DOCSLIB.ORG

Msc Informatics Eng

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Msc Informatics Eng Computing Systems & Performance Beyond Vector/SIMD architectures • Vector/SIMD-extended architectures are hybrid approaches MSc Informatics Eng. – mix scalar + vector operation capabilities on a single device – highly pipelined approach to reduce memory access penalty – tightly-closed access to shared memory: lower latency 2011/12 • Evolution of Vector/SIMD-extended architectures – CPU cores with wider vectors and/or SIMD cores: A.J.Proença • DSP VLIW cores with vector capabilities: Texas Instrument • PPC cores coupled with SIMD cores: Cell Broadband Engine • ARM64 cores coupled with SIMD cores: project Denver (NVidia) • future x86 hybrid cores: Intel, AMD ... Data Parallelism 2 (Cell BE, GPU, ...) – devices with no scalar processor: accelerator devices • penalty on disjoint physical memories (most slides are borrowed) • based on the GPU architecture: SIMD—>SIMT to hide memory latency • ISA-free architecture, code compiled to silica: FPGA AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 1 AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 2 Texas Instruments: Keystone DSP architecture Cell Broadband Engine (1) AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 3 AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 4 Cell Broadband Engine (2) Cell Broadband Engine (3) AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 5 AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 6 Cell Broadband Engine (4) NVidia: Project Denver • Pick a successful SoC: Tegra 3 • Replace the 32-bit ARM Cortex 9 cores by 64-bit ARM cores • Add some Fermi SIMT cores into the same chip AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 7 AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 8 What is an FPGA FPGA as a multiple configurable ISA Field-Programmable Gate Arrays (FPGA) A fabric with 1000s of simple configurable logic cells with LUTs, on-chip SRAM, configurable routing and I/O cells AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 9 AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 10 The GPU as a compute device: the G80 The CUDA programming model • Compute Unified Device Architecture • CUDA is a recent programming model, designed for – Manycore architectures – Wide SIMD parallelism – Scalability • CUDA provides: – A thread abstraction to deal with SIMD – Synchr. & data sharing between small groups of threads • CUDA programs are written in C with extensions • OpenCL inspired by CUDA, but hw & sw vendor neutral – Programming model essentially identical AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 11 AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 12 Graphical Processing Units CUDA Devices and Threads Terminology (and in NVidia) ! Threads of SIMD instructions (warps) • A compute device ! Each has its own PC (up to 48 per SIMD processor) – Is a coprocessor to the CPU or host ! Thread scheduler uses scoreboard to dispatch – Has its own DRAM (device memory)! ! No data dependencies between threads! – Runs many threads in parallel ! Threads are organized into blocks & executed in groups – Is typically a GPU but can also be another type of parallel ! processing device ! of 32 threads (thread block) • Data-parallel portions of an application are expressed as ! Blocks are organized into a grid device kernels which run on many threads - SIMT ! The thread block scheduler schedules blocks to • Differences between GPU and CPU threads SIMD processors (Streaming Multiprocessors) – GPU threads are extremely lightweight ! Within each SIMD processor: • Very little creation overhead, requires LARGE register bank ! 32 SIMD lanes (thread processors) – GPU needs 1000s of threads for full efficiency ! Wide and shallow compared to vector processors • Multi-core CPU needs only a few AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 13 Copyright © 2012, Elsevier Inc. All rights reserved. 