Marching Memory マーチングメモリ. UCAS-6 6 > Stanford > Imperial > Verify 中村維男 Based on Patent Application by Tadao Nakamura and Michael J.

Transcription

1 UCAS-6 6 > Stanford > Imperial > Verify 2011 Marching Memory マーチングメモリ Tadao Nakamura 中村維男 Based on Patent Application by Tadao Nakamura and Michael J. Flynn 1 Copyright 2010 Tadao Nakamura

2 C-M-C Computer Communication Memory CPU 2 Copyright 2010 Tadao Nakamura

3 HPC Today and beyond! 1 EFlops Ideal This Jump is realized by the Marching Memory Processing Speed 100 PFlops 10 PFlops 1 PFlops 100 TFlops Maximum of HPC Today s 10 TFlops SX Mainframe Today s Applications 3 Copyright 2010 Tadao Nakamura

4 Computer System Tunable Energy Efficiency η 1, η 2,, η 9 at Levels 1-9 Level 9: Algorithms for Applications Level 8: High Level Languages Level 7: Programs Level Level 6: Operating Systems Level 5: Compilers Level 4: Instruction Set Architectures Level 3: Microarchitectures Level 2: Logic Level 1: Circuits Level 0: Devices COOL Software Software Domain Instruction Set Architecture (Old Definition of Architecture) Hardware Domain COOL Chips Architecture Given Energy Efficiency η 0 at Level 0 Energy Efficiency at All the Levels for COOL Systems 4 Copyright 2010 Tadao Nakamura

5 Energy Consumption Power Dynamic Static Device level energy consumption Time Power Net energy consumed Overhead energy consumed Power Real energy consumption Net energy consumed Time Ideal energy consumption Time 5 Copyright 2010 Tadao Nakamura

6 Difference Among Shift Registers, Delay Line Memory, CCD and Marching Memory Examples of streaming memory are shift registers, delay line memory and CCDs. However, these differ from marching memory in the content, and scale s of streaming memory. Shift registers: Bit manipulation in a least unit, word of memory, performing multiplication and division in the binary system and serial parallel and parallel- serial conversions Delay line Memory: Feedbacked delay lines are used to keep one bit in it. CCD: transmitting analog signalized electric charge to the output after photo-electric conversion on the image receptor. Marching memory: DRAM type memory with storage and marching functions of information / data. 6 Copyright 2010 Tadao Nakamura

7 Concept of Conventional Memory Data accessed by addressing with several procedures and complicated hardware and lots of wires Von Neumann bottleneck Conventional memory C P U Memory access time is 1000 times the clock cycle of the CPU The memory bottleneck as a memory wall 7 Copyright 2010 Tadao Nakamura

8 Concept of Marching Memory Data marches, column by column, to the DRAM pins / edge No long addressing / sense lines. Marching memory Information / Data marching C P U 1 memory unit marching time = CPU s clock cycle Any portion available to be off for energy saving, which means that t all the cells / memory units are not always active through the control. c 8 Copyright 2010 Tadao Nakamura

9 Computer Organization Alternatives Cache Memory CPU Marching Memory memory CPU CPU (a) Computer organizations Increasing bandwidth The memory bottleneck as a memory wall (b) Bandwidths 9 Copyright 2010 Tadao Nakamura

10 Organization Consisted of Marching Memory and ALUs / Arithmetic Pipelines. Conventional main memory Conventional cache Conventional Register file Marching memory appears as a huge DRAM vector register file ALU / Arithmetic pipeline 10 Copyright 2010 Tadao Nakamura

11 Kinds of Marching Memory 1. Simple Marching Memory for vector data and streaming data Extended Simple Marching Memory for random access with high locality of data 2. Complex Marching Memory for random access 11 Copyright 2010 Tadao Nakamura

12 Classification of Marching Memory Definition: The abbreviation of Marching Memory is MM 1) Simple MM Sequential Access Mode As a core of a complex MM <<<<< As an independent MM unit for streaming/vector data only As an independent MM unit for sequential programs only 2) Extended simple MM Random Access Mode with High Locality As a core of a complex MM <<<<< As an independent MM unit for data with high locality As an independent MM unit for programs without long distance branches 3) Complex MM Random Access Mode with Low Locality As an independent MM unit for huge programs with long distance branches As an independent MM unit for general types of data 12 Copyright 2010 Tadao Nakamura

13 Structure of a Simple Marching Memory From the previous chip To the arithmetic units Almost no long wires on a chip because of minimum addressing overhead, no precharge,, no activation Information / Data marching Any portion available to be off for energy saving, which means that all the cells / memory units are not always active through the control. 13 Copyright 2010 Tadao Nakamura

14 Circuit for a simple marching memory for vector and streaming data Space Clock AND gate with delay Memory Write Memory Read Capacitor Cell 1 Cell 2 Cell 3 Cell 4 time=t time = t+1 (a) Memory scheme Logical value Cell 5 Cell position & Time 1 0 Time t t+τ Memory shift Memory stay (a) Clock 14 Copyright 2010 Tadao Nakamura

15 One stage of Marching Memory consisting of an AND gate and one capacitor Vague part! Space Clock AND gate with delay Memory Write Capacitor Cell 1 Cell 2 time=t time = t+1 Cell position & Time 15 Copyright 2010 Tadao Nakamura

16 How to implement Marching Memories 1. Reliable implementation by 2-dimensional 2 arrays of flip-flops flops 2. A definite way of using memory cells that have one normal AND gate consisting of several CMOS FETs,, and one capacitor 3. An expected way of using memory cells that have one (special) AND gate function MOS FET and one capacitor 16 Copyright 2010 Tadao Nakamura

