基于GPU与神经网络加速器的 异构计算 兼顾算力效率与算法普适性的平台 时昕 2020 July World Leading Technologies in GPU, AI, Wireless Connectivity IP and More
Total Page:16
File Type:pdf, Size:1020Kb
基于GPU与神经网络加速器的 异构计算 兼顾算力效率与算法普适性的平台 时昕 2020 July World leading technologies in GPU, AI, Wireless Connectivity IP and more An independent worldwide provider of strategic silicon >11 Billion ~38% Intellectual Property Mobile GPU IP market share >900 employees worldwide – 80% engineers Cumulative chip shipment with Imagination IPs An original IP portfolio with a significant, long-present & long-term, patent portfolio underpinning it ~43% Domain expertise in GPU, AI, CPU & Connectivity $108M Automotive GPU IP market 2018 Revenue share Targeting the fastest growing market segments including Mobile, Automotive, AIoT, Compute, Gaming, Consumer Thousands #2 Customers include Apple, MediaTek, Rockchip, Fundamental patents and the In Wireless Connectivity IP UNISOC, TI, Renesas, Socionext, and more only non-US core GPU patents holder 2 Global Team 3 Imagination The best solution for embedded graphics, AI, compute and connectivity PowerVR GPU PowerVR Ensigma Broad suite of products covering Vision and AI Connectivity and broadcast embedded graphics needs across all communications markets Dedicated AI and vision hardware High performance, low power AXE/XM/XT GPU IP Ray Tracing AI Compute Software Neural Network SDK RF Accelerators Scalable cores with Wi-Fi, Bluetooth Architecture for best PPA Tools and libraries for Ethernet Packet advanced modelling deep learning Wi-Fi, Bluetooth + Safety Critical Hardware NNA Processor of light framework integration Automotive Cores Digital Radio 4 Emerging technologies PowerVR Ray Tracing IMG Edge New IP projects PowerVR Ray Tracing is a Performance and datapath Imagination has a number of revolutionary 3D graphics analysis, design verification new IP projects for 2020, technology that mimics how and validation remain highly available for discussion light behaves in the real complex and challenging under NDA and lead partner world to create visuals with parts of any SoC project. engagement now. astonishing realism, while IMG Edge solves this also enabling developers and challenge, enabling content creators to simplify customers to save 10+ their workflow. millions of dollars and months of project time. 5 Imagination’s investment in AI compute Researching in AI for 7 years Silicon Proven IP ~400 Man years of development Over 80 patents filed/granted in AI 神经网络生态 Deep Learning Neural Networks at the EDge AI Revenue growth Edge inference: The big opportunity Tech team of the year 2018 Design team of the year 2018 Best IP Product 2018 Confidential, Subject to NDA, Imagination Technologies 2020 9 Benchmark leading AI (Quant) Top benchmark quant performance on ai-benchmark.com CPU Q CPU F QUANT QUANT FP16 FP16 FP32 FP PAR Processor CPU Cores AI Accelerator Year Accuracy AI-Score AI Score AI Score Score Accuracy Score Accuracy Score Score 4x2 GHz Cortex-A75 & PowerVR AX2185 & Unisoc Tiger T710 2019 1400 2113 8255 96 15603 79 394 314 85 280971,5 4x1.8 GHz Cortex-A55 PowerVR GM9446 GPU 1x2.96 + 3x2.42 + 4x1.8 DSP (Hex. 690) + GPU Snapdragon 855 Plus 2019 2213 4132 6375 40 8960 30 1357 1078 34 24553 1,4 GHz Kryo 485 (Adreno 640) 2x2.27 GHz Cortex-A76 & HiSilicon Kirin 810 NPU / n.a. 2019 1085 2488 1695 80 17635 95 463 348 90 23944 1 6x1.88 GHz A55 1x2.84 + 3x2.41 + 4x1.78 DSP (Hex. 690) + GPU Snapdragon 855 2018 2106 3431 5535 55 7342 37 1156 714 43 20554 1 GHz Kryo 485 (Adreno 640) 2x2.2 GHz Cortex-A75 & DSP x 2 + APU / PowerVR Mediatek Helio P90 2018 1067 2038 6566 98 10120 94 140 69 96 20050 1 6x2 GHz Cortex-A55 GM9446 Exynos 9825 Octa n.a. GPU / n.a. + NPU x 2 2019 1591 2956 2077 60 9170 45 780 747 