile  , futhi ukuthuthukisa i-AI kumngcele ngakho-ke kuhlukile kokuqhathaniswa nokuthuthukisa i-AI efwini. Ukuhlaziya idatha eduze nomthombo wayo kungaba yindida ngenxa yamandla aphansi edivayisi, ukungabi bikho kwezinsiza ezitholakalayo, kanye nokuhlukahluka kwezinkundla zemingcele. Ukunciphisa amandla okubala kusho ukunciphisa ukukhiqiza okubalwe ngaphansi, okwenza kube nzima nakakhulu ukubuyisela imithwalo esindayo esikhungweni sezimiso kumngcele. Kwezinye izicelo, ukukhiqiza ukubala, isb., ama-inferences angaphansi ngomzuzwana we-DNN, kungase kuxhunyaniswe nokunemba kwe-DNN, isb., ngokuqapha izinto kumafutha evidiyo, lapho ukwehlisa isivinini sokubala kusuka ku-15 kuye ku-5 ama-inferences ngomzuzwana kubonakale kunciphisa isikolo se-F1__3__wesitholi sezinto ngo-10%  . Ngamanye amazwi, izinhlelo ezinezinga eliphansi lokukhiqiza ukubala zingabangela ukulahlekelwa ulwazi olubalulekile, isb., ukulahlekelwa ukutholwa ezicelo zokuqapha ividiyo. [](javascript:void()) [](javascript:void()) Kunezindlela ezimbili eziyinhloko ezitholakalayo ukuze kwanelise izidingo zokusebenza ezinkundleni zemingcele enemikhawulo: ukucindezela imodeli noma ukusheshisa kwezingcingo.   kuhlanganisa ukunciphisa nokwenza kahle izimanzi ukuze kube inguqulo elula, ngokuvamile engasebenzi kahle, yemodeli. Izindlela ezijwayelekile zokucindezela imodeli zihlanganisa ukususa  , isb., ukususa amapharamitha angadingekile kumodeli, ukuqokwa  , isb., ukunciphisa inkumbulo ebekelwe ukugcina imodeli, kanye nokufundisa ngolwazi  , isb., ukuqeqesha imodeli encane yenziwa ngolwazi oluthathwe kumodeli enkulu. Ukwenza kahle kwemodeli kudinga imizamo emikhulu yokuklama futhi kungaba isithiyo ekunembeni kwemodeli.   emadivayisini emingcele kusho ukuthola izingcingo ezingasebenzisa amandla kancane futhi ezilula, kodwa zisazokwazi ukusebenzisa ukubala ngokusebenza okwanele. Kunezinhlobo eziningana zezinto ezingakwazi ukusheshisa kumngcele, ezihlukahlukene ngamandla nokukhetheka, kodwa amayunithi okubala ajwayelekile ahluleka ukwanelisa izimlindelo futhi izingcingo ezikhethekile njenge-ASICs kuye kwatholakala ukuthi azisebenzi futhi azibizi  . Ukucindezela imodeli [](javascript:void()) [](javascript:void()) [](javascript:void()) Ukusheshisa kwezingcingo [](javascript:void()) Ukuhlukaniswa kwe-DNN kuyindlela ehambisanayo yokusheshisa ukubala ngokusebenzisa izimfanelo zamadivayisi akhona ezinkundleni zemingcele ukusabalalisa ukubala kokubala – okungasetshenziswa uwedwa, noma ngokuhlangana nezinye izindlela ezimbili. Ukuhlukaniswa kwe-DNN kuhlanganisa ukubheka inani lezimanzi njengomjikelezo wokuhlukanisa ukuze uhlukaniswe, futhi uhlukanise lezi zinhlanganisela kumadivayisi eseduze. Indawo yalezi zinhlanganisela isekelwe kokubili ku-DNN kanye nezimfanelo zenethiwekhi engaphansi. Ukuhlukaniswa kwe-DNN kusekelwe ekuhlonzeni amaphuzu okwehlukanisa, okuyizindawo kumdwebo wemodeli lapho imodeli ihlukaniswa khona izinhlanganisela. Ngesikhathi sokusebenza, izinhlanganisela ziyasetshenziswa ngokulandelana, ngamunye ethumela imiphumela yokubala yesikhashana kuye kwinhlangano elandelayo. Lokhu kuvumela inhlangano ngayinye ukuba iqale ukubala idatha elandelayo ngenkathi amanye amadivayisi eqhubeka nokucubungula eyathunyelwe, ngaleyo ndlela kuthuthukiswe ukukhiqiza ukubala, njengoba kuboniswe kumdwebo wezithombe wezithombe 1. Ekupheleni kwaleli phepha, ukuhlukaniswa kwe-DNN kuzovezwa ngokwesibonelo sokubala kwesikhathi sangempela kumafutha evidiyo.  