njengamanje, i-GPU ichitha iningi lamandla ayo hhayi ezibalweni, kodwa ekuhambeni kwezinombolo. Ukulanda izisindo kumemori, ukuzifaka kuyunithi yekhompyutha, ukwenza ukuphindaphinda, nokubuyisela umphumela. Yenza lokho izikhathi eziyizigidigidi ngomzuzwana, izinyanga, ngesikhathi sokuqeqeshwa, futhi uqala ukuqonda ukuthi kungani isikweletu sikagesi se-AI sesiyinkinga ye-boardroom kunokuba umbhalo waphansi wobunjiniyela.
Izinombolo zenza lokhu. U-Gartner ulindele ukuthi isidingo samandla esikhungo sedatha yomhlaba wonke sikhuphuke sisuka kumagigawathi angu-104 ngo-2025 siye kumagigawathi angu-132 ngo-2026, futhi sifinyelele cishe amagigawathi angu-290 ngo-2030, ngomthwalo okhiqizayo we-AI obizwa ngokuthi umshayeli oyinhloko. Amaseva enziwe nge-AI ewodwa kulindeleke ukuthi aphendule cishe ingxenye yesithathu yokusetshenziswa kwamandla esikhungo sedatha kulo nyaka.
Lesi ingemuva lokuvela kabusha okuthulile kombono wakudala: ikhompuyutha ye-analog. Amasekhethi angempela e-analog, kusetshenziswa ama-voltage nama-currents aqhubekayo esikhundleni sika-1s no-0s, ukuze kwenziwe ukuphindaphinda kwe-matrix okubusa ukuchazwa kwenethiwekhi ye-neural. I-pitch iyayenga: esikhundleni se-computing mayelana i-physics ngokwedijithali, kungani ungavumeli i-physics ukubala ngokuqondile?
Akuwona umqondo omusha, vele, amakhompiyutha e-analog andulela awadijithali. Lalahlwa emashumini eminyaka adlule ngenxa yezizathu ezinhle. Futhi le ndatshana imayelana nohhafu waleyo ndaba: kungani iphimbo liphoqa ngempela, nokuthi kungani okubulale ikhompuyutha ye-analog okokuqala kungakasuki. Ngizohamba ngomshini, bese ngisebenzisa izifaniso ezincane, ezithembekile ezibonisa lapho iphuka khona, nokuthi abacwaningi baqala ukuyipeca kanjani.
Isebenza kanjani i-analog in-memory computing
Ku-chip evamile, inkumbulo nokubala kuhlukene. Ngaso sonke isikhathi uma uphindaphinda isisindo ngokokufaka, leso sisindo kufanele sihambe sisuka kuhlu lwenkumbulo, sinqamule ibhasi, siye kuyunithi yekhompyutha, futhi umphumela kufanele ubuyele emuva. Abacwaningi balinganisela ukuthi lokhu kunyakaza kungabiza noma yikuphi ukusuka ku-3 kuya ku-10,000 amandla angaphezu kwezibalo ngokwazo, lokho okubizwa ngokuthi i-von Neumann bottleneck.
I-Analog in-memory computing (AIMC) ibhidliza lolo hambo. Izisindo zigcinwa njengamavelu okuziphatha okuphathekayo kugridi yamaseli enkumbulo, ngokuvamile akhiwa kusuka kumemori yoshintsho lwesigaba noma i-RAM eqinile. Faka okokufaka njenge-voltage kuleyo grid, futhi imithetho emibili yefiziksi yenza okunye: Umthetho ka-Ohm (yamanje = i-voltage × conductance) ibala ukuphindaphinda ngakunye, kanye Umthetho wamanje kaKirchhoff (i-currents sum uma izintambo zihlangana) ibala ukunqwabelana. Ukuphindaphinda kwe-matrix-vector egcwele, amanethiwekhi e-neural okusebenza achitha isikhathi esiningi ekhona, kwenzeka ngesinyathelo esisodwa somzimba, alikho ibhasi elidingekayo.
Kukhona okucashile kulowo musho, nokho: ukuphindaphinda ngokwako kumahhala, kodwa okokufaka nokukhiphayo kusafanele kweqe umngcele phakathi komhlaba we-analog nedijithali: ama-voltage phakathi, ukuphuma kwemisinga, kokubili kudinga ukuguqulwa. Empeleni, ama-ADC kanye nama-DAC enza lokho kuguqulwa kuvame ukuba lapho okwangempela isabelomali samandla nendawo siyahamba, hhayi i-crossbar array ngokwayo. Kuyindida kancane enhliziyweni yenkundla: i-physics yenza izibalo mahhala, bese kuchithwa isabelo esiphusile se-chip ukuze sikhulume nayo.
