Downloaded 182 times
Uploaded byRichard Kuo
5g, gpu and fpga
The document discusses trends in telecom infrastructure including digitalization, virtualization, and mobility that are increasing data volumes and requiring more advanced data processing. It highlights edge computing, graphics processing units (GPUs), and field programmable gate arrays (FPGAs) as technologies that can help meet these demands. Edge computing distributes computing power and data storage closer to end users to reduce latency and network loads. GPUs and FPGAs can provide increased computing power for tasks like machine learning and help customize network nodes. The document also examines how these technologies may support the evolution to 5G networks through features like network slicing and distributed analytics.
5g, gpu and fpga
- 1. 5G Network, GPUand FPGA Notes Sharing Richard T. Kuo
- 2. Intent • Highlight fewtrends affect telecom infrastructure architecture: • Digital Society • Big Data and Machine Learning • List some enablers for evolving demand: • Edge Computing • Graphic Process Unit(GPU) • Field Programmable Gate Array (FPGA) • Discuss some supporting architecture options: • Customize Nodes • Network Slicing 20161004 CC 4.0 SA, NC 2
- 3. 20161004 CC 4.0SA, NC 3
- 4. 20161004 CC 4.0SA, NC 4
- 5. Trends • Digitalization • Virtualization •Physical -> digital -> data driven • IoE -> data 4V (volume, variety, velocity, veracity) -> analytics, optimization, ML • Mobilization • Connectivity • Unconnected -> mobile -> mobile optimized • Increase system scope/domain/complexity/interactions • Place -> People -> Devices • Automation -> self-directed/managing/healing • Mobility • Time => anytime • space => anywhere • data => more available information/inference 20161004 CC 4.0 SA, NC 5
- 6. 20161004 CC 4.0SA, NC 6
- 7. What Should 5GBring 20161004 CC 4.0 SA, NC 7
- 8. 20161004 CC 4.0SA, NC 8
- 9. 5G Use Cases 20161004CC 4.0 SA, NC 9
- 10. 5G Main Enablers •Dynamic RAN provides a RAN that can adapts to rapid changes in user needs and the mix of the generic 5G services. • Ultra-Dense Networks, • Moving Networks, • Devices acting as temporary as access nodes, • D2D communication for both access and backhaul. • Lean System Control Plane (LSCP) provides new lean signaling/control information and allows separation of data and control information and support large variety of devices with very different capabilities. • Localized Contents and Traffic Flows allow offloading, aggregation and distribution of real-time and cached content. • Spectrum Toolbox contains a set of enablers (tools) to allow 5G systems to operate under different regulatory framework and share enablers. 20161004 CC 4.0 SA, NC 10
- 11. Data Process atEdge 20161004 CC 4.0 SA, NC 11 from: http://ubiquity.acm.org/article.cfm?id=2822875
- 12. Edge Computing • Pushesapplications, data and computing power (services) to the logical extremes of a network. • Replicates fragments of information across distributed networks. • Place customized nodes (SDR, SDN, NFV, …) closer to the client whenever it is possible. 20161004 CC 4.0 SA, NC 12
- 13. 20161004 CC 4.0SA, NC 13
- 14.
- 15. Advantages of EdgeComputing • Decrease the data volume that must be moved, the consequent traffic, and the distance the data must go, thereby reducing transmission costs, shrinking latency. • Reduces or eliminates the network bottlenecks. • Fast response times. • Reduce security risk by reducing the non-essential data travels into network core, only necessary data forward for analysis. • Deeper insights, with privacy control: Analyze sensitive data locally instead of sending it to the cloud for analysis. 20161004 CC 4.0 SA, NC 15
- 16.
