DOCSLIB.ORG

High-Performance Processors' Design Choices

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
High-Performance Processors' Design Choices High-Performance Processors’ Design Choices Ramon Canal PD Fall 2013 1 High-Performance Processors’ Design Choices 1 Motivation 2 Multiprocessors 3 Multithreading 4 VLIW 2 Outline • Motivation • Multiprocessors – SISD, SIMD, MIMD, and MISD – Memory organization – Communication mechanisms • Multithreading •VLIW 3 Motivation Instruction-Level Parallelism (ILP): What all we have covered so far: – simple pipelining – dynamic scheduling: scoreboarding and Tomasulo’s alg. – dynamic branch prediction – multiple-issue architectures: superscalar, VLIW – compiler techniques and software approaches Bottom line: There just aren’t enough instructions that can actually be executed in parallel! – instruction issue: limit on maximum issue count – branch prediction: imperfect – # registers: finite – functional units: limited in number – data dependencies: hard to detect dependencies via memory 4 So, What do we do? Key Idea: Increase number of running processes – multiple processes: at a given “point” in time • i.e., at the granularity of one (or a few) clock cycles • not sufficient to have multiple processes at the OS level! Two Approaches: – multiple CPU’s: each executing a distinct process • “Multiprocessors” or “Parallel Architectures” – single CPU: executing multiple processes (“threads”) • “Multi-threading” or “Thread-level parallelism” 5 Taxonomy of Parallel Architectures Flynn’s Classification: – SISD: Single instruction stream, single data stream • uniprocessor – SIMD: Single instruction stream, multiple data streams • same instruction executed by multiple processors • each has its own data memory • Ex: multimedia processors, vector architectures – MISD: Multiple instruction streams, single data stream • successive functional units operate on the same stream of data • rarely found in general-purpose commercial designs • special-purpose stream processors (digital filters etc.) – MIMD: Multiple instruction stream, multiple data stream • each processor has its own instruction and data streams • most popular form of parallel processing – single-user: high-performance for one application – multiprogrammed: running many tasks simultaneously (e.g., servers) 6 Multiprocessor: Memory Organization Centralized, shared- memory multiprocessor: – usually few processors – share single memory & bus – use large caches 7 Multiprocessor: Memory Organization Distributed-memory multiprocessor: – can support large processor counts • cost-effective way to scale memory bandwidth • works well if most accesses are to local memory node – requires interconnection network • communication between processors becomes more complicated, slower 8 Communication Mechanisms • Shared-Memory Communication – around for a long time, so well understood and standardized • memory-mapped – ease of programming when communication patterns are complex or dynamically varying – better use of bandwidth when items are small – Problem: cache coherence harder • use “Snoopy” and other protocols • Message-Passing Communication (i.e. intel’s Knight… family) – simpler hardware because keeping caches coherent is easier – communication is explicit, simpler to understand • focuses programmer attention on communication – synchronization: naturally associated with communication • fewer errors due to incorrect synchronization 9 Multiprocessor: Hybrid Organization • Use distributed-memory organization at top level • Each node itself may be a shared-memory multiprocessor (2-8 processors) 10 Multiprocessor: Hybrid Organization • Use distributed-memory organization at top level • Each node itself may be a shared-memory multiprocessor (2-8 processors) • What about Big Data? Is it a “game changer”? – Next slides based on the following works: • M. Ferdman et al. “Clearing the clouds” ASPLOS’12 • P.Lotfi-Kamran et al.‘‘Scale-OutProcessors” ISCA’12 • B. Grot et al. “Optimizing Datacenter TCO with Scale-Out Processors”, IEEE MICRO 2012 – Next couple of slides © of Prof. Babak Falsafi (EPFL) 11 Multiprocessors and Big Data PD, 2013 12 PD, 2013 13 PD, 2013 14 PD, 2013 15 PD, 2013 16 Scale-out Processors • Small LLC. Just to capture instructions. • More cores for higher throughput • “Pods” for small distance to memory PD, 2013 17 Performance • Iso server power (20MW) PD, 2013 18 Summary Multiprocessors • Need to tailor chip design to applications – Big Data applications are too big for data caches. Best solution is too eliminate them. – Big Data applications in need of coarse grain parallelism (i.e. At the request level) – Still single-thread performance is STILL important for other applications (i.e. Computation intensive) PD, 2013 19 Multithreading Threads: multiple processes that share code and data (and much of their address space) • recently, the term has come to include processes that may run on different processors and even have disjoint address spaces, as long as they share the code Multithreading: exploit thread-level parallelism within a processor – fine-grain multithreading • switch between threads on each instruction! – coarse-grain multithreading • switch to a different thread only if current thread has a costly stall – E.g., switch only on a level-2 cache miss 20 Multithreading • How can we guarantee no dependencies between instructions in a pipeline? – One way is to interleave execution of instructions from different program threads on same pipeline Interleave 4 threads, T1-T4, on non-bypassed 5-stage pipe T1: LW r1, 0(r2) T2: ADD r7, r1, r4 T3: XORI r5, r4, #12 T4: SW 0(r7), r5 T1: LW r5, 12(r1) 21 Simple Multithreaded Pipeline • Have to carry thread select down pipeline to ensure correct state bits read/written at each pipe stage 22 Multithreading Fine-grain multithreading – switch between threads