Lokhu akukona ukuqagela. I-chip yephrojekthi ye-IBM yoCwaningo ye-HERMES iyisibonelo esisebenzayo, i-analogi esekwe kumemori esekwe esigabeni se-analog esesivele sisetshenziswa ukukhombisa ukunemba kwesoftware egcwele imisebenzi yangempela. I-prototype yangaphambilini esekwe ku-IBM PCM, eyabikwa ngo-2023, yapakisha amaseli enkumbulo okushintsha isigaba ayizigidi ezingama-35 futhi yabamba amapharamitha afinyelela ezigidini eziyi-17 ngqo ku-chip, engxenyeni yamandla esheshisi sedijithali. Iziqalo ezifana ne-EnCharge AI (ephuma e-Princeton, esebenza ne-TSMC ekukhiqizeni) kanye ne-Mythic ikhuphule imali yangempela yokubheja i-physics efanayo ingakhiqizelwa ukucabangela onqenqemeni.
Ukulingisa ngokwakho
Awudingi i-hardware yangaphandle ukuze ubone indlela esebenza ngayo. Nasi ukulingisa okuncane: ukuphindaphinda kwe-matrix yedijithali “ephelele”, uma kuqhathaniswa nokusebenza okufanayo nokungapheleli okungokoqobo kwe-analogi okubekwe ngaphakathi, umsindo wohlelo olusuka kudivayisi kuya kudivayisi (izisindo azibhalelwa ngendlela eqondile oyicelile), funda umsindo (umsindo oshisayo/ oshisayo njalo uma usebenzisa uhlelo), futhi ukuguqula okulinganiselwe kwe-ADC kusenomphumela wedijithali (umphumela we-analogi).
import numpy as np
rng = np.random.default_rng(42)
n_in, n_out = 64, 32
W = rng.uniform(-1, 1, size=(n_in, n_out)) # "ideal" weights
x = rng.uniform(-1, 1, size=n_in) # input activations
digital_result = x @ W # ground truth
def analog_matmul(x, W, programming_noise_std=0.0, read_noise_std=0.0, adc_bits=None):
# weights are physically stored as conductances -> noise is baked in
# until the cell is reprogrammed
G = W + rng.normal(0, programming_noise_std, size=W.shape)
# Ohm's law + Kirchhoff's current law, in one step
out = x @ G
# thermal/read noise added at measurement time
out = out + rng.normal(0, read_noise_std, size=out.shape)
# the analog result still needs to be digitized
if adc_bits is not None:
lo, hi = out.min(), out.max()
step = (hi - lo) / (2 ** adc_bits)
out = np.round((out - lo) / step) * step + lo
return out
Ukwenza lokhu ngamanani akhulayo okubona ngokoqobo kukhiqiza ijika lephutha elihlanzekile, elikhulayo:
Ukusuka ekuphindaphindeni okuphelele kuye kokukodwa ngayo yomibili imithombo yomsindo kanye ne-4-bit ADC eshibhile yethula iphutha elihlobene elingaphezu kuka-8%, ungqimba olulodwangaphambi kokuba lelo phutha libe nethuba lokuhlanganisa kunethiwekhi ejulile. Okuphakamisa umbuzo wangempela: kubaluleke kangakanani lokho ngokunemba kwemodeli eqeqeshiwe?
Ikhompuyutha ye-analogi ibilokhu inenkinga yomsindo
Yingakho ikhompuyutha ye-analog yalahleka yaba yidijithali kwasekuqaleni, amashumi eminyaka ngaphambi kokuthi i-AI iyenze isebenze futhi. Izimpawu zomzimba eziqhubekayo zinzima ngokwemvelo ukuzicindezela kunezimo ezimbambili. Isekhethi yedijithali kufanele ihlukanise u-0 koku-1; isifunda se-analogi kufanele silondoloze ubukhulu obunembile ngokumelene nomsindo oshisayo, ukuhlukahluka kwedivayisi, kanye nokukhukhuleka ngokuhamba kwesikhathi. Kuyinkinga yobunjiniyela enzima kakhulu, futhi ayisuki ngenxa yokuthi uhlelo lokusebenza lushintshile ukusuka ekuxazululeni izilinganiso ezihlukene ukuya ekuqaliseni amanethiwekhi e-neural. Umsebenzi wakamuva wokwenza ikhathalogi le miphumela ubala umsindo wohlelo, umsindo ofundwayo, ukukhukhuleka kokuziphatha kwesikhashana, ukunemba kancane kancane, nokulahlekelwa kokuguqulwa kwe-analog kuya kwedijithali njengezinkinga ezivulekile ezimile kulo mkhakha. I-Drift iyinqaba kakhulu ukuhlala nayo: ngokungafani nebhithi yedijithali, isisindo se-analog siyinto ebonakalayo, futhi leyo nto ikhululeka ngokuhamba kwesikhathi, ukuze isisindo “esifanayo” sifunde ngokuhlukile ngesonto ngemva kokubhalwa, yingakho ama-analog chips ngokuvamile adinga ukulungiswa kabusha kwezikhathi noma ukuhlela kabusha ukuze nje ahlale enemba, inkumbulo yedijithali engaphezulu ayikhokhi.