A. Imisebenzi Ehlobene Izindlela zokuhlukaniswa kwe-DNN, ezithathwe ekukhishweni kwe-cloud-edge mobile, zifuna ukwenza kahle izinkomba ezahlukene zokusebenza, njenge-computing latency, ukusetshenziswa kwamandla, ukusetshenziswa kwezinsiza, izindleko, noma ukukhiqiza. Ukuhlukaniswa kwe-DNN kusekelwe ekuhlangeni kwamunye noma okunye kwalawa malithi ukuze kuchazwe umgomo wokwenza kahle, kanye nendlela eyenza inzuzo kuleli methriki ukuthola izikimu zokuhlukanisa. Isibonelo, i-Neurosurgeon   ifuna iphuzu elilodwa lokwehlukanisa, igcina inhlangano yokuqala eseduze futhi ikhiphe inhlangano yesibili efwini, ukuze kuncishiswe i-latency namandla. Izicelo ku-IoT nazo ziye zacabangela ukuhlukaniswa okuhlangene nokukhishwa kwezimanzi eziningi ukuze kwenziwe kahle amandla, ukubambezeleka, kanye/noma izindleko, kusetshenziswa izixazululi ezahlukene njenge-SPSO-GA   ehlanganisa i-Particle Swarm Optimization (PSO) kanye ne-Genetics Algorithm (GA), noma i-DDPQN   esebenzisa i-Deep Reinforcement Learning ukuthola izikimu zokuhlukanisa. Ezinye izibonelo zihlanganisa i-DINA  , echaza inkinga ye-Integer Non Linear Programming (INLP), futhi isebenzisa isixazululi esisekelwe kumqondo wokufanisa, ukuze kwenziwe kahle ukubambezeleka nokusetshenziswa kwezinsiza ezinkundleni zenkungu. Izindlela zokukhiqiza ukubala okuthuthukisiwe zihlanganisa i-DNN surgery  , esebenzisa usayizi ophansi wemiphumela ye-DNN ye-interim ukuhlukanisa ukubala kokubala phakathi komngcele nefu. Zombili lezi zibalo zenza izinhlanganisela ezimbili, enye eseduze nenye efwini. Ezinye izifundo emkhakheni we-cloud-edge mobile   zihlanganisa nokubala kwezinto eziningi efwini ukuze kukhuliswe kakhulu ukukhiqiza okubalwe kuso sonke. Ukuthembela efwini ekubalweni kwenhlanganisela yesibili kuthatha ukuxhumana nenethiwekhi okuhle, okungaba okungaqiniseki ngezinga lokuxhumana elibi lezinhlelo ezithile eziseceleni. I-Edgent   ilungele indlela yangaphambilini kumadivayisi eselula kanye namaseva aseduze atholakalayo, kunciphisa umkhawulo wokukhishwa efwini. Le ndlela ihlobene nokuqeqeshwa okuthile kanye nesakhiwo semodeli, futhi isacela ukuhlukaniswa kube izinhlanganisela ezimbili. Ezinye izindlela ziye zacabangela amaphuzu amaningi okwehlukanisa,  , i-IONN  , echaza inkinga yokuhlukanisa ukuze kuncishiswe i-latency nokusetshenziswa kwamandla, iphinde icubungule isikhathi sokulayisha amamodeli kumaseva aseduze. [](javascript:void()) [](javascript:void()) [](javascript:void()) [](javascript:void()) [](javascript:void()) [](javascript:void()) [](javascript:void()) isb. [](javascript:void()) Lezi zindlela zisebenzisana namalifu akude, noma zikhawulelwe kuhlobo lokuhlukanisa,  , zidingeka ukuba zihlukaniswe ngezingxenye ezimbili ngqo, ngezinto eziningi noma ngaphandle kwazo. Ayikho neyodwa yezifundo ezishiwoyo ze-DNN partitioning ezicubungula kokubili amaphuzu amaningi okwehlukanisa kanye nezinto eziningi. isb. B. Okufakiwe Leli phepha lethula uhlelo lokubala oluhlukahlukene ukuze kwenziwe kahle ukukhiqiza ukubala kwe-DNN yesikhathi sangempela kudatha yokusakaza enemiphumela eminingi yokwehlukanisa kanye nezinto eziningi ezihlukahlukene. Okufakiwe kwale dokhumenti yilezi ezilandelayo: 
 