on each instruction! – multiple threads executed in interleaved manner – interleaving is usually round-robin – CPU must be capable of switching threads on every cycle! • fast, frequent switches – main disadvantage: • slows down the execution of individual threads • that is, traded off latency for better throughput 23 CDC 6600 Peripheral Processors (Cray, 1965) • First multithreaded hardware • 10 “virtual” I/O processors • fixed interleave on simple pipeline • pipeline has 100ns cycle time • each processor executes one instruction every 1000ns • accumulator-based instruction set to reduce processor state 24 Denelcor HEP (Burton Smith, 1982) • First commercial machine to use hardware threading in main CPU – 120 threads per processor – 10 MHz clock rate – Up to 8 processors – precursor to Tera MTA (Multithreaded Architecture) 25 Tera MTA (Cray, 1997) • Up to 256 processors • Up to 128 active threads per processor • Processors and memory modules populate a sparse 3D torus interconnection fabric • Flat, shared main memory – No data cache – Sustains one main memory access per cycle per processor • 50W/processor @ 260MHz 26 Tera MTA (Cray) • Each processor supports 128 active hardware threads – 128 SSWs, 1024 target registers, 4096 general-purpose registers • Every cycle, one instruction from one active thread is launched into pipeline • Instruction pipeline is 21 cycles long • At best, a single thread can issue one instruction every 21 cycles – Clock rate is 260MHz, effective single thread issue rate is 260/21 = 12.4MHz 27 Multithreading Coarse-grain multithreading – switch only if current thread has a costly stall • E.g., level-2 cache miss – can accommodate slightly costlier switches – less likely to slow down an individual thread • a thread is switched “off” only when it has a costly stall – main disadvantage: • limited in ability to overcome throughput losses – shorter stalls are ignored, and there may be plenty of those • issues instructions from a single thread – every switch involves emptying and restarting the instruction pipeline 28 IBM PowerPC RS64-III (Pulsar) • Commercial coarse-grain multithreading CPU • Based on PowerPC with quad-issue in-order five stage pipeline • Each physical CPU supports two virtual CPUs • On L2 cache miss, pipeline is flushed and execution switches to second thread – short pipeline minimizes flush penalty (4 cycles), small compared to memory access latency – flush pipeline to simplify exception handling 29 Simultaneous Multithreading (SMT) Key Idea: Exploit ILP across multiple threads! – Share CPU to multiple threads – i.e., convert thread-level parallelism into more ILP – exploit following features of modern processors: • multiple functional units – modern processors typically have more functional units available than a single thread can utilize • register renaming and dynamic scheduling – multiple instructions from independent threads can co-exist and co-execute! 30 Multithreading: Illustration (a) (b) (c) (d) (a) A superscalar processor with no multithreading (b) A superscalar processor with coarse-grain multithreading (c) A superscalar processor with fine-grain multithreading (d) A superscalar processor with simultaneous multithreading (SMT) 31 From Superscalar to SMT • SMT is an out-of-order superscalar extended with hardware to support multiple executing threads 32 Simultaneous Multithreaded Processor 33 Simultaneous Multithreaded Processor • Add multiple contexts and fetch engines to wide out-of- order superscalar processor – [Tullsen, Eggers, Levy, University of Washington, 1995] • OOO instruction window already has most of the circuitry required to schedule from multiple threads • Any single thread can utilize whole machine • First examples: – Alpha 21464 (DEC/Compaq) – Pentium IV (Intel) – Power 5 (IBM) – Ultrasparc IV (Sun) 34 SMT: Design Challenges •
Load more

Recommended publications
  • 2.5 Classification of Parallel Computers
    52 // Architectures 2.5 Classification of Parallel Computers 2.5 Classification of Parallel Computers 2.5.1 Granularity In parallel computing, granularity means the amount of computation in relation to communication or synchronisation Periods of computation are typically separated from periods of communication by synchronization events. • fine level (same operations with different data) ◦ vector processors ◦ instruction level parallelism ◦ fine-grain parallelism: – Relatively small amounts of computational work are done between communication events – Low computation to communication ratio – Facilitates load balancing 53 // Architectures 2.5 Classification of Parallel Computers – Implies high communication overhead and less opportunity for per- formance enhancement – If granularity is too fine it is possible that the overhead required for communications and synchronization between tasks takes longer than the computation. • operation level (different operations simultaneously) • problem level (independent subtasks) ◦ coarse-grain parallelism: – Relatively large amounts of computational work are done between communication/synchronization events – High computation to communication ratio – Implies more opportunity for performance increase – Harder to load balance efficiently 54 // Architectures 2.5 Classification of Parallel Computers 2.5.2 Hardware: Pipelining (was used in supercomputers, e.g. Cray-1) In N elements in pipeline and for 8 element L clock cycles =) for calculation it would take L + N cycles; without pipeline L ∗ N cycles Example of good code for pipelineing: §doi =1 ,k ¤ z ( i ) =x ( i ) +y ( i ) end do ¦ 55 // Architectures 2.5 Classification of Parallel Computers Vector processors, fast vector operations (operations on arrays). Previous example good also for vector processor (vector addition) , but, e.g. recursion – hard to optimise for vector processors Example: IntelMMX – simple vector processor.