Ukuze ngibone ukuthi lokhu kubalulekile yini, ngiqeqeshe inethiwekhi encane ye-neural ngendlela evamile, ngokunemba okugcwele kwedijithali, kungekho msindo noma kuphi, kumsebenzi wokuhlukanisa amadijithi obhalwe ngesandla wakudala, ngabe sengisebenzisa ukusikisela ngongqimba olulingisiwe lwe-analog emazingeni akhulayo omsindo.
def forward(params, X, layer2_noise_std=0.0):
h = np.maximum(0, X @ params["W1"] + params["b1"]) # digital hidden layer
W2 = params["W2"]
if layer2_noise_std > 0:
# this line is the analog crossbar: weights corrupted by device noise
W2 = W2 + rng.normal(0, layer2_noise_std, size=W2.shape)
logits = h @ W2 + params["b2"]
return softmax(logits), h
Inethiwekhi yaqeqeshwa ingenalo ulwazi lokuthi izoke iwubone umsindo. Nakhu okwenzekile lapho ngiyihlola ngokusebenzisa ihadiwe elilingisa ngokuqhubekayo:
Inethiwekhi ibekezelela ngokuphawulekayo amanani amancane omsindo, njengoba ukunemba kuhamba kancane ngaphansi kwe-std yomsindo ≈ 0.1. Kodwa esedlule lelo phuzu, alehlisi isithunzi ngomusa, iwela eweni: 83% ku-0.2, 64% ku-0.4, futhi ngo-0.8 isondela ekuqageleni okungahleliwe. Lona umumo wangempela “we-analog AI inenkinga yomsindo”, hhayi intela yokunemba emnene, kodwa umkhawulo inethiwekhi ayizange ilungele ukuwela.
Ukuqeqesha inethiwekhi yezingxenyekazi zekhompuyutha ezosebenza kuyo
Impendulo yamanje yocwaningo kulokhu ayikho (nje) mayelana nokwakha ihadiwe ethule, imayelana nokungaqeqesheli inethiwekhi ukuthi ilindele isignali ehlanzekile kwasekuqaleni. Inqubo ihamba ngamagama ambalwa kuye ngokuthi ubani oyishicilelayo, i-IBM ibiza uhlobo lwayo lokuqeqeshwa kolwazi lwehadiwe, kanye nomugqa wakamuva wocwaningo ngalokho eliwabiza ngokuthi “amamodeli esisekelo se-analog” ukubiza ngokuthi ukuqeqeshwa komjovo womjovo womsindo, kodwa umqondo uyafana: linganisa umsindo wehadiwe okuhlosiwe. ngesikhathi ukuqeqeshwa, hhayi nje ekuthunyelweni, ngakho-ke ukwehla kwe-gradient kuyaphoqeleka ukuthi kutholwe izisindo eziqinile kukho.
Ngiphinde ngenza ukuhlola okufanayo, kodwa kulokhu ngiqeqeshe ikhophi yesibili yenethiwekhi ngomsindo ofakwe kuwo wonke amaphasi oya phambili ngesikhathi sokuqeqeshwa:
def train(params, X, Y, epochs=300, lr=0.1, train_noise_std=0.0):
for epoch in range(epochs):
# train_noise_std > 0 means the network never sees a clean signal —
# it has to learn weights that tolerate the noise from day one
probs, h = forward(params, X, layer2_noise_std=train_noise_std)
# ...gradient descent as normal from here
Umphumela uyishadi elithakazelisa kakhulu kulesi siqeshana:
Ngomsindo onguziro, amamodeli amabili afana ngokwezibalo: ukuqeqeshwa kokwazi umsindo akubizi lutho uma i-hardware ihlanzekile. Kodwa njengoba umsindo ukhula, igebe livuleka ngokushesha: ku-std=0.6, imodeli ejwayele ukuqeqeshwa ihle yaba ku-39% kuyilapho imodeli eqaphela umsindo isabambe u-61%. Ayikho imodeli amasosha omzimba ukwenza umsindo, njengenjwayelo, akukho ukudla kwasemini kwamahhala, futhi ukunemba kusancipha, kodwa enye yazo yehlisa isithunzi njengohlelo olwakhelwe lokhu, kanti olunye ludicilela phansi njengohlelo olungakaze lutshelwe ukuthi luzoqhubeka lunjani.