 
 
 
 
 Imodeli yokubala kanye nokubala kokudlulisa kwe-DNN esatshalaliswe, lapho kunganikezwa khona isilinganiso sokukhiqiza ukubala kwe-DNN enikeziwe. Ukuhlelwa kwenkinga yokwenza kahle yokuhlukaniswa kwe-DNN, ukusetshenziswa kwesixazululi se-branch and bound salezi nkinga, kanye nokuhlola kwayo ubunzima. Izibalo zokuhlola ukuthuthukiswa kokusebenza okungenzeka ngokushintsha izimfanelo zenethiwekhi kanye ne-DNN. Ukuhlonzwa kwezifunda eziqinisekile ezintweni zenethiwekhi kanye ne-DNN eholela ekubeni khona kwezixazululo ezenziwe kahle kanye nezindleko zokubala lezi zimiphumela, isb., izimo lapho ukuhlukaniswa kwe-DNN kungenza khona inzuzo. Imiphumela yokuhlola ebonisa izifunda ezichazwe ezibalweni, kanye nokunemba kokubikezela kanye nokukhiqiza okukhulu kokubala kwezikhathi ezithile zokuphakathi kwezingcingo kanye nezindawo ezingafani nezindawo ezingafani. Ngokulandelayo kwale dokhumenti, ukuze kugwenywe ukudideka, i-deep neural network ehlangene izobhekiselwa kuyo njenge  , futhi igama elithi   lizobhekisela kunethiwekhi yokuxhumana yomzimba engaphansi. DNN inethiwekhi C. Uhlu Lwadokumenti Okulandelayo kwale dokhumenti kuhlelwe kanje.   siletha ingemuva le-Deep Learning nendlela ama-DNN amelelwa ngayo kulesi sifundo.   sichaza izindlela zokubala kanye nokubala kokudlulisa ezidingekayo   esichaza inkinga yokwenza kahle, kanye ne-algorithm ye-branch and bound ukuze yenziwe izinqumo zokuhlukanisa.   siletha imiphumela yezibalo ukuze kulinganiswe ukusebenza nobubanzi be-algorithm ye-branch and bound, kanye nezimo eziholela ekubeni khona kokuhlukanisa okuthuthukisa ukukhiqiza ukubala kunethiwekhi efanayo.   siletha imiphumela yokuhlola eqinisekisa izibalo ze  , futhi ibonisa ukuthuthukiswa kokukhiqiza ukubala kusethi yezinto zokuhlola ezingafani. Ekugcineni,   siyaxoxisana ngemiphumela futhi sichaza ukusetshenziswa okubanzi kwale ndlela. Isigaba II Isigaba III Isigaba IV Isigaba V Isigaba VI Isigaba V Isigaba VII ISIGABA II. Ingemuva: Ukufunda Okujulile Emkhakheni we-Machine Learning, i-Deep Learning (DL) ibhekisela ezindleleni ezisekelwe ekusetshenziswa kwe-DNNs, isb., amanethiwekhi enziwe omuntu (noma izindlela ezihlobene zokufunda komshini) ezinemigubuzelo engaphezu komunye. Ezinye zezinkomba zakuqala ezikhuluma ngezinhlelo zokufunda ezijulile zaphrintwa ngo-1967  , kodwa kwaze kwaba ngama-2010 lapho i-DL yazuza udumo, kuqala ekuqashelweni kwenkulumo, bese kuthi ekutholeni izithombe emikhiqizweni, ngenxa yamandla okubamba ubudlelwano obuyinkimbinkimbi kumaloli aphezulu wedatha. I-Deep Learning ilinganiselwa ukuba kube ingxenye enkulu yezicelo ze-AI ngo-2022, ngaphezu kwama-75% ezinhlangano ezisebenzisa ama-DNNs ezicelo ezingasebenzisa izindlela zakudala. Izindlela ze-DL ziye zachazwa njenge “   ”  . Le mijikelezo ye-algebra ngokuvamile ihlelwa ngemigubuzelo, eyenza izinyathelo ekubalweni. Igama elithi   libhekisela ezinhlelweni ezihlanganisa imigubuzelo engaphezu komunye. [](javascript:void()) ubuchwepheshe bokufunda komshini lapho izicelo zithatha uhlobo lwemijikelezo ye-algebra eyinkimbinkimbi enezinhlobonhlobo eziguquguqukayo [](javascript:void()) deep Kunezinhlobo ezimbili eziyinhloko