Wuce dokar Moore: in-memory compute da chiplets don edge robotics

TL;DR. A Economist Enterprise Physical AI and Robotics Summit a London, Taner Ozcelik na Mythic AI ya bayyana yadda in-memory compute ke rushe bangon ƙwaƙwalwa na kwamfuta.

Mahimman Bayanai

Bangon thermodynamic: dalilin da yasa physical AI ba za ta iya rayuwa a cloud ba. A zaman fireside mai taken Moving past Moore’s law: the power behind high-performance robotics, Kenneth Cukier ya tsara babbar matsalar injiniya da ke fuskantar cin gashin kai na kamfanoni: samfuran artificial…
Ra'ayin tsarin gini: Compute-In-Memory (CIM). Domin magance matsalar thermodynamic, Mythic AI ta ciyar da wata falsafar tsarin gini gaba: maimakon kai-da-kawon data tsakanin sassan sarrafawa da wuraren ajiyar ƙwaƙwalwa daban-daban, ƙwaƙwalwar kanta ce ya kamata…
Dabarun modular chiplet. Ko da yake analog compute-in-memory yana warware lissafin makamashi da latency, faɗaɗa silicon a fannonin kasuwanci daban-daban yana buƙatar sassauci a packaging.
Kwatanta tsarin gini: hanyoyin compute don edge robotics. Tushe: Haɗin bayanan tattaunawar fireside da ƙayyadaddun tsarin semiconductor da aka gabatar a Economist Enterprise Physical AI and Robotics Summit (Convene 200 Aldersgate, London, 13 ga Oktoba 2026).

Fiye da rabin ƙarni, ci gaban kwamfuta ya bi dokar Moore: cusa ninki biyu na transistors a kan silicon duk bayan wata goma sha takwas zuwa ashirin da huɗu. Amma a sabon fannin physical AI da robots masu cin gashin kansu, dokar Moore ta yi karo da dokokin thermodynamics. Jirage marasa matuƙi, robots masu ƙafa huɗu, hannayen tiyata na robot da motoci masu tuƙa kansu ba za su iya ɗaukar GPUs na matakin server masu cin watt 700 ko hasumiyoyin sanyaya ba. A Economist Enterprise Physical AI and Robotics Summit a London a ranar 13 ga Oktoba 2026, Taner Ozcelik, Shugaba kuma Babban Jami'in Mythic AI, ya haɗu da Kenneth Cukier a wata tattaunawar fireside da ta duba juyin juya halin tsarin da ake buƙata domin karya wannan tarnaƙi: wuce dokar Moore ta hanyar analog compute-in-memory (CIM) da modular chiplets.

Bangon thermodynamic: dalilin da yasa physical AI ba za ta iya rayuwa a cloud ba #

A zaman fireside mai taken Moving past Moore’s law: the power behind high-performance robotics, Kenneth Cukier ya tsara babbar matsalar injiniya da ke fuskantar cin gashin kai na kamfanoni: samfuran artificial intelligence suna ƙaruwa a girman parameters da ƙarfin fahimta, amma injuna na zahiri suna aiki ƙarƙashin tsauraran ƙa'idojin SWaP-C (Size, Weight, Power, and Cost).

Iyakar AI ta yau tana da tushe a manyan hyperscale data centres da ke cin megawatt 150 ko fiye a kan manyan gungun servers. Kamar yadda Ozcelik ya nuna, wannan tsarin yana aiki a matakan makamashi da suka yi nisa da ma'aunin halittu sau miliyoyi. Ƙwaƙwalwar ɗan adam tana gudanar da daidaita gani da motsi a ainihin lokaci, tunani mai ma'ana da kewayawa ta gani da sarari cikin power envelope na jiki na kusan watt 20.

A tarihi, injiniyoyin robots sun yi ƙoƙarin kauce wa iyakokin compute da ke cikin na'ura ta hanyar tura inference zuwa cloud. Sai dai a wuraren masana'antu, sararin samaniya da tsaro masu muhimmancin gaske, cin gashin kai da ya dogara da cloud ba zai yiwu ba a tsarin gini:

Bambancin latency da jitter: Sarrafa robot mai sauri a rufaffiyar madauki yana buƙatar sabunta motoci tabbatattu a 50Hz zuwa 1,000Hz (zagaye ƙasa da millisecond 20). Hanyoyin sadarwa na cellular da Wi-Fi suna kawo packet jitter, ɓatattun frames da tsallen latency da ke sa hannayen robot su saki abin da suke riƙe ko motoci masu cin gashin kansu su kasa kaiwa iyakar birki.
Cikar bandwidth: Watsa bidiyon 4K da yawa da ba a matse ba da LiDAR point clouds daga robots masu motsi da dama a cikin masana'anta yana saurin cinye RF bandwidth na wurin kuma yana jawo kuɗin cloud egress masu tsada sosai.
Ikon gudanar da aiki da haɗin sadarwa da ake takarar sa: A haƙar ma'adinai a ƙarƙashin ƙasa, ramuka marasa GPS, wuraren gine-gine masu nisa ko wuraren soja, hanyoyin sadarwa na waje ko dai babu su ko ana toshe su da gangan. Dole ne tsarin masu cin gashin kansu su ɗauki cikakken basirar aiki kai tsaye a jikin na'urar edge.

Saka GPUs masu ƙarfi na yau da kullum a cikin robots yana kawo nasa hukuncin na zahiri: thermal throttling, manyan heat sinks da ke ƙara nauyi mara amfani, da yawan cin wuta da ke rage lokacin tashin batir daga awanni zuwa mintuna.

Ra'ayin tsarin gini: Compute-In-Memory (CIM) #

Domin magance matsalar thermodynamic, Mythic AI ta ciyar da wata falsafar tsarin gini gaba: maimakon kai-da-kawon data tsakanin sassan sarrafawa da wuraren ajiyar ƙwaƙwalwa daban-daban, ƙwaƙwalwar kanta ce ya kamata ta zama kwamfutar.

Bayanin tsarin Next-Gen In-Memory AI Compute Platform da ke nuna compute arrays na transistor guda ɗaya, dabarun modular chiplet, ninkin inganci da manyan abokan kasuwanci .class="img-fluid clearfix"

Kimiyyar lissafi ta matrix multiplication a transistor guda ɗaya #

A digital computing na gargajiya, lissafin matrix dot product yana buƙatar ɗauko dubban ƙimar weight na 8-bit ko 16-bit daga DRAM na waje ta hanyoyin bus masu high-capacitance, loda su cikin cache registers, yin multiplication a cikin digital ALUs, sannan rubuta sakamakon zuwa ƙwaƙwalwa. Fiye da kashi 90% na makamashi yana tafiya ne wajen cika da zubar da cajin layukan copper interconnect tsakanin chips.

Tsarin in-memory compute yana maye gurbin wannan motsin data na digital da kimiyyar lissafi ta analog circuits:

Shirya weights. Ana shirya weights na neural network kai tsaye cikin non-volatile memory cells (kamar flash da aka gyara ko resistive RAM) a matsayin analog conductances ($G$). Da zarar an shirya su, weights suna zama a tsaye a cikin memory array; ba a taɓa ɗauko su ko motsa su yayin inference ba.

Daidaita voltage na input. Ana mayar da input activations (kamar visual tokens daga kyamarori) zuwa matakan analog voltage ($V$) kuma a tura su ta word lines na array.

Multiplication ta dokar Ohm. Yayin da voltage na input ke wucewa ta kowane memory cell, dokar Ohm ($I = V \times G$) tana yin analog multiplication nan take, tana samar da wutar lantarki daidai da sakamakon ninka input da weight da aka ajiye.

Tarawa ta Kirchhoff's Current Law. A kan bit lines na tsaye, wutar lantarki daga dubban memory cells masu layi ɗaya tana haɗuwa ta halitta bisa Kirchhoff's Current Law ($\sum I = 0$).

Don haka flash memory array guda ɗaya yana aiwatar da cikakken matrix-vector multiplication a cikin clock cycle guda ɗaya, yana aiki a matakin kimiyyar lissafi ba tare da ɗauko ƙwaƙwalwa ko sau ɗaya ba. Daga nan ana mayar da wutar analog da aka tara zuwa ƙimar digital ta hanyar high-speed analog-to-digital converters (ADCs) a gefen array.

Dabarun modular chiplet #

Ko da yake analog compute-in-memory yana warware lissafin makamashi da latency, faɗaɗa silicon a fannonin kasuwanci daban-daban yana buƙatar sassauci a packaging. Maimakon ƙera manyan system-on-chips (SoCs) guda ɗaya - waɗanda ke fama da ƙaruwar lahani mai saurin gaske, ƙarancin wafer yield da tsarin samfura marasa sassauci - Ozcelik ya bayyana dabarun modular chiplet.

A ƙarƙashin wannan tsari, ana ƙera ƙananan in-memory compute tiles masu daidaitaccen tsari daban-daban sannan a manna su a kan interposers ta amfani da advanced packaging:

Daidaitattun high-speed die-to-die interconnects: Ta amfani da sabbin ƙa'idoji kamar Universal Chiplet Interconnect Express (UCIe), chiplets suna sadarwa ta micro-bumps masu yawan bandwidth density da latency a matakin picosecond.
Haɗuwar abubuwa daban-daban: Mai tsara tsarin zai iya haɗa in-memory chiplets guda biyu da ARM control processor mai arha don ƙaramin nauyin jirgi mara matuƙi, ko ya haɗa chiplets goma sha shida tare da high-speed PCIe interfaces don gina server accelerator card don edge data centres.
Compilation na gama-gari da rarraba ayyuka: Compiler layer guda ɗaya yana karɓar daidaitattun deep learning graphs (ONNX, PyTorch) kuma yana raba layers na samfurin kai tsaye a kan chiplet topology da ke akwai, yana inganta hanyar data domin rage sadarwa tsakanin chiplets.

Kwatanta tsarin gini: hanyoyin compute don edge robotics #

Ɓangaren tsarin gini	Digital GPU na yau da kullum (NVIDIA Jetson / Orin)	Keɓaɓɓen edge NPU (Google Coral / Hailo)	Analog Compute-In-Memory chiplet (Mythic AI)
Tsarin aiki na tushe	Digital SIMD/tensor cores masu ɗauko weights daga LPDDR5/HBM	Digital systolic arrays tare da ajiyar weights a on-chip SRAM	Analog arrays na transistor guda ɗaya masu yin matrix multiply a wurin
Nauyin motsin data	Mai yawa (cajin DRAM/SRAM bus ba tsayawa; ɓarnar makamashi 85%+)	Matsakaici (iyakancewar ƙarfin SRAM cache na ciki)	Babu motsin data na ciki; weights suna zama a ƙwaƙwalwa har abada
Power envelope na inference	30W zuwa 75W+ (yana buƙatar fanka mai aiki da manyan heat sinks)	5W zuwa 15W (sanyaya mara aiki ko kaɗan)	10W zuwa 20W (fitar da zafi mara aiki a ƙaramin wuri)
Ingancin makamashi (TOPS/W)	2 zuwa 5 TOPS/Watt (a matakin tsari)	10 zuwa 25 TOPS/Watt	50 zuwa 200+ TOPS/Watt (ingancin ka'ida na 10x-1000x)
Bill-of-materials (BOM) na na'ura	£500 zuwa £2,500+ kowane compute module	£100 zuwa £350 kowane accelerator board	Dala goma-goma kowane chiplet a samarwa mai yawa
Memory bottleneck (Von Neumann)	Mai tsanani; iyakancewar memory bandwidth (GB/s zuwa TB/s)	Matsakaici; iyakancewar ƙarfin on-chip SRAM (MBs)	Babu; bandwidth na array yayi daidai da memory matrix na ciki
Fannin amfani da aka nufa	Motoci masu cin gashin kansu masu tsada da robots na duba masana'antu	Kyamarori masu wayo da ƙananan na'urorin IoT	Robots masu taɓawa da sauri, jirage marasa matuƙi na soja da AI na mota

Tushe: Haɗin bayanan tattaunawar fireside da ƙayyadaddun tsarin semiconductor da aka gabatar a Economist Enterprise Physical AI and Robotics Summit (Convene 200 Aldersgate, London, 13 ga Oktoba 2026).

Abin da wannan ke nufi a kowane fanni #

Tabbatar da in-memory compute a kasuwanci ya dogara ne kan shiga masana'antun da a tarihi iyakokin wuta da farashi suka hana amfani da manyan foundation models.

Tsaron soja da tsarin marasa matuƙi #

A aikace-aikacen tsaro na edge, ingancin makamashi yana fassara kai tsaye zuwa tsira a fagen aiki da tsawon aikin. Ƙananan jirage marasa matuƙi (sUAS) da rovers na ƙasa na soja suna da tsauraran iyakoki saboda nauyin batir. Saka GPU na yau da kullum a jirgin leƙen asiri mara matuƙi yana rage tsawon lokacinsa a sama da kashi 30% zuwa 50% kawai don ba da wuta ga AI perception stack.

Haɗin gwiwar Mythic AI da Lockheed Martin ya mayar da hankali kan saka in-memory compute a cikin tsarin edge na soja, gungun jirage marasa matuƙi da wuraren sa ido masu nisa. Ta hanyar rage cin wutar compute daga 50W zuwa ƙasa da 10W, jirage marasa matuƙi na soja suna samun ƙarin lokacin tashi yayin da suke gudanar da multi-modal object classification, sarrafa siginar yaƙin lantarki da visual-inertial odometry a wurin, a fagagen aiki da ake takara a kai kuma babu GPS.

Wayar da kai kan yanayin mota da AI na cikin mota #

A fannin motoci, manyan masana'antun tier-one suna fuskantar matsin riba mai tsanani. Ko da yake motocin lantarki masu tsada za su iya ɗaukar farashin kwamfutocin tuƙi mai cin gashin kai masu sanyaya da ruwa na £2,000, motocin talakawa ba za su iya ba.

Haɗin gwiwar Honda Motor da Mythic AI yana magance wannan cikas na kasuwanci. Masu ƙera motoci suna buƙatar AI chips masu cancanta ga mota, da ake samarwa da yawa, masu farashin dala goma-goma maimakon ɗaruruwa, waɗanda za su iya gudanar da tsarin sa ido kan direba, blind-spot object segmentation da 360-degree surround perception ba tare da buƙatar madaukai masu rikitarwa na sanyaya da ruwa da ke ƙara farashin ƙera da sarƙaƙiyar haɗawa ba.

Edge data centres na masana'antu da mobile fulfillment #

Ko da yake an inganta in-memory compute don edge, Mythic tana aiwatar da dabarun gada da gangan ta hanyar ba da PCIe accelerator cards don edge data centres da cibiyoyin fulfillment na ɗakunan ajiya. A wuraren da ke ɗauke da ɗaruruwan autonomous mobile robots (AMRs), edge server racks masu gudanar da samfuran daidaita gungun robots suna fuskantar iyakar layin wutar lantarki da iyakar sanyaya na wurin. In-memory accelerator cards suna ba wa wurare damar ninka inference throughput na cikin gida ba tare da inganta ababen more rayuwa na wutar lantarki ba.

Juriyar supply chain da yanayin ƙasa na ƙera kayayyaki #

Muhimmin ɓangare na bayanin Ozcelik ya ta'allaka ne kan ikon supply chain. Ƙirƙira a hardware ba ta da ma'ana idan ba za a iya ƙera silicon, a haɗa shi a packaging kuma a tura shi da yawa cikin aminci ba.

Mythic AI ta kafa tsarin ƙera kayayyaki da ya bazu a ƙasashen duniya:

Ƙera silicon: Babban samar da in-memory wafers yana da tushe a Tokyo, Japan, yana amfana da shugabancin Japan da aka kafa a ƙera semiconductor mai inganci da balagaggun analog/flash process nodes.
Advanced packaging da haɗawa: Ana tura wafers zuwa Turai (musamman Jamus) don advanced chiplet dicing, multi-die substrate bonding da automated optical inspection.
Tabbatar da sabon bugu: Kamfanin yana kammala tabbatar da sabon bugun silicon ɗinsa a yanzu, yana mai da hankali kan fassara ma'aunin yield na dakin gwaje-gwaje zuwa Statements of Work (SOWs) masu maimaituwa a matakin samarwa don samar da yawa na mota da tsaro.

Wannan hanya ta yankuna da yawa tana kauce wa raunin siyasa na wuri guda, yayin da take ba wa abokan kamfanoni na Turai da Asiya supply chains masu bambancin yanayin ƙasa.

Kammalawa #

Tattaunawar fireside tsakanin Taner Ozcelik da Kenneth Cukier a Economist Enterprise Physical AI and Robotics Summit ta bayyana gaskiyar zahiri da za ta jagoranci shekaru goma masu zuwa na tsarin masu cin gashin kansu. Dokar Moore, a matsayin ma'auni mai sauƙi na ƙaruwar transistors, ba ta isa ta ba da ƙarfi ga juyin juya halin physical AI ba.

Muddin tsarin kwamfuta ya manne wa rabuwar logic da ƙwaƙwalwa irin ta Von Neumann, robotics za ta ci gaba da kasancewa a ƙarƙashin iyakokin nauyin batir, fitar da zafi da tsadar bill-of-materials. Ta hanyar komawa daga motsin data na digital zuwa analog in-memory compute arrays da modular chiplet topologies, injiniyar semiconductor tana samar da tsallen inganci da tsarin masu cin gashin kansu ke buƙata.

Sauyi daga gungun data centre masu cin 700W zuwa edge silicon na 15W shi ne ainihin ƙofar cin gashin kai na zahiri. A ƙarshe dai, robots da ke aiki a duniyar gaske dole su bi dokokin makamashi na duniyar gaske; in-memory compute shi ne tsarin da ke haɗa silicon da ilimin halittu.

Tambayoyin da Ake Yawan Yi #

Menene Von Neumann bottleneck a AI hardware? Von Neumann bottleneck yana nufin iyakar throughput da makamashi da rabuwar zahiri tsakanin central processing unit (ALU) da ƙwaƙwalwar tsari (DRAM) ke haifarwa. A ayyukan artificial intelligence, waɗanda suka ƙunshi biliyoyin matrix multiplications, processors suna kashe fiye da kashi 85% na jimillar makamashinsu da clock cycles masu yawa kawai wajen kai-da-kawon data ta copper buses, wanda ke haifar da tsananin matsalolin zafi da latency.

Ta yaya analog Compute-In-Memory (CIM) ke aiki? Compute-In-Memory yana yin ayyukan lissafi kai tsaye a cikin memory cells ba tare da motsa data ba. Ana ajiye weights na neural network a matsayin analog conductance states a cikin non-volatile memory cells. Idan aka sanya input voltages a kan array, wutar lantarki tana gudana ta cells bisa dokar Ohm ($I = V \times G$), kuma tana haɗuwa a kan bit lines bisa Kirchhoff's Current Law. Wannan yana kammala cikakken matrix-vector multiplication a mataki ɗaya na zahiri.