- 17. KT – NetworkArchitecture Evolution 4G to 5G from http://www.netmanias.com/en/?m=attach&no=13956 20161004 CC 4.0 SA, NC 17
- 18. KT - NetworkArchitecture Evolution 4G to 5G from http://www.netmanias.com/en/?m=attach&no=13955 20161004 CC 4.0 SA, NC 18
- 19. Edge Node Architecture (needdesign improvement, this is only a place holder) 20161004 CC 4.0 SA, NC 19 from: http://ubiquity.acm.org/article.cfm?id=2822875
- 20. 20161004 CC 4.0SA, NC 20
- 21. Graphics Processing Unit(GPU) • A GPU is a computer chip that performs rapid mathematical calculations, primarily for the purpose of rendering images. • A GPU generally has a large number of slow and weak processors (lower operating frequency, lower number of registers, simpler ALU's etc.) • GPU's come strapped with lots of memory and generally have high memory bandwidth to support the hundreds of small processors that make up the GPU. • GPUs are special purpose and can compute vector math, matrix math, pixel transforms and rendering jobs about 10-100x faster than the equivalent CPU performance as all these tasks are embarrassingly parallel. 20161004 CC 4.0 SA, NC 21
- 22. CPU vs GPU fromhttp://www.electronicspecifier.com/communications/vivante-es-design-magazine-gpus-the-next-must-have 20161004 CC 4.0 SA, NC 22
- 23. Parallel Computing andStreaming The Landscape of Parallel Computing Research: A View from Berkeley The 13 application areas where OpenCL can be used 1. Dense Linear Algebra 2. Sparse Linear Algebra 3. Spectral Methods 4. N-Body Methods 5. Structured Grids 6. Unstructured Grids 7. Monte Carlo 8. Combinational Logic 9. Graph traversal 10. Dynamic Programming 11. Backtrack and Branch + Bound 12. Construct Graphical Models 13. Finite State Machine 14. … 20161004 CC 4.0 SA, NC 23
- 24. Current VPP HardwareAcceleration from https://fd.io/technology 20161004 CC 4.0 SA, NC 24Q QS1`
- 25. Field Programmable GateArrays (FPGAs) • Field Programmable Gate Arrays (FPGAs) are semiconductor devices that are based around a matrix of configurable logic blocks (CLBs) connected via programmable interconnects. • FPGAs can be reprogrammed to desired application or functionality requirements after manufacturing. 20161004 CC 4.0 SA, NC 25
- 26. FPGA Applications PartialList from: http://www.xilinx.com/training/fpga/fpga-field-programmable-gate-array.htm • Aerospace & Defense - Radiation-tolerant FPGAs along with intellectual property for image processing, waveform generation, and partial reconfiguration for SDRs. • ASIC Prototyping - ASIC prototyping with FPGAs enables fast and accurate SoC system modeling and verification of embedded software • Audio - Xilinx FPGAs and targeted design platforms enable higher degrees of flexibility, faster time-to-market, and lower overall non-recurring engineering costs (NRE) for a wide range of audio, communications, and multimedia applications. • Automotive - Automotive silicon and IP solutions for gateway and driver assistance systems, comfort, convenience, and in-vehicle infotainment. - Learn how Xilinx FPGA's enable Automotive Systems • Broadcast - Adapt to changing requirements faster and lengthen product life cycles with Broadcast Targeted Design Platforms and solutions for high- end professional broadcast systems. • Consumer Electronics - Cost-effective solutions enabling next generation, full-featured consumer applications, such as converged handsets, digital flat panel displays, information appliances, home networking, and residential set top boxes. • Data Center - Designed for high-bandwidth, low-latency servers, networking, and storage applications to bring higher value into cloud deployments. • High Performance Computing and Data Storage - Solutions for Network Attached Storage (NAS), Storage Area Network (SAN), servers, and storage appliances. • Medical - For diagnostic, monitoring, and therapy applications, the Virtex FPGA and Spartan® FPGA families can be used to meet a range of processing, display, and I/O interface requirements. • Video & Image Processing - Xilinx FPGAs and targeted design platforms enable higher degrees of flexibility, faster time-to-market, and lower overall non-recurring engineering costs (NRE) for a wide range of video and imaging applications. • Wired Communications - End-to-end solutions for the Reprogrammable Networking Linecard Packet Processing, Framer/MAC, serial backplanes, • and more 20161004 CC 4.0 SA, NC 26