    [Show full text]
  • Data-Flow Prescheduling for Large Instruction Windows in Out-Of-Order Processors
    Data-Flow Prescheduling for Large Instruction Windows in Out-of-Order Processors Pierre Michaud, Andr´e Seznec IRISA/INRIA Campus de Beaulieu, 35042 Rennes Cedex, France {pmichaud, seznec}@irisa.fr Abstract We introduce data-flow prescheduling. Instructions are sent to the issue buffer in a predicted data-flow order instead The performance of out-of-order processors increases of the sequential order, allowing a smaller issue buffer. The with the instruction window size. In conventional proces- rationale of this proposal is to avoid using entries in the is- sors, the effective instruction window cannot be larger than sue buffer for instructions which operands are known to be the issue buffer. Determining which instructions from the yet unavailable. issue buffer can be launched to the execution units is a time- In our proposal, this reordering of instructions is accom- critical operation which complexity increases with the issue plished through an array of schedule lines. Each schedule buffer size. We propose to relieve the issue stage by reorder- line corresponds to a different depth in the data-flow graph. ing instructions before they enter the issue buffer. This study The depth of each instruction in the data-flow graph is de- introduces the general principle of data-flow prescheduling. termined, and the instruction is inserted in the correspond- Then we describe a possible implementation. Our prelim- ing schedule line. Lines are consumed by the issue buffer inary results show that data-flow prescheduling makes it sequentially. possible to enlarge the effective instruction window while Section 2 briefly describes issue buffers and discusses re- keeping the issue buffer small.
    [Show full text]
  • Pipelining 2
    Instruction-Level Parallelism Dynamic Scheduling CS448 1 Dynamic Scheduling • Static Scheduling – Compiler rearranging instructions to reduce stalls • Dynamic Scheduling – Hardware rearranges instruction stream to reduce stalls – Possible to account for dependencies that are only known at run time - e. g. memory reference – Simplifies the compiler – Code built for one pipeline in mind could also run efficiently on a different pipeline – minimizes the need for speculative slot filling - NP hard • Once a common tactic with supercomputers and mainframes but was too expensive for single-chip – Now fits on a desktop PC’s 2 1 Dynamic Scheduling • Dependencies that are close together stall the entire pipeline – DIVD F0, F2, F4 – ADDD F10, F0, F8 – SUBD F12, F8, F14 • The ADD needs the DIV to finish, so there is a stall… which also stalls the SUBD – Looong stall for DIV – But the SUBD is independent so there is no reason why we shouldn’t execute it – Or is there? • Precise Interrupts - Ignore for now • Compiler could rearrange instructions, but so could the hardware 3 Dynamic Scheduling • It would be desirable to decode instructions into the pipe in order but then let them stall individually while waiting for operands before issue to execution units. • Dynamic Scheduling – Out of Order Issue / Execution • Scoreboarding 4 2 Scoreboarding Split ID stage into two: Issue (Decode, check for Structural hazards) Read Operands (Wait until no data hazards, read operands) 5 Scoreboard Overview • Consider another example: – DIVD F0, F2, F4 – ADDD F10, F0,
    [Show full text]
  • GPU Implementation Over IPTV Software Defined Networking
    Esmeralda Hysenbelliu. Int. Journal of Engineering Research and Application www.ijera.com ISSN : 2248-9622, Vol. 7, Issue 8, ( Part -1) August 2017, pp.41-45 RESEARCH ARTICLE OPEN ACCESS GPU Implementation over IPTV Software Defined Networking Esmeralda Hysenbelliu* Information Technology Faculty, Polytechnic University of Tirana, Sheshi “Nënë Tereza”, Nr.4, Tirana, Albania Corresponding author: Esmeralda Hysenbelliu ABSTRACT One of the most important issue in IPTV Software defined Network is Bandwidth Issue and Quality of Service at the client side. Decidedly, it is required high level quality of images in low bandwidth and for this reason it is needed different transcoding standards (Compression of image as much as it is possible without destroying the quality of it) as H.264, H265, VP8 and VP9. During a test performed in SMC IPTV SDN Cloud Network, it was observed that with a server HP ProLiant DL380 g6 with two physical processors there was not possible to transcode in format H.264 more than 30 channels simultaneously because CPU’s achieved 100%. This is the reason why it was immediately needed to use Graphic Processing Units called GPU’s which offer high level images processing. After GPU superscalar processor was integrated and done functional via module NVENC of FFEMPEG Program, number of channels transcoded simultaneously was tremendous increased (more than 100 channels). The aim of this paper is to real implement GPU superscalar processors in IPTV Cloud Networks by achieving improvement of performance to more than 60%. Keywords - GPU superscalar processor, Performance Improvement, NVENC, CUDA --------------------------------------------------------------------------------------------------------------------------------------- Date of Submission: 01 -05-2017 Date of acceptance: 19-08-2017 --------------------------------------------------------------------------------------------------------------------------------------- I.