14 © David Kirk/NVIDIA and Wen-mei W. Hwu, 2007-2009 W. and Wen-mei © David Kirk/NVIDIA ECE 498AL, University of Illinois, Urbana-Champaign CUDA basic model: Single-Program Multiple-Data (SPMD) Programming Model: SPMD + SIMT/SIMD • CUDA integrated CPU + GPU application C program • Hierarchy – Device => Grids – Serial C code executes on CPU – Grid => Blocks – Block => Warps – Parallel Kernel C code executes on GPU thread blocks – Warp => Threads • Single kernel runs on multiple blocks ! CPU Code ! (SPMD) Grid 0 • Threads within a warp are executed in a lock-step way called single- GPU Parallel Kernel instruction multiple-thread (SIMT) KernelA<<< nBlk, nTid >>>(args); . • Single instruction are executed on multiple threads (SIMD) CPU Code – Warp size defines SIMD granularity Grid 1 (32 threads) • Synchronization within a block using GPU Parallel Kernel shared memory KernelB<<< nBlk, nTid >>>(args); . Courtesy NVIDIA AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 15 AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 16 © David Kirk/NVIDIA and Wen-mei W. Hwu, 2007-2009 W. and Wen-mei © David Kirk/NVIDIA ECE 498AL, University of Illinois, Urbana-Champaign The Computational Grid: CUDA Thread Block Block IDs and Thread IDs • A kernel runs on a computational • Programmer declares (Thread) Block: grid of thread blocks – Block size 1 to 512 concurrent – Threads share global memory threads CUDA Thread Block • Each thread uses IDs to decide – Block shape 1D, 2D, or 3D what data to work on – Block dimensions in threads ! threadID 0 1 2 3 4 5 6 7 ! – Block ID: 1D or 2D ! • All threads in a Block execute the ! – Thread ID: 1D, 2D, or 3D same thread program • A thread block is a batch of … • Threads share data and synchronize float x = input[threadID]; threads that can cooperate by: while doing their share of the work float y = func(x); output[threadID] = y; – Sync their execution w/ barrier • Threads have thread id numbers … – Efficiently sharing data through a within Block low latency shared memory • Thread program uses thread id to – Two threads from two different select work and address shared data blocks cannot cooperate AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 17 AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 18 © David Kirk/NVIDIA and Wen-mei W. Hwu, 2007-2009 W. and Wen-mei © David Kirk/NVIDIA ECE 498AL, University of Illinois, Urbana-Champaign Hwu, 2007-2009 W. and Wen-mei © David Kirk/NVIDIA ECE 498AL, University of Illinois, Urbana-Champaign Parallel Memory Sharing CUDA Memory Model Overview Thread • Local Memory: per-thread (Device) Grid – Private per thread • Each thread can: Local Memory – Auto variables, register spill – R/W per-thread registers Block (0, 0) Block (1, 0) • Shared Memory: per-block – R/W per-thread local memory Block Shared Memory Shared Memory – Shared by threads of the same – R/W per-block shared memory block Registers Registers Registers Registers ! – R/W per-grid global memory ! Shared – Inter-thread communication ! ! Memory – Read only per-grid constant • Global Memory: per-application Thread (0, 0) Thread (1, 0) Thread (0, 0) Thread (1, 0) – Shared by all threads memory Grid 0 – Inter-Grid communication – Read only per-grid texture Local Local Local Local memory Memory Memory Memory Memory . Host Global • The host can R/W Memory Global Sequential Grid 1 Memory global, constant, and Constant Grids Memory in Time texture memories . Texture Memory AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 19 AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 20 © David Kirk/NVIDIA and Wen-mei W. Hwu, 2007-2009 W. and Wen-mei © David Kirk/NVIDIA ECE 498AL, University of Illinois, Urbana-Champaign Hwu, 2007-2009 W. and Wen-mei © David Kirk/NVIDIA ECE 498AL, University of Illinois, Urbana-Champaign Graphical Processing Units Hardware Implementation: NVIDIA GPU Memory Structures Memory Architecture ! Each SIMD Lane has private section of off-chip Device DRAM • Device memory (DRAM) Multiprocessor N – Slow (2~300 cycles) ! “Private memory” (Local Memory) Multiprocessor 2 ! – Local, global, constant, Multiprocessor 1 Contains stack frame, spilling registers, and private and texture memory variables Shared Memory ! Each multithreaded SIMD processor also has Registers Registers Registers • On-chip memory Instruction … Unit local memory (Shared Memory) – Fast (1 cycle) Processor 1 Processor 2 Processor M ! Shared by SIMD lanes / threads within a block – Registers, Constant Cache shared memory, ! Memory shared by SIMD processors is GPU Texture constant/texture cache Cache Memory (Global Memory) ! Device memory Host can read and write GPU memory Courtesy NVIDIA AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 21 Copyright © 2012, Elsevier Inc. All rights reserved. 