zezakhiwo ze-DNN: amanethiwekhi okudluliswa phambili kanye namanethiwekhi abuyayo. Zombili izinhlobo ze-DNN zingameliwe njengemidwebo yokubala, nomehluko omkhulu ukuthi amanethiwekhi okudluliswa phambili angameliwe njengemidwebo eqondile eyenziwe ngokucacile (DAGs), ngenkathi amanethiwekhi abuyayo angase aqukethe imijikelezo. Amanethiwekhi abuyayo akhiwe ngamachungechunge edatha, lapho iphuzu ngalinye ledatha lingaba nobudlelwano namaphuzu adlule echungechungeni, ngenkathi amanethiwekhi okudluliswa phambili angacubunguli ubudlelwano phakphakathi kwamaphuzu echungechunge ledatha. Okulandelayo kwale dokhumenti kuzobheka isimo sama-DNN okudluliswa phambili. 5 ISIGABA III. Imodeli Yokubala Ehlukahlukene Lesi sigaba sethula imodeli yokumela izici ze-DNN kanye nenethiwekhi ukuze kubikezelwe izikhathi zokubala, izikhathi zokudlulisa, kanye nokukhiqiza kokugcina kokubala kwesixazululo esithile sokuhlukanisa. A. Ukumelwa kwe-DNN kanye Nenethiwekhi I-DNN yokudluliswa phambili enezinkomba ezingu-N imelwe njenge-DAG GA=(L,E) nenemibhalo L=(L1,…,LL) amagebe e-DNN. Imikhawulo (Li,Lj)∈E izixhumanisi phakathi kwamagebe Li kanye Lj . Ilalelana Li > has a compute consumption ci >, measured in the number of floating-point operations required to compute a forward pass through the layer. Edges (Li,Lj) > are assigned a weight si,j > corresponding to the size of the data transiting between layers Li > and Lj > in bytes. Inethiwekhi yomzimba imelwe njengomdwebo oxhunywe ngokuphelele G=(N,V) > lapho N=(N1,…,NN) > is the set of compute nodes and V is the set of links between nodes. It is assumed that compute nodes N1,…,NN > have processing rates η1,…,ηN , respectively, measured in floating-point operations per second. Finally, the link throughput between two adjacent nodes Na > and Nb > is denoted as θa,b , measured in bytes per second. Every node Ni > is connected to itself with infinite throughput to represent the loopback link, and links are assumed to be symmetrical,  ,  θa,b=θb,a . i.e. Izinhlelo zokuhlukanisa zichazwa njengama-maphu P:L→N ,  , a partitioning assigns a node number to each layer in the DNN. A partitioning can be described as a matrix P of dimension (N×L) , N > being the number of nodes on the network and L > the number of DNN layers and, with Pa,i=1 > if layer Li > is placed on node Na >, 0 otherwise. With the example given in  , L=7 > layers and N=2 > compute nodes, the displayed partitioning with layers {L1,L2,L3,L6,L7} > on node N1 , and layers {L4,L5} > on node N2 , is represented as: isb. figure 2 Given a partitioning, a   > is defined as a group of consecutive layers between two split points, run sequentially on the same node. In the example above, node N1 > will run two threads, the first containing layers {L1,L2,L3} > and the second one containing {L6,L7} . thread With these notations defined, the remainder of this Section derives a closed expression of the inference throughput achieved by a given partitioning, as a function of the inference ( ) and transmission ( ) latencies. section III-B section III-C B. Ukubikezela i-Inference Latency Given a partitioning, this Section looks at