50 17514 1,5 2x2.6 GHz + 2x1.92 GHz HiSilicon Kirin 980 NPU x 2 / n.a. 2018 1817 3447 222 60 10750 85 139 64 76 16684 2 A76 & 4x1.8 GHz A55 2x2.7 GHz M4 & 2x2.3 GHz GPU (Mali-G76 MP12) + Exynos 9820 Octa 2018 1491 2645 1656 60 8367 45 744 701 50 15713 1,5 A75 & 4x2 GHz A55 NPU x 2 4x2.8 GHz Kryo 385/G & DSP (Hex. 685) + GPU Snapdragon 845 2018 1605 2172 2031 58 6327 38 916 683 45 13894 1 4x1.7 GHz Kryo 385/S (Adreno 630) 2x2.2 GHz Kryo 470/G & DSP (Hex. 688) + GPU Snapdragon 730 2019 1066 2082 2765 58 3581 37 450 374 44 10453 1 6x1.8 GHz Kryo 470/S (Adreno 618) More details can be found at: http://ai- Confidential, Subject to NDA, Imagination Technologies 2020 benchmark.com/news_2019_04_18_spreadtrum_ud710.html 10 灵活高效普适的异构平台 NX-F及AI Synergy NX-F Neural Compute Solution for the Edge Programmable Extensibility • High efficiency neural network support • Series3NX-F introduces programmability and floating point support • Compute optimised GPGPU • PowerVR Rogue architecture • Built on years of GPU compute experience • Leveraging the software tools and compute API investment Series3NX cores: Use case examples PERFORMANCE 0.6 TOPS 1.2 TOPS 2.4 TOPS 5 TOPS 10 TOPS 20 TOPS 40 TOPS 80 TOPS 160 TOPS TOPS AX3125 AX3145 AX3365 AX3385 AX3595 UH2X40 UH4X40 UH8X40 UH16X40 IoT Smart camera Wearables Speech rec TOSpPeeSch NLP Smartphone Home Assistant AR/VR ADAS/AV NNA: Series3NX highlights Faster Improved PPA Features • Up to • Up to • Up to 70% 40% •35% more inf/s 2 better inf/s/mm • lower bandwidth Lossless Variable Advanced weight security Bit-depth compression enablement Introducing Imagination’s Series3NX family Matching a range of required performance for vision and audio task Series3NX Series3NX-F Series3NX-MC Series3NX-A • Ultimate performance • Enables pre- • Multicore designs for • Dedicated density for AI processing and the most demanding automotive cores workloads custom layer support AI market • ASIL certification • Single core designs • Built upon years of requirements packages • Provides 256 to parallel compute • Delivering 8,192 to • Dedicated long term 4,096 MAC/clock expertise 65,536 MAC/clk support • Up to 10 TOPS • Programmable with • Up to 160 TOPS performance OpenCL Confidential, Subject to NDA, Imagination Technologies 2020 15 AI Synergy gpu gpu sram nna graphics AI AI Programmable AI & Graphics powerhouse Series3NX and Imagination GPUs Imagination GPUs are fully programable, highly efficient compute processors Supports the highest efficiency for graphics and compute applications Industry standard Khronos APIs available, including OpenGLES, OpenCL, and Vulkan Imagination GPUs extend the capability and flexibility of Series3NX when combined together Allows a complete Imagination AI solution Single source of tools and dedicated AI APIs between both cores Including common debug and profiling tools Single shared API – IMGDNN enables developers to easily port their applications from GPU to NNA with minimal code change Imagination GPU Hyperlane technology for dedicated AI resource allocation Enables developers to be sure the minimum performance levels are achieved Confidential, Subject to NDA, Imagination Technologies 2020 January 2020 17 PowerVR AI: One combined API Your network, one API, optimised for GPU and NNA IMG DNN API PowerVR GPU PowerVR NNA Cost Lowering effort to support a wide range of networks Performance Zero copy memory sharing between IP, driver-level synchronisation Ease of use Common tools across GPU and NNA Use Case - Automotive Cluster (GPU) IVI / Navigation (GPU) Surround view (GPU) Object Detection (GPU/NNA) Mirror Replacement (GPU) Heads Up Display (GPU) Driving Monitoring (GPU/NNA) Collision Detection (GPU/NNA) Autonomous Driving (GPU/NNA) Thank You.