22 Families in NVidia GPU NVidia GPU structure & scalability AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 23 AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 24 The NVidia Fermi architecture GT200 and Fermi SIMD processor Fermi Multithreaded SIMD Processor (Streaming Fermi Multiprocessor) Architecture AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 25 AJProença, Computer Systems & Performance, MEI, UMinho, 2011/12 26 Graphical Processing Units Fermi Architecture Innovations Fermi: Multithreading and Memory Hierarchy ! Each SIMD processor has ! Two SIMD thread schedulers, two instruction dispatch units ! 16 SIMD lanes (SIMD width=32, chime=2 cycles), 16 load-store units, 4 special function
Load more

Recommended publications
  • Gpus: the Hype, the Reality, and the Future
    Uppsala Programming for Multicore Architectures Research Center GPUs: The Hype, The Reality, and The Future David Black-Schaffer Assistant Professor, Department of Informaon Technology Uppsala University David Black-Schaffer Uppsala University / Department of Informaon Technology 25/11/2011 | 2 Today 1. The hype 2. What makes a GPU a GPU? 3. Why are GPUs scaling so well? 4. What are the problems? 5. What’s the Future? David Black-Schaffer Uppsala University / Department of Informaon Technology 25/11/2011 | 3 THE HYPE David Black-Schaffer Uppsala University / Department of Informaon Technology 25/11/2011 | 4 How Good are GPUs? 100x 3x David Black-Schaffer Uppsala University / Department of Informaon Technology 25/11/2011 | 5 Real World SoVware • Press release 10 Nov 2011: – “NVIDIA today announced that four leading applicaons… have added support for mul<ple GPU acceleraon, enabling them to cut simulaon mes from days to hours.” • GROMACS – 2-3x overall – Implicit solvers 10x, PME simulaons 1x • LAMPS – 2-8x for double precision – Up to 15x for mixed • QMCPACK – 3x 2x is AWESOME! Most research claims 5-10%. David Black-Schaffer Uppsala University / Department of Informaon Technology 25/11/2011 | 6 GPUs for Linear Algebra 5 CPUs = 75% of 1 GPU StarPU for MAGMA David Black-Schaffer Uppsala University / Department of Informaon Technology 25/11/2011 | 7 GPUs by the Numbers (Peak and TDP) 5 791 675 4 192 176 3 Intel 32nm vs. 40nm 36% smaller per transistor 2 244 250 3.0 Normalized to Intel 3960X 130 172 51 3.3 2.3 2.4 1 1.5 0.9 0 WaEs GFLOP Bandwidth
    [Show full text]
  • Accelerating Applications with Pattern-Specific Optimizations On
    Accelerating Applications with Pattern-specific Optimizations on Accelerators and Coprocessors Dissertation Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the Graduate School of The Ohio State University By Linchuan Chen, B.S., M.S. Graduate Program in Computer Science and Engineering The Ohio State University 2015 Dissertation Committee: Dr. Gagan Agrawal, Advisor Dr. P. Sadayappan Dr. Feng Qin ⃝c Copyright by Linchuan Chen 2015 Abstract Because of the bottleneck in the increase of clock frequency, multi-cores emerged as a way of improving the overall performance of CPUs. In the recent decade, many-cores begin to play a more and more important role in scientific computing. The highly cost- effective nature of many-cores makes them extremely suitable for data-intensive computa- tions. Specifically, many-cores are in the forms of GPUs (e.g., NVIDIA or AMD GPUs) and more recently, coprocessers (Intel MIC). Even though these highly parallel architec- tures offer significant amount of computation power, it is very hard to program them, and harder to fully exploit the computation power of them. Combing the power of multi-cores and many-cores, i.e., making use of the heterogeneous cores is extremely complicated. Our efforts have been made on performing optimizations to important sets of appli- cations on such parallel systems. We address this issue from the perspective of commu- nication patterns. Scientific applications can be classified based on the properties (com- munication patterns), which have been specified in the Berkeley Dwarfs many years ago. By investigating the characteristics of each class, we are able to derive efficient execution strategies, across different levels of the parallelism.
    [Show full text]
  • NVIDIA Corp NVDA (XNAS)
    Morningstar Equity Analyst Report | Report as of 14 Sep 2020 04:02, UTC | Page 1 of 14 NVIDIA Corp NVDA (XNAS) Morningstar Rating Last Price Fair Value Estimate Price/Fair Value Trailing Dividend Yield % Forward Dividend Yield % Market Cap (Bil) Industry Stewardship Q 486.58 USD 250.00 USD 1.95 0.13 0.13 300.22 Semiconductors Exemplary 11 Sep 2020 11 Sep 2020 20 Aug 2020 11 Sep 2020 11 Sep 2020 11 Sep 2020 21:37, UTC 01:27, UTC Morningstar Pillars Analyst Quantitative Important Disclosure: Economic Moat Narrow Wide The conduct of Morningstar’s analysts is governed by Code of Ethics/Code of Conduct Policy, Personal Security Trading Policy (or an equivalent of), Valuation Q Overvalued and Investment Research Policy. For information regarding conflicts of interest, please visit http://global.morningstar.com/equitydisclosures Uncertainty Very High High Financial Health — Moderate Nvidia to Buy ARM in $40 Billion Deal with Eyes Set on Data Center Source: Morningstar Equity Research Dominance; Maintain FVE Quantitative Valuation NVDA Business Strategy and Outlook could limit Nvidia’s future growth. a USA Abhinav Davuluri, CFA, Analyst, 19 August 2020 Undervalued Fairly Valued Overvalued Nvidia is the leading designer of graphics processing units Analyst Note that enhance the visual experience on computing Abhinav Davuluri, CFA, Analyst, 13 September 2020 Current 5-Yr Avg Sector Country Price/Quant Fair Value 1.67 1.43 0.77 0.83 platforms. The firm's chips are used in a variety of end On Sept. 13, Nvidia announced it would acquire ARM from Price/Earnings 89.3 37.0 21.4 20.1 markets, including high-end PCs for gaming, data centers, the SoftBank Group in a transaction valued at $40 billion.
    [Show full text]
  • Massively Parallel Computation Using Graphics Processors with Application to Optimal Experimentation in Dynamic Control
    Munich Personal RePEc Archive Massively parallel computation using graphics processors with application to optimal experimentation in dynamic control Morozov, Sergei and Mathur, Sudhanshu Morgan Stanley 10 August 2009 Online at https://mpra.ub.uni-muenchen.de/30298/ MPRA Paper No. 30298, posted 03 May 2011 17:05 UTC MASSIVELY PARALLEL COMPUTATION USING GRAPHICS PROCESSORS WITH APPLICATION TO OPTIMAL EXPERIMENTATION IN DYNAMIC CONTROL SERGEI MOROZOV AND SUDHANSHU MATHUR Abstract. The rapid growth in the performance of graphics hardware, coupled with re- cent improvements in its programmability has lead to its adoption in many non-graphics applications, including a wide variety of scientific computing fields. At the same time, a number of important dynamic optimal policy problems in economics are athirst of computing power to help overcome dual curses of complexity and dimensionality. We investigate if computational economics may benefit from new tools on a case study of imperfect information dynamic programming problem with learning and experimenta- tion trade-off, that is, a choice between controlling the policy target and learning system parameters. Specifically, we use a model of active learning and control of a linear au- toregression with the unknown slope that appeared in a variety of macroeconomic policy and other contexts. The endogeneity of posterior beliefs makes the problem difficult in that the value function need not be convex and the policy function need not be con- tinuous. This complication makes the problem a suitable target for massively-parallel computation using graphics processors (GPUs). Our findings are cautiously optimistic in that new tools let us easily achieve a factor of 15 performance gain relative to an implementation targeting single-core processors.
    [Show full text]
  • Parallelization Schemes & GPU Acceleration
    Parallelization schemes & GPU Acceleration Erik Lindahl / Szilárd Páll GROMACS USA workshop September 13, 2013 Outline ● MD algorithmic overview ● Motivation: why parallelize/accelerate MD? ● Acceleration on CPUs ● Parallelization schemes ● Heterogeneous/GPU acceleration GROMACS USA workshop 2 Important technical terms ● Hardware: – CPU core: physical, logical, "Bulldozer" module – Node: compute machine (!= mdrun “node”) – Accelerators: GPUs, Intel MIC ● Algorithmic/parallelization: – Accelerated code, compute kernels: CPU SIMD: SSE, AVX; GPU SIMT: CUDA – multi-threading: OpenMP (thread-MPI) – SPMD, MPMD: MPI GROMACS USA workshop 3 Molecular dynamics: algorithm overview Neighbor search/DD step: every 10-50 iterations MD step NS, DD Bonded F Non-bonded F PME Integration Constraints ~ milliseconds GROMACS USA workshop 4 Why parallelize/accelerate? ● Need for speed: – MD is computationally demanding – but a fast* simulation needs: → short time-step → strong scaling Neighbor search/DD step: every 10-50 iterations MD step NS, DD Bonded F Non-bonded F PME Integration Constraints Goal: making it as short as possible currently at peak: ~100s of microsec. GROMACS USA workshop 5 Motivation: hardware evolution ● Hardware is increasingly: ● parallel on multiple levels ● heterogenous: accelerators ` + x2 x100-1000s SIMD & SIMT multicore, NUMA network: topology, memory & cache accelerator:GROMACS PCI-E, USA topologyworkshop bandwidth, latency 6 Motivation: need to keep up! ● Multiple levels of harware parallelism → need to address each level with suitable
    [Show full text]
  • The GPU Computing Revolution
    The GPU Computing Revolution From Multi-Core CPUs to Many-Core Graphics Processors A Knowledge Transfer Report from the London Mathematical Society and Knowledge Transfer Network for Industrial Mathematics By Simon McIntosh-Smith Copyright © 2011 by Simon McIntosh-Smith Front cover image credits: Top left: Umberto Shtanzman / Shutterstock.com Top right: godrick / Shutterstock.com Bottom left: Double Negative Visual Effects Bottom right: University of Bristol Background: Serg64 / Shutterstock.com THE GPU COMPUTING REVOLUTION From Multi-Core CPUs To Many-Core Graphics Processors By Simon McIntosh-Smith Contents Page Executive Summary 3 From Multi-Core to Many-Core: Background and Development 4 Success Stories 7 GPUs in Depth 11 Current Challenges 18 Next Steps 19 Appendix 1: Active Researchers and Practitioner Groups 21 Appendix 2: Software Applications Available on GPUs 23 References 24 September 2011 A Knowledge Transfer Report from the London Mathematical Society and the Knowledge Transfer Network for Industrial Mathematics Edited by Robert Leese and Tom Melham London Mathematical Society, De Morgan House, 57–58 Russell Square, London WC1B 4HS KTN for Industrial Mathematics, Surrey Technology Centre, Surrey Research Park, Guildford GU2 7YG 2 THE GPU COMPUTING REVOLUTION From Multi-Core CPUs To Many-Core Graphics Processors AUTHOR Simon McIntosh-Smith is head of the Microelectronics Research Group at the Univer- sity of Bristol and chair of the Many-Core and Reconfigurable Supercomputing Conference (MRSC), Europe’s largest conference dedicated to the use of massively parallel computer architectures. Prior to joining the university he spent fifteen years in industry where he designed massively parallel hardware and software at companies such as Inmos, STMicro- electronics and Pixelfusion, before co-founding ClearSpeed as Vice-President of Architec- ture and Applications.
    [Show full text]
  • Understanding Throughput- Oriented Architectures
    Understanding Throughput- Oriented Architectures Cedric Nugteren GPU Mini Symposium 25th of January, 2011 Contents • Motivation • GPUs in supercomputers • Performance per Watt • Example applications • Throughput-oriented architectures • Programming GPUs • Current and future architectures • Summary / GPU Mini Symposium 25-01-2011 PAGE 2 GPUs in supercomputers Taken from: Top500 Supercomputer Websites Found at ‘www.top500.org’, November 2010 / GPU Mini Symposium 25-01-2011 PAGE 3 Performance per Watt Taken from: Bill Dally In ‘GPU Computing: To Exascale and Beyond’, 2010 / GPU Mini Symposium 25-01-2011 PAGE 4 Performance per Watt NVIDIA GeForce 10000 ] GTX480 1000 GTX285 GTS250 GFLOPS GT220 100 G210M 10 Performance [ Performance 1 0,1 1 10 100 1000 Power [W] / GPU Mini Symposium 25-01-2011 PAGE 5 Performance per Watt NVIDIA GeForce AMD Radeon CPU 10000 ] HD5870 HD5770 GTX480 1000 GTX285 GTS250 GFLOPS GT220 100 Core i7-960 HD5430 G210M 10 Performance [ Performance Atom 330 1 0,1 1 10 100 1000 Power [W] / GPU Mini Symposium 25-01-2011 PAGE 6 Performance per Watt NVIDIA GeForce AMD Radeon PowerVR CPU 10000 ] HD5870 HD5770 GTX480 1000 GTX285 GTS250 GFLOPS GT220 100 Core i7-960 HD5430 G210M 543MP8 10 Performance [ Performance Atom 330 1 545 0,1 1 10 100 1000 Power [W] / GPU Mini Symposium 25-01-2011 PAGE 7 Example applications / GPU Mini Symposium 25-01-2011 PAGE 8 Contents • Motivation • Throughput-oriented architectures • Example: The NVIDIA Fermi GPU • Hardware multithreading • Many simple processing units • SIMD execution • Programming GPUs
    [Show full text]
  • Modern GPU Architectures
    4/22/12 Agenda/GPU Decoder Ring • Fermi / GF100 / GeForce GTX 480 – “Fermi Refined” / GF110 / GeForce GTX 580 – “LiPle Fermi” / GF104 / GeForce GTX 460 Modern GPU Architectures • Cypress / Evergreen / RV870 / Radeon HD 5870 – Cayman / Northern Islands / Radeon HD 6970 • TahiV / Southern Islands / GCN / Radeon HD 7970 Varun Sampath • Kepler / GK104 / GeForce GTX 680 University of Pennsylvania • Future CIS 565 - Spring 2012 – Project Denver – Heterogeneous System Architecture From G80/GT200 to Fermi Unified Address Space • GPU Compute becomes a driver for innovaon – Unified address space – Control flow advancements – ArithmeVc performance – Atomics performance – Caching – ECC (is this seriously a graphics card?) – Concurrent kernel execuVon & fast context switching • PTX 2.0 ISA supports 64-bit virtual addressing (40-bit in Fermi) • CUDA 4.0+: Address space shared with CPU • Advantages? Image from NVIDIA 1 4/22/12 Unified Address Space Control Flow Advancements cudaMemcpy(d_buf, h_buf, sizeof(h_buf), • Predicated instrucVons cudaMemcpyDefault) • RunVme manages where buffers live – avoid branching stalls (no branch predictor) • Enables copies between different devices (not • Indirect funcVon calls: only GPUs) via DMA call{.uni} fptr, flist; – Called GPUDirect • What does this enable support for? – Useful for HPC clusters • Pointers for global and shared memory are equivalent Control Flow Advancements ArithmeVc • Predicated instrucVons • Improved support for IEEE 754-2008 floang – avoid branching stalls (no branch predictor) point standards • Indirect
    [Show full text]
  • Trends in Heterogeneous Systems Architectures (And How They'll Affect Parallel Programming Models)
    Trends in Heterogeneous Systems Architectures (and how they'll affect parallel programming models) Simon McIntosh-Smith [email protected] Head of Microelectronics Research University of Bristol, UK 1 ! "Moore’s Law today Average Moore’s Law = 2x/2yrs 20-30B transistors 2x/3yrs 6-7B transistors 2x/2yrs ~2B transistors High-performance MPU, e.g. Intel Nehalem Cost-performance MPU, e.g. Nvidia Tegra http://www.itrs.net/Links/2011ITRS/2011Chapters/2011ExecSum.pdf 2 ! "Herb Sutter’s new outlook http://herbsutter.com/welcome-to-the-jungle/ “In the twilight of Moore’s Law, the transitions to multicore processors, GPU computing, and HaaS cloud computing are not separate trends, but aspects of a single trend – mainstream computers from desktops to ‘smartphones’ are being permanently transformed into heterogeneous supercomputer clusters. Henceforth, a single compute-intensive application will need to harness different kinds of cores, in immense numbers, to get its job done.” “The free lunch is over. Now welcome to the hardware jungle.” 3 ! "Four causes of heterogeneity •" Multiple types of programmable core •" CPU (lightweight, heavyweight) •" GPU •" Others (accelerators, …) •" Interconnect asymmetry •" Memory hierarchies •" Software (OS, middleware, tools, …) 4 ! "Heterogeneous Systems AMD Llano Fusion APUs Intel MIC FP7 Mont Blanc ARM + GPU NVIDIA Tegra, Project Denver 5 ! "Heterogeneity is mainstream Quad-core ARM Cortex A9 CPU Dual-core ARM 1.4GHz, ARMv7s CPU Quad-core SGX543MP4+ Imagination GPU Triple-core SGX554MP4 Imagination GPU Most
    [Show full text]
  • 3-1-11: Master Deck- Please Save to Your Desktop
    The Evolution of Modern Parallel Computing Sanford H. Russell Director of CUDA Marketing, NVIDIA Corporation Taipei | May 19 , 2011 GPU Computing Is not CPU vs. GPU It is CPU + GPU 2 GPU Computing Application Code Rest of Sequential Only Critical Functions CPU Code GPU Parallelize using CUDA CPU Programming Model + 3 GPU Computing Milestones 2002 2007 2007 GPGPU G80 NVIDIA First Parallel C Compiler Programming on Computing Architecture top of OGL (Single Precision) C SDK for GPU 4 GPU Computing Milestones 2009 2010 1st True HPC Class GPU Industry Standard IDE Fermi Architecture Parallel Nsight for DP, ECC and C++ support Microsoft Visual Studio 5 Tesla GPUs Power 3 of Top 5 Supercomputers #1 : Tianhe-1A #3 : Nebulae #4 : Tsubame 2.0 7168 Tesla GPU’s 2.5 PFLOPS 4650 Tesla GPU’s 1.2 PFLOPS 4224 Tesla GPU’s 1.194 PFLOPS 6 Tesla in 3 of Top 5 Supercomputers Performance 2500 2000 1500 Gigaflops 1000 500 0 Tianhe-1A Jaguar Nebulae Tsubame Hopper II 7 Tesla Best Performance/Watt 2500 8 Power 7 2000 6 1500 5 4 Gigaflops 1000 3 Megawatts 2 500 1 0 0 Tianhe-1A Jaguar Nebulae Tsubame Hopper II 8 World’s Fastest HPC Processor Tesla M2090: The 512 Core Fermi 512 CUDA Cores 665 GFlops 178 GB/s memory B/W 9 Industry and Research Partners Oil and gas Edu/Research Government Life Sciences Finance Manufacturing Reverse Time Astrophysics Signal Processing Bio-chemistry Risk Analytics Structural Migration Molecular Satellite Imaging Bio-informatics Monte Carlo Mechanics Kirchoff Time Dynamics Video Analytics Material Science Options Pricing Computational
    [Show full text]
  • Low Overhead Dynamic Binary Translation for ARM
    Low Overhead Dynamic Binary Translation for ARM A THESIS SUBMITTED TO THE UNIVERSITY OF MANCHESTER FOR THE DEGREE OF DOCTOR OF PHILOSOPHY IN THE FACULTY OF SCIENCE AND ENGINEERING. 2016 Bernard Amanieu d’Antras School of Computer Science 2 Contents Abstract 11 Declaration 12 Copyright 13 Acknowledgments 14 1 Introduction 15 1.1 Binary translation . 16 1.2 Contributions . 19 2 Dynamic Binary Translation 23 2.1 Code caches . 24 2.2 Multi-threading . 27 2.3 Environment . 28 2.4 Transparency . 30 2.5 MAMBO-X64 . 32 2.5.1 Binary translator . 34 2.5.2 System emulator . 36 2.6 Summary . 39 3 3 Optimizing indirect branches in dynamic binary translators 40 3.1 Hardware-assisted function returns . 43 3.1.1 Software return address stack . 44 3.1.2 Hardware return address prediction . 45 3.1.3 Return address stack elision . 46 3.1.4 Overflow and underflow handling . 50 3.1.5 Misprediction handling . 51 3.1.6 Unlinking . 54 3.2 Branch table inference . 55 3.2.1 Detecting branch tables . 56 3.2.2 Translating branch tables . 58 3.3 Fast atomic hash tables . 60 3.3.1 Hash table operations . 60 3.3.2 SPC and TPC packing . 64 3.4 Evaluation . 65 3.4.1 Experimental setup . 65 3.4.2 MAMBO-X64 . 67 3.4.3 Hardware-assisted function returns . 68 3.4.4 Branch table inference . 72 3.4.5 Fast atomic hash tables . 72 3.5 Related work . 76 3.5.1 Indirect branch handling . 76 3.5.2 Function return handling .
    [Show full text]
  • ACORN RISC MACHINE Jitendra Marathe ARM Is a Reduced
    SRJIS / BIMONTHLY/JITENDRA MARATHE. (581-585) ACORN RISC MACHINE Jitendra Marathe ARM is a reduced instruction set computer (RISC) instruction set architecture (ISA) developed by ARM Holdings. It was named the Advanced RISC Machine and, before that, the Acorn RISC Machine. The ARM architecture is the most widely used 32-bit instruction set architecture in numbers produced. Originally conceived by Acorn Computers for use in its personal computers, the first ARM-based products were the co-processor modules for the BBC series of computers. Features and applications: In 2005 about 98% of the more than one billion mobile phones sold each year used at least one ARM processor. As of 2009 ARM processors accounted for approximately 90% of all embedded 32-bit RISC processors and were used extensively in consumer electronics, including personal digital assistants (PDAs), tablets, mobile phones, digital media and music players, hand-held game consoles, calculators and computer peripherals such as hard drives and routers. Licensees: The ARM architecture is licensable. Companies that are current or former ARM licensees include Alcatel-Lucent, Apple Inc., AppliedMicro, Atmel, Broadcom, Cirrus Logic, CSR_plc, Digital Equipment Corporation, Ember, Energy Micro, Freescale, Intel (through DEC), LG, Marvell Technology Group, Microsemi, Microsoft, NEC, Nintendo, Nuvoton, Nvidia, Sony, NXP (formerly Philips Semiconductor), Oki, ON Semiconductor, Psion, Qualcomm, Renesas, Samsung, Sharp, Silicon Labs, STMicroelectronics, Symbios Logic, Texas Instruments, VLSI Technology, Yamaha, Fuzhou Rockchip, and ZiiLABS. In addition to the abstract architecture, ARM offers several microprocessor core designs, including the ARM7, ARM9, ARM11, Cortex- A8, Cortex-A9, and Cortex-A15. Companies often license these designs from ARM to manufacture and integrate into their own system on a chip (SoC) with other components like RAM, GPUs, or radio basebands (for mobile phones).
    [Show full text]