17 Extended simple marching memory Information shifting (a) Shifting (marching) right (b) Staying Information shifting (c) Shifting (marching) left 17 Copyright 2010 Tadao Nakamura

18 Circuit for extended simple marching memory Clock Selector Space Program shifted left Clock not issued then Program Stayed AND gate Program with delay shifted right AND gate with delay Memory Read/Write Memory Read/Write Capacitor Cell 1 Cell 2 Cell 3 Cell 4 Cell 5 time=t time = t+1 Cell position & Time 18 Copyright 2010 Tadao Nakamura

19 Accessing data in marching memory requires marching to the correct column Position indexes (tags) relative in marching information / data, that are created dynamically with variables by a counter Contents 19 Copyright 2010 Tadao Nakamura

20 Operation of marching memories in action Instruction position index I v I s (a) For instructions b a Scalar / vector instruction Position index for data of the scalar instruction Contents t s r (b) For scalar data q p o Starting point position index for vector data and streaming data (c) For vector data and streaming data 20 Copyright 2010 Tadao Nakamura

21 Pragmatics of marching memories in general uses Position used Contents Instruction from/to the next column Scalar data from/to the next column Position not used Vector data from the next column Instructions marching (a) For instructions Scalar data marching (b) For scalar data Vector data marching (c) For vector data Instruction to the CPU,, and from/to the next column Scalar data to ALU, and from/to the next column Position used at the start point of a vector data Vector data to a pipeline 21 Copyright 2010 Tadao Nakamura

22 Comparison of MM with conventional DRAM in a memory cycle in terms of hardware quantity Sense amplifiers Memory space Column decoder A wordline Conventional DRAM Row decoder Bitline 1 or Bitline 2 or Bitline 3 or Bitline 4 One stage of MM Marching memory MM 22 Copyright 2010 Tadao Nakamura

23 Pipelining in MM but not so serious in energy consumption due to shorter processing time and if any, lower clock frequency even still higher Addressing Data sense Column access Data restore Bitline precharge time Pipelining in MM storage process time 23 Copyright 2010 Tadao Nakamura

24 Comparison of MM with conventional memory in a typical read cycle Addressing Data sense Possible? And besides refresh! Column access Data restore Memory access of ordinal DRAM Bitline precharge time Data access / One stage march in Marching Memory Memory access of Marching memory time 24 Copyright 2010 Tadao Nakamura

25 (Extended) Simple marching memory speed 1 memory unit time in existing memory Memory unit in existing memory s s speed 100 memory unit time in marching memory Memory units in marching memory s s speed The other 99 memory units of marching memory could be available in 100 memory units. 25 Copyright 2010 Tadao Nakamura

26 Scalar data for threads in a process, needs compiler support Memory unit in existing memory Time multiplex thread data Memory units in marching memory 26 Copyright 2010 Tadao Nakamura

27 Complex marching memory 1. Sometimes it is required to access a remote column 2. Use multiple (extended) simple marching memory cores with interconnection to a single simple marching memory to interface to arithmetic 3. Provides random access facility among multiple (extended) simple marching memory cores 27 Copyright 2010 Tadao Nakamura

28 Comparing conventional DRAM Device to complex (multi-core) marching memory Row decoder Bit line Word line Sense amplifiers 256 Kbit marching memory core which is either a simple or extended marching memory 256 Mbit DRAM Device / 256 Mbit Marching memory consisting of 1000 marching memory cores Column decoder 28 Copyright 2010 Tadao Nakamura

29 Structure of conventional DRAM architecture 256 Kbit-core DRAM device Rank Rank Dual in-line memory module 29 Copyright 2010 Tadao Nakamura

30 Structure of conventional DRAM core 256 Kbit-core Bitline pitch is 90 nm 90 nm*512 bitlines = 46 um Wordline Wordline pitch 135 nm 135 nm*512 wordlines = 69 um DRAM device Bitline Core 30 Copyright 2010 Tadao Nakamura

31 Structure of complex (multi-core) marching memory 256 Kbit-simple simple/extended marching memory core pins Complex marching memory Rank Rank Dual in-line marching memory module 31 Copyright 2010 Tadao Nakamura

32 Complex Marching Memory Core -Its Basics- 256 Kbit simple / extended marching memory core that has 1000 marching columns, each of which consists of 256 bits (32 Bytes). Contents Position Index/Tag (For example, The token of the column that means the first address of the column bytes) 32 Copyright 2010 Tadao Nakamura

33 Interconnecting 1000 simple MM cores in complex marching memory One solution : Simple marching memory cores are interconnected to a single simple marching memory as the interface between the multiple cores and the arithmetic units Data is transferred column by column c between cores. Simple marching memory contents transfer interconnect Arithmetic units Simple marching memory Multi-core controller based upon compiler analysis 256 Kbit (extended) simple marching memory core 33 Copyright 2010 Tadao Nakamura

34 Features of Marching Memory Row decoder Wordlines Column decoder Bitlines Sense Amplifiers And then Disappeared! None of addressing, sensing, restore, precharge and refresh exist MM s Downsizing with much less hardware No long wires even among complex MMs, Simple structure of MM systems total As a result High speed Large capacity Low energy consumption Low cost 34 Copyright 2010 Tadao Nakamura

35 Conclusions We have presented two types of marching memory (MM) using DRAM type technology 1. Simple MM is fast and suitable for streaming applications 1-1. Extended Simple MM for random access with high locality of data 2. Complex MM for random access in a program uses multiple (extended) simple MM cores This is slower and requires multi-core interconnection networks. More research is needed. MM has potential for faster memory technology especially with good compiler support. 35 Copyright 2010 Tadao Nakamura