inference latency prediction on an isolated node. The latency induced by the computation of a layer Li > with consumption ci >, to be computed on node Na >, with processing rate ηa is expressed as Tci(Na)=ci/ηa . Given a set of layers L′⊂L > across all threads running on node Na > with processing rate ηa >, the inference latency can be expressed as: This expression is a first order approximation of an inference latency estimation based on the number of floating-point operations (FLOPs) required to process the DNN,  , the number of multiply-add operations in the model. i.e. Estimation of inference latency on edge devices is complicated by run-time optimizations. Existing solutions such as nn-Meter   predict inference latency based on FLOPs. Other solutions rely on a look-up table with pre-computed inference times for latency inference. For example, BRP-NAS   uses a pre-trained graph convolutional network to predict inference latencies, while taking run-time model optimizations into account. [](javascript:void()) [](javascript:void()) Ku  , kuphinde kucatshangelwe ukuthi abacubungula basebenzisa amandla abo aphelele, okusho ukuthi abacubungula abaningi bahlanganiswa njengabacubungula abodwa abanamandla okucubungula alingana nesamba sezikhombisi zabo zokucubungula. Kuphinde kucatshangelwe ukuthi insiza yokubala ihlukaniswe ngokulinganayo phakathi kwezintambo, isb., umcibisheli uhloka isabelo esifanayo sokuphila kwawo zonke izintambo. isibalo 1 Kuzo zonke izindlu, intambo enesikhathi eside kakhulu ibeka i-inference throughput ye-DNN ehlukahlukene. Ukukhipha izikhathi zokudlulisa, lokhu kusho ukuthi uma le ntambo eyinhloko isiqedile ukubala ifutho ledatha, izoqala ngokushesha ukucubungula okulandelayo. Ezinye izintambo zizoba nesikhathi sokuphumula ngaphambi kokucubungula ifutho ledatha elilandelayo ngoba izikhathi zazo zokubala zincane kunesikhathi esiphezulu, njengoba kuboniswe kumdwebo wezithombe 1. Given a partitioning matrix P , the inference latencies can be expressed as a single vector Tc of size N > where the a -th component is the highest thread inference latency on node Na ,  ,  Tca=Tc(Na) : i.e. where H > is the diagonal matrix of inverse node processing rates diag(η−11,…,η−1N) and c > is the column vector of individual layer consumptions (c1,…,cL)&

Part of HackerNoon's growing list of open-source research papers, promoting free access to academic material.

Read My Stories

Cisco Systems

Codkan waxa lagu soo saaray afkii asalka ahaa ee sheekada!

Mwanamume Kwanini AI Yako Haipati Tena Wingu Ili Kuendesha Haraka

About Author

FAALLO

HANG TAGS

MAQAALKAN AYAA LAGU SOO BANDHIGAY

Related Stories

Scary-good Writing Tips: Learn How to Write a Listicle

Meet the Writer: HackerNoon's Contributor Aayush Naik, Software Engineer @ Apple

10 Stories To Learn About Medium

69 Stories To Learn About Productivity Tools

Scary-good Writing Tips: Learn How to Write a Listicle

Meet the Writer: HackerNoon's Contributor Aayush Naik, Software Engineer @ Apple

10 Stories To Learn About Medium

69 Stories To Learn About Productivity Tools

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps