Simplif lifyin ing A AI for or C Com ommunic icatio ions, Radar, a , and W Wirele less Systems John D D. Fergu guson info@deep eepwaved edigi gital.com
Deep L Learn rning a and R Radio F Frequency (RF) ) Systems Deep eep L Learning i g is Emer erging Radio F Freq equen ency T Technology i is Pervasive Cyber Medicine Auton onom omy Inte ternet UAV Wi Wireless C Control Radar ar Satellite lite Electronic W Warfare Wireless N Networki king Communicat ations • Intrusion • Tumor • Pedestrian / • Image c on D Detection on or D Detection on / obst stac acle cla lassif ificatio tion • Threat • Medical • Speech r at c clas assi sificat ation al d data a a anal alysis detection on recognitio ition • Facia • Diag • Navig • Languag ial r recognitio ition agnosis igatio tion age t translat ation Navig igatio tion RF A Abla latio tion Milita litary • Imag • Drug d • Street s • Document / agery a anal alysis dis iscovery sign r read ading / d datab abas ase Internet o of Th Things (Medical) l) Communicat ations • Speech r recognitio ition sear arching Tele lecommunic icatio tions Enabl bled b d by low-cost, h hig ighly capable le g gen eneral l ma Gamma ible pur urpose g se graphi hics processi essing u uni nits ( s (GPUs) s) Visib X-Ray ay UV UV Infr frared Radio F Frequencies pm pm nm nm nm 1 mm 1 mm km 1 p 10 p 10 n 400 n 700 n 100 k Deep l p learni ning ng techno nology e enabl bled a ed and d accel eler erated ed b by GP GPU U pr processo essors s - Has yet t to impa pact de design a and nd appl pplicatio ions in wir ireless and r nd radio dio frequ quenc ncy s systems 2
Where to o Use se Deep eep Lea earnin ing in RF RF Systems Transm smit Recei eive e Frequ equenc ncy Frequ equenc ncy Conv nvert Conv nvert • Spectrum um m moni nitoring ng ( (thr hreats) Spectru rum / / Ne Network rk • Intel elligen ent s spec ectrum usage • Electronic p protection ( (anti-jam) m) Centric A Appl pplications • Cognitive s e system em c control • Advanced m modulati tion t techniques Devi evice / e / Basesta tation • Adapti Modu Mo dula late Demodula ulate 1 0 1 1 0 0 1 tive wa waveforms 1 0 1 1 0 0 1 Centric A Appl pplications • Encrypti tion a and s security ty • Voic oice / / imag age r recognit itio ion User A r App • Multi User A r Apps User A r Apps ti-sensor or f fusion ion Centric A Appl pplications • Decision m making g and d data ta r reduction 3
Deep L Learn rning C Compari rison Image and Vi Video Audio a and L Language Systems a and S Sign gnals • Multiple channels (RGB) • Single channel • Multiple channels • x, y spatial dependence • Frequency, phase, amplitude • Frequency, phase, amplitude • Temporal dependence (video) • Temporal dependence • Temporal dependence • Complex data (I/Q) • Large Bandwidths • Human engineered Existing ng d deep l learni ning p poten entially a ada daptabl ble t to systems s and s nd signa nals • Must t contend wi with wi wideb eband s sign gnals and complex d data t types 4
Hardware for D Deep L Learning i in RF Systems Training Infer eren ence Pro ros Cons Pro ros Cons • Slower than GPU • Adaptable architecture • Low parallelism • Supported by ML Frameworks CPU • Fewer software • Software programmable • Limited real-time bandwidth • Lower power consumption • Medium latency • Medium power requirements architectures • Supported by ML Frameworks • Overall power • Adaptable architecture • Medium power requirements • Widely utilized consumption GPU • High real-time bandwidth • Not well integrated into RF • Highly parallel / adaptable • Requires highly parallel • Software programmable • Higher latency • Good throughput vs power algorithms • High power efficiency • Long development / upgrades Not widely utilized, not well suited FPGA • High real-time bandwidth • Limited reprogrammability (yet) • Low latency • Requires special expertise • Extremely power efficient • Extremely expensive • High real-time bandwidth • Long development time ASIC Not widely utilized, not well suited • Highly reliable • No reprogrammability • Low latency • Requires special expertise 5
Critical al Performan ance P Parameters f for D Deep Lear arning i in RF Systems Ada daptabilit ity / / Deplo ployment Real al Ti Time Compute / e / Upg pgradabil bilit ity Time me Lifecy cycle cle C Cost Band ndwidt idth Watt Latenc ncy CPU PU GPU PU FPGA FPG ASIC GPU PU s signa nal pr l processing ng can pr n provide w wide deba band d capa pability and nd software upg upgradabil ilit ity a at lower cost and de nd develo lopment t time - Mus ust conten end w with i inc ncrea eased ed l laten ency ( (~2 m 2 microsec second nd) 6
Outl tline • Introduction to Deep Learning in RF • Deepwave’s Technology • Signal Detection and Classification • Real-time Benchmarks on Embedded GPUs • Summary
Why Has Deep Learn rning i in RF N Not B Been Ad Addressed Bandwidth Limite ted C Compute te Complica cated Li Limitati tions Resources Softw tware remote processing not possible at field site for RF and AI independently • AI requires large data sets • No RF systems exist with • Disjointed software integrated AI computational • Insufficient bandwidth to • Difficult to program and processors send to remote data center understand 8
Deepwave’s Software Defined R Radio A Pl Platform f for a a Mul ultit itude of A Appl pplications ns The P Platform orm The S he Soft ftware Complete Edge-compute AI Platform for RF Simply build AI into wireless technology Artific icia ial I l Intell llig igence R Radio T Transceiv iver* WAVELEARNER Software Deep / Machine GPU Parallel Learning Computing 3 rd Party AI Signal Software Frameworks Processing Hardware Abstraction AIR-T Hardware *Patent P Pendi ding ng
Arti tifi fici cial Intelligence Radio T Transceiver (AIR-T) T) Ha Hardware S e Spec pecifications AI AIR-T • 2x2 MIMO T Transcei eiver er • Analog Devices 9371 chip Mini I i ITX F X For orm F Fac actor • Tunable from 300 MHz to 6 GHz • 100 MHz bandwidth per channel • Digital S Signa nal / l / Deep L Learning ning Processors • Xilinx Artix 7 FPGA • NVIDIA Jetson TX2 • ARM Cortex-A57 (quad-core) • Denver2 (dual core) • Nvidia Pascal 256 Core GPU • Shared GPU/CPU memory • Conne nectiv ivity • 1 PPS / 10 MHz for GPS Synchronization • External LO input • HDMI, USB 2.0/3.0, SATA, Ethernet, SD Card, GPIO 13
Arti tifi fici cial Intelligence Radio T Transceiver (AIR-T) T) Block D k Diag agram am 2x2 M MIMO 1 G GigE gE NVIDIA Jetson SAT ATA CPU HDMI MI RF Transceiver FPGA GPU/CPU Arm A57 (4 Core) JESD SD PCIe Ie Analog Devices 9371 Xilinx Artix 7 Shared USB 3 B 3.0 GPU Memory GPI PIO NVIDIA Parker (256 Core) GPI PIO Clock Incorp orpor oration of of GPU PU i in RF system a allows f for or wideband REF EF TIME pr processi essing o of signal da data i in n software en environment - Reduces es dev evel elopmen ent t time e and c cost
Simplified d Programming Deep eep TensorRT Te Learning or or or or VHDL, Verilog Cust stom S Software Dig igit ital l Signal al Proces essing 14
FFT T Perfor ormance e Testing FPGA GA PCI CIe Sha hared ed M Mem emory Compl plex int16 t 16 to Intel 7500U complex f float32 32 Sha hared ed M Mem emory cuFFT uFFT NVIDA DA Te Tegra TX2 X2 13
Infer eren ence e at the E Edge e with GR GR-Wavelearner TensorRT Optimi mizer Runtime me Train N Neural N Net etwork Optimize N e Neural N Net etwork Deploy A oy Application on 14
GR GR-Wavel elea earner er Software • Goal i l is to he help lp the he o ope pen s source c communit ity e easily ly de deplo loy deep l learning wi ng within s signal processing a g applications • Wel ell docu cumented R README w with th d dep ependency cy i installati tion instruction ons t to g get et started ed q quick ckly • Ubuntu 16.04 recommended, Windows 10 supported • NVIDIA Docker Container 18.08* • Signal al c clas assifier er e exam ample p e provided ed: • GNU Radio Flowgraph • Python source code • PLAN files that are executable on the AIR-T and Maxwell • Signal data file example for testing • Support f for Te TensorRT 5.0 • Avai ailab able a e at: d deep epwaved edigital al.com om/wavel elear earner er https://docs.nvidia.com/deeplearning/sdk/tensorrt-container-release-notes/rel_18.08.html 15
GN GNU R U Radio – Sof oftware D Defin ined R Radio io ( (SDR) F ) Framework • Popular o open s source s software d defined radio io ( (SDR) t ) toolkit it: • RF Hardware optional • Can run full software simulations • Python A API PI • C++ under the hood • Easily ly c create D DSP a algorit ithms • Custom user blocks • Pri rimarily ly us uses C CPU • Advanced parallel instructions • Recent development: RFNoC for FPGA processing • De Deepwave i is integr grating G g GPU s support f for bot oth D DSP a and M ML 16
Recommend
More recommend