- 27. GPU and FPGAConsiderations • GPU has good penetration on ML community. It needs to overcome big inertia to get people move away from GPUs CUDA. • FPGAs uses less power, but new NVIDIA chips use as few as 10-15 watts for a teraflop. • Verification of complex designs implemented on FPGA is a big challenge. In contrast testing and validating CUDA code is relatively easier. • Fast digital design, no wait to obtaining a target chip. • The design can be implemented on the FPGA and tested at once. • FPGAs are good for prototyping. Design change change can be absorbed in field. 20161004 CC 4.0 SA, NC 27
- 28. Observations and Interpretations •Add flexible Edge Computing Nodes for 5G network. • Use high computing power and customized SoC+FPGA for BTS/eNode- B, and form a cluster/cloud to share resources. • Accelerate performance with Field Programmable Gate Array. • Adapt Software Defined Radio and dynamic RAN. • Leverage GPUs for distributed data analytics and ML. • Consider Micro-HPC (High Performance Computing) data center for deployment. 20161004 CC 4.0 SA, NC 28
- 29. Supplementary Material 20161004 CC4.0 SA, NC 29
- 30. 5G 1000xMobile Data Volumes 10x-100x Connected Devices5xLower Latency 10x-100x End-user Data Rates 10x Battery Life for Low Power Devices Source: METIS Evolution Towards 2020 4G3G2G © Telefonaktiebolaget LM Ericsson 2015 | Ericsson June 2015
- 31. 5G Requirements Business Requirements •Massive broadband (xMBB) that delivers gigabytes of bandwidth on demand <- velocity • Massive machine-type communication (mMTC) that connects billions of sensors and machines <- variety + volume • Critical machine-type communication (uMTC) that allows immediate feedback with high reliability and enables for example remote control over robots and autonomous driving. <- velocity + veracity Technology Requirements • 1-10Gbps connections to end points in the field (i.e. not theoretical maximum) • 1 millisecond end-to-end round trip delay (latency) • 1000x bandwidth per unit area • 10-100x number of connected devices • (Perception of) 99.999% availability • (Perception of) 100% coverage • 90% reduction in network energy usage • Up to ten year battery life for low power, machine-type devices 20161004 CC 4.0 SA, NC 31
- 32. Comparison between 1g,2g, 3g, 4g and 5g 20161004 CC 4.0 SA, NC 32
- 33. 20161004 CC 4.0SA, NC 33
- 34. 20161004 CC 4.0SA, NC 34
- 35. 5G mMBB Requirementsand Technology Enablers • Software Defined Radio (SDR) -> multiple radio technologies. • Massive multiple-input and multiple-output (MIMO) antennas -> data throughput and capacity • Dynamic RAN 20161004 CC 4.0 SA, NC 35
- 36. 20161004 CC 4.0SA, NC 36
- 37. Logic Gates Truth TableFPGA 20161004 CC 4.0 SA, NC 37
- 38. Binary Adder from: http://cpuville.com/adder.htm TruthTable Gates for Adder 20161004 CC 4.0 SA, NC 38
- 39. FPGA Development Board MercuryFPGA From: http://www.micro-nova.com/mercury/ Terasic SoCKit Altera Cyclone V + Dual Core A9 From: http://www.cnx-software.com/2013/08/15/249-terasic-sockit- development-kit-features-altera-cyclone-v-sx-dual-core-a9-fpga-soc/ 20161004 CC 4.0 SA, NC 39
- 40. Programming FPGA 20161004 CC4.0 SA, NC 40
- 41. 20161004 CC 4.0SA, NC 41
- 42. 20161004 CC 4.0SA, NC 42
- 43. 20161004 CC 4.0SA, NC 43
- 44. 20161004 CC 4.0SA, NC 44
Editor's Notes
- #3 5G, GPU, FPGA
- #8 International Mobile Telecommunications for the year 2000 (IMT-2000) is a worldwide set of requirements for a family of standards for the 3rd generation of mobile communications. International Mobile Telecommunications-Advanced (IMT-Advanced Standard) are requirements issued by the ITU-R of the International Telecommunication Union (ITU) in 2008 for what is marketed as 4G (Or sometimes as 4,5G) mobile phone and Internet access service.
- #16 When data analysis is done at the edge of a network, that's known as "edge analytics.” Edge computing does not replace cloud computing, however. In reality, an analytic model or rules might be created in a cloud then pushed out to edge devices. Some edge devices are also incapable of doing analysis. Edge computing is also closely related to "fog computing," which also entails data processing from the edge to the cloud. (Let's also not forget the data warehouse, which is used for massive storage of data and slower analytic queries). Solution accelerators: The cornerstone of the system is the use of semantic data models to represent and visualize real-world edge devices, such as smart garbage cans or street lamps. “You can have a rule at the edge that says, when a vehicle comes along, turn the street lamp on,” Bates said. “So you might be saving power.” IoT Solution Engine: A cloud-ready, scalable system that supports stream management. Distributed Fog Intelligence: Allows business rules to be dynamically distributed across a device network and supports edge analytics as well as machine protocols. According to Plat. One, an edge agent also allows monitoring and control of devices, including configuration, software updates, and lifecycle management.
- #20 This is only a placeholder, apply SDN/NFV and uniKernal here; may combined with enodeb possible high performance micro-datacenter.
- #21 SoC FPGA Board management Management Security Protocol bridging FPGA/GPU Compute Acceleration Machine learning Video transcode Custom routines Storage Acceleration Cryptography Compression Indexing FPGA Network Processing Inline processing Pre-processing Pre-filtering Exception processing Cryptography Compression I/O expansion Protocol bridging
- #22 Gpgpu-General-purpose computing on graphics processing units From: https://www.quora.com/Whats-the-difference-between-a-CPU-and-a-GPU
- #23 http://superuser.com/questions/308771/why-are-we-still-using-cpus-instead-of-gpus TL;DR answer: GPUs have far more processor cores than CPUs, but because each GPU core runs significantly slower than a CPU core and do not have the features needed for modern operating systems, they are not appropriate for performing most of the processing in everyday computing. They are most suited to compute-intensive operations such as video processing and physics simulations. GPGPU is still a relatively new concept. GPUs were initially used for rendering graphics only; as technology advanced, the large number of cores in GPUs relative to CPUs was exploited by developing computational capabilities for GPUs so that they can process many parallel streams of data simultaneously, no matter what that data may be. While GPUs can have hundreds or even thousands of stream processors, they each run slower than a CPU core and have fewer features (even if they are Turing complete and can be programmed to run any program a CPU can run). Features missing from GPUs include interrupts and virtual memory, which are required to implement a modern operating system. In other words, CPUs and GPUs have significantly different architectures that make them better suited to different tasks. A GPU can handle large amounts of data in many streams, performing relatively simple operations on them, but is ill-suited to heavy or complex processing on a single or few streams of data. A CPU is much faster on a per-core basis (in terms of instructions per second) and can perform complex operations on a single or few streams of data more easily, but cannot efficiently handle many streams simultaneously. As a result, GPUs are not suited to handle tasks that do not significantly benefit from or cannot be parallelized, including many common consumer applications such as word processors. Furthermore, GPUs use a fundamentally different architecture; one would have to program an application specifically for a GPU for it to work, and significantly different techniques are required to program GPUs. These different techniques include new programming languages, modifications to existing languages, and new programming paradigms that are better suited to expressing a computation as a parallel operation to be performed by many stream processors. For more information on the techniques needed to program GPUs, see the Wikipedia articles on stream processing and parallel computing. Modern GPUs are capable of performing vector operations and floating-point arithmetic, with the latest cards capable of manipulating double-precision floating-point numbers. Frameworks such as CUDA and OpenCL enable programs to be written for GPUs, and the nature of GPUs make them most suited to highly parallelizable operations, such as in scientific computing, where a series of specialized GPU compute cards can be a viable replacement for a small compute cluster as inNVIDIA Tesla Personal Supercomputers. Consumers with modern GPUs who are experienced with Folding@home can use them to contribute with GPU clients, which can perform protein folding simulations at very high speeds and contribute more work to the project (be sure to read the FAQsfirst, especially those related to GPUs). GPUs can also enable better physics simulation in video games using PhysX, accelerate video encoding and decoding, and perform other compute-intensive tasks. It is these types of tasks that GPUs are most suited to performing. AMD is pioneering a processor design called the Accelerated Processing Unit (APU) which combines conventional x86 CPU cores with GPUs. This approach enables graphical performance vastly superior to motherboard-integrated graphics solutions (though no match for more expensive discrete GPUs), and allows for a compact, low-cost system with good multimedia performance without the need for a separate GPU. The latest Intel processors also offer on-chip integrated graphics, although competitive integrated GPU performance is currently limited to the few chips with Intel Iris Pro Graphics. As technology continues to advance, we will see an increasing degree of convergence of these once-separate parts. AMD envisions a future where the CPU and GPU are one, capable of seamlessly working together on the same task. Nonetheless, many tasks performed by PC operating systems and applications are still better suited to CPUs, and much work is needed to accelerate a program using a GPU. Since so much existing software use the x86 architecture, and because GPUs require different programming techniques and are missing several important features needed for operating systems, a general transition from CPU to GPU for everyday computing is very difficult.
- #32 http://blog.cnccookbook.com/2015/12/08/need-know-now-iot-internet-things-cnc-manufacturing/
- #41 Contrary to regular computer programs which are sequential, its statements are inherently concurrent (parallel). For that reason, VHDL is usually referred to as a code rather than a program. In VHDL , only statements placed inside a PROCESS, FUNCTION, or PROCEDURE are executed sequentially.
- #45 ### C-RAN From: http://www.fujitsu.com/downloads/TEL/fnc/whitepapers/CloudRANwp.pdf Operators who choose wireless architectures have several options to choose from: small cells (SC-RAN); carrier Wi-Fi (CW-F); and Distributed Antenna System (DAS). These and a host of other solutions are being introduced by network operators as methods of expanding their network to accommodate data growth. A Centralized-RAN, Cloud-RAN, or C-RAN architecture addresses capacity and coverage issues, while supporting mobile xHaul (Fronthaul and Backhaul) solutions as well as network self-optimization, self-configuration, self-adaptation with software control and management through SDN and NFV. Cloud RAN also provides great benefits in controlling ongoing operational costs, improving network security, network controllability, network agility and flexibility. Growth in data traffic also severely impacts power consumption, with consequent cost burdens. Most of the power consumption is in the radio access networks, specifically at base stations. These consume more than 80% of the total power drawn by a typical mobile network system [4]. Reducing energy cost and shrinking the carbon footprint to transform to an efficient power management paradigm are increasingly urgent imperatives, especially in combination with demand for increased capacity, better coverage, and all-time-high throughput. New and alternative techniques and architectures that favour efficient operation, low power consumption, agile traffic management, and high reliability are not just nice-to-haves; they are business essentials. Remote Radio Heads (RRHs) connects to the baseband unit (BBU) using CPRI (Common Public Radio Interface) or OBSAI (Open Base Station Architecture Initiative) interfaces. The RRHs include the radio, the associated amplification/filtering, and the antenna. The BBU is implemented separately and performs the centralized signal processing functionality of the RAN. The decentralized BBU enables agility, faster service delivery, cost savings, and improved coordination of radio capabilities across a set of RRHs. Check https://www.crowdsupply.com/lime-micro/limesdr/updates/remote-radio-head Airinterface https://www.crowdsupply.com/lime-micro/limesdr/updates/oai-lte-demo Octave https://www.crowdsupply.com/lime-micro/limesdr/updates/lab-gnu-ocatve ### V-RAN The Virtualized-RAN (V-RAN) architecture virtualizes the BBU functionality and services in a centralized BBU pool (V-BBU) in the Central Office (CO) that can effectively manage on-demand resource allocation, mobility, and interference control for a large number of interfaces using programmable software layers. V-RAN architecture enjoys software-defined capacity and scaling limits. ### CPRI (Common Public Radio Interface) CPRI (which is more widely adopted in the industry than OBSAI) is a digital interface standard for encapsulating radio samples between the RRH and the Baseband Unit (BBU). The interface burden is not packet-based; signals are multiplexed in a low-latency timeslot-like fashion. CPRI offers maximum latency, near-zero jitter, and a near-zero bit error rate. In practice, a value of 0.4 milliseconds for transport leaves an acceptable delay budget for processing requirements and propagation delay. The CPRI capacity required is up to 10 GbE, with distances of up to 40 km between the RRH and the BBU.