Me yasa in-memory compute ke da muhimmanci musamman ga robots da jirage marasa matuƙi? Robots masu motsi da jirage marasa matuƙi suna aiki ƙarƙashin tsauraran iyakokin Size, Weight, Power, and Cost (SWaP-C). Suna dogara ne da batir a cikinsu kuma ba za su iya ɗaukar manyan radiators na sanyaya ko kayan wuta masu yawan watt ba. In-memory compute yana ba da ingancin makamashi mafi girma da 10x zuwa 1,000x, yana ba robots damar gudanar da samfuran vision-language-action masu rikitarwa a cikin su a power envelope na 10W zuwa 20W yayin da suke kiyaye tsawon batir.

Menene amfanin hanyar chiplet idan aka kwatanta da chip guda ɗaya? Chips guda ɗaya suna haɗa dukkan sassan aiki a kan babban silicon die guda ɗaya. Yayin da girman die ke ƙaruwa, yawan lahani a ƙera yana ƙaruwa cikin sauri sosai, yana rage yield na samarwa kuma yana ƙara farashin kowane guda. Tsarin chiplet yana raba tsarin zuwa ƙananan silicon dies masu daidaitaccen tsari da ake ƙerawa daban-daban da yield mai yawa, kuma ana haɗa su ta high-speed micro-bumps. Wannan yana ba da damar tsari mai sassauci da aka tsara bisa buƙatun wuta da form factor na masana'antu.

Su wane ne manyan abokan kasuwanci na wannan fasaha? A yanzu dandamalin in-memory compute na Mythic AI yana da tushe a manyan haɗin gwiwar kamfanoni guda biyu: Lockheed Martin, wadda ke haɗa tsarin a cikin jirage marasa matuƙi na tsaro, dandamalin robots marasa matuƙi da wuraren tsaron iyaka don tsawaita rayuwar batir; da Honda Motor, wadda ke neman AI chips masu arha da suka cancanci mota don wayar da kan yanayin mota da fahimtar cikin mota.

Manazarta #

Economist Enterprise, (2026). Physical AI & Robotics Summit: Fifiko a aiki ta hanyar tsarin masu cin gashin kansu ⧉. Economist Enterprise.
Ozcelik, T., (2026). Wuce dokar Moore: ƙarfin da ke bayan high-performance robotics ⧉. Mythic AI.
Cukier, K., (2026). Iyakokin thermodynamic na basirar masu cin gashin kansu ⧉. The Economist.
Marković, D., Mizrahi, A., Querlioz, D., & Grollier, J., (2020). Kimiyyar lissafi don neuromorphic computing ⧉. Nature Reviews Physics, 2, 499-510.
Universal Chiplet Interconnect Express, (2025). UCIe Specification 2.0: Ƙa'idar die-to-die interconnect ⧉. UCIe Consortium.
IEEE, (2026). International Roadmap for Devices and Systems (IRDS): Bayan CMOS da Compute-in-Memory ⧉. IEEE Standards Association.

Bita ta ƙarshe 2026-10-13.

Bita ta ƙarshe 2026-10-27.

Sake buga wannan labari

Wuce dokar Moore: in-memory compute da chiplets don edge robotics — Sebastien Rousseau

A Economist Enterprise Physical AI and Robotics Summit a London, Taner Ozcelik na Mythic AI ya bayyana yadda in-memory compute ke rushe bangon ƙwaƙwalwa na kwamfuta.

An lasisin wannan labari a karkashin Creative Commons Attribution 4.0 International. Sake bugawa na bukatar nuna asalin URL na asali.

Wuce dokar Moore: in-memory compute da chiplets don edge robotics — Sebastien Rousseau

A Economist Enterprise Physical AI and Robotics Summit a London, Taner Ozcelik na Mythic AI ya bayyana yadda in-memory compute ke rushe bangon ƙwaƙwalwa na kwamfuta.

Originally published at https://sebastienrousseau.com/ha/2026-10-13-wuce-dokar-moore-in-memory-compute/ by Sebastien Rousseau.
Licensed under CC-BY-4.0.

SEBASTIEN ROUSSEAU FOUNDER · ENGINEER