Lesi isibonelo nje samathoyizi, inethiwekhi enezingqimba ezimbili ezithombeni ezinamadijithi angu-8×8, hhayi imodeli yangempela yokukhiqiza, kodwa iyindlela efanayo abacwaningi bemishini manje abayisebenzisa kuma-LLM, nenkinga eyengeziwe yokuthi amamodeli ezilimi amakhulu anezendlalelo eziningi kakhulu zomsindo ozohlanganisana kuwo wonke, kanye nokubekezelela okuncane kakhulu kohlobo lwewa lokunemba eliboniswe ngenhla.
Lapho lokhu kumi ngempela
Kafushane nje, kufanelekile ukwethembeka mayelana nesimo senkundla. Izinto ezimbalwa zibonakala ziqinile:
Kubukeka kungokoqobo ngokuqonda, hhayi okwamanje ukuqeqeshwa. Cishe yonke imithombo ebucayi kulokhu idonsa umugqa ofanayo: ihadiwe ye-analog iyasebenza ekusebenziseni imodeli esivele iqeqeshiwe, ikakhulukazi emaphethelweni, kodwa izimfuno ezinembayo zokusatshalaliswa kwe-backpropagation zenza ukuqeqeshwa kuma-analog substrates kube yinkinga engaxazululeki enzima kakhulu.
Irekhodi lethrekhi yezohwebo lixubile. I-Mythic, enye yezinhlelo zokuqala ze-analog AI chip startups, isichithe iminyaka izulazula emakethe lapho ama-chips edijithali anamandla aphansi eqhubeka ethuthuka nawo, njengoba inyuse umzuliswano wamaRandi ayizigidi ezingu-125 ngasekupheleni kuka-2025. I-EnCharge AI, eyasungulwa ngo-2022, iqhubekele phambili ngokushesha ekukhiqizeni ubambiswano, isebenzisana ne-TSMC futhi ifuna ukusetshenziswa kwedijithali cishe okukhulu kunama-chips aphansi angama-20x. ngaphambi kokuhweba. I-IBM ihlala kakhulu kumodi yocwaningo, isanda kushicilela izendlalelo ze-LLM ezixubene nochwepheshe ekwakhiweni kwenkumbulo ye-analog ye-3D.
Ukubheja okukodwa phakathi kokuningana, hhayi indaba yonke. I-Photonic computing (esebenzisa ukukhanya esikhundleni se-voltage, elandelwa izinkampani ezifana ne-Lightmatter kanye ne-Celestial AI) i-wager ehlobene kodwa ehlukile, futhi namuhla ihloselwe kakhulu ukuxazulula izingqinamba ze-chip-to-chip interconnect kunokushintsha ikhompuyutha ngokwayo. Ama-Neuromorphic chips afana nolayini we-Intel's Loihi athatha enye indlela, eboleka isiginali eqhutshwa ubuchopho, esekwe ku-spike esikhundleni sezibalo eziminyene ze-analog matrix. Bobathathu bahlanganiswa ndawonye ekusakazweni kwabezindaba njengokuthi “ikhompyutha ephefumulelwe ingqondo,” kodwa baxazulula izingxenye ezihlukene zenkinga.
Akukho kulokhu okulungisa inkinga yamandla e-AI ngokwayo, futhi akufanele ithengiswe ngaleyo ndlela. Kodwa kuwukucatshangelwa kabusha kwezakhiwo zangempela ngesikhathi lapho izinzuzo zedijithali ziphelelwa ukuwina okulula – futhi ngokungafani nemigwaqo eminingi yezingxenyekazi zekhompiyutha, ungakwazi ngempela ukuqinisekisa isimangalo esimaphakathi, nenkinga emaphakathi, kukhompyutha yakho ephathekayo emigqeni engaba ngamashumi amathathu ye-numpy.
Imithombo nokufunda okuqhubekayo
- Gartner, Ukusetshenziswa kukagesi kwesikhungo sedatha kuzokhula ngo-26% ngo-2026
- I-IBM Research, i-Analog in-memory computing ingaba namandla kumamodeli e-AI akusasa
- Ucwaningo lwe-IBM, Umndeni we-AIU wama-chips
- Ucwaningo lwe-IBM, NorthPole: i-architecture ye-neural network inference ene-12nm chip (idijithali, hhayi i-analog – ifakiwe ekuhlukaniseni)
- I-Büchel et al., I-Kernel Approximation isebenzisa i-Analog In-Memory Computing – i-chip yephrojekthi ye-HERMES, i-analog chip ye-IBM ye-PCM
- IEEE Spectrum, Isithembiso se-Femtojoule se-Analog AI
- I-IEEE Spectrum, EnCharge's Analog AI Chip Ithembisa Amandla Aphansi Nokunemba
- I-TechCrunch, i-EnCharge inyusa u-$100M+ ukusheshisa i-AI isebenzisa ama-analog chips
- “Amamodeli we-Analog Foundation Abhekana Nomsindo we-AI Hardware for LLMs” – neurotechnus.com