    [Show full text]
  • Jetson TX2 • NVIDIA Jetson Xavier • GPU Programming • Algorithm Mapping: • Convolutions Parallel Algorithm Execution
    GPU and multicore CPU architectures. Algorithm mapping Contributors: N. Tsapanos, I. Karakostas, I. Pitas Aristotle University of Thessaloniki, Greece Presenter: Prof. Ioannis Pitas Aristotle University of Thessaloniki [email protected] www.multidrone.eu Presentation version 1.3 GPU and multicore CPU architectures. Algorithm mapping • GPU and multicore CPU processing boards • Graphics cards • NVIDIA Jetson TX2 • NVIDIA Jetson Xavier • GPU programming • Algorithm mapping: • Convolutions Parallel algorithm execution • Graphics computing: • Highly parallelizable • Linear algebra parallelization: • Vector inner products: 푐 = 풙푇풚. • Matrix-vector multiplications 풚 = 푨풙. • Matrix multiplications: 푪 = 푨푩. Parallel algorithm execution • Convolution: 풚 = 푨풙 • CNN architectures, linear systems, signal filtering. • Correlation: 풚 = 푨풙 • template matching, tracking. • Signal transforms (DFT, DCT, Haar, etc): • Matrix vector product form: 푿 = 푾풙 • 2D transforms (matrix product form): 푿’ = 푾푿. Processing Units • Multicore (CPU): • MIMD. • Focused on latency. • Best single thread performance. • Manycore (GPU): • SIMD. • Focused on throughput. • Best for embarrassingly parallel tasks. Pascal microarchitecture https://devblogs.nvidia.com/inside-pascal/gp100_block_diagram-2/ Pascal microarchitecture https://devblogs.nvidia.com/inside-pascal/gp100_sm_diagram/ GeForce GTX 1080 • Microarchitecture: Pascal. • DRAM: 8 GB GDDR5X at 10000 MHz. • SMs: 20. • Memory bandwidth: 320 GB/s. • CUDA cores: 2560. • L2 Cache: 2048 KB. • Clock (base/boost): 1607/1733 MHz. • L1 Cache: 48 KB per SM. • GFLOPs: 8873. • Shared memory: 96 KB per SM. GPU and multicore CPU architectures. Algorithm mapping • GPU and multicore CPU processing boards • Graphics cards • NVIDIA Jetson TX2 • NVIDIA Jetson Xavier • GPU programming • Algorithm mapping: • Convolutions ARM Cortex-A57: High-End ARMv8 CPU • ARMv8 architecture • Architecture evolution that extends ARM’s applicability to all markets. • Full ARM 32-bit compatibility, streamlined 64-bit capability.
    [Show full text]
  • Computer Hardware Architecture Lecture 4
    Computer Hardware Architecture Lecture 4 Manfred Liebmann Technische Universit¨atM¨unchen Chair of Optimal Control Center for Mathematical Sciences, M17 [email protected] November 10, 2015 Manfred Liebmann November 10, 2015 Reading List • Pacheco - An Introduction to Parallel Programming (Chapter 1 - 2) { Introduction to computer hardware architecture from the parallel programming angle • Hennessy-Patterson - Computer Architecture - A Quantitative Approach { Reference book for computer hardware architecture All books are available on the Moodle platform! Computer Hardware Architecture 1 Manfred Liebmann November 10, 2015 UMA Architecture Figure 1: A uniform memory access (UMA) multicore system Access times to main memory is the same for all cores in the system! Computer Hardware Architecture 2 Manfred Liebmann November 10, 2015 NUMA Architecture Figure 2: A nonuniform memory access (UMA) multicore system Access times to main memory differs form core to core depending on the proximity of the main memory. This architecture is often used in dual and quad socket servers, due to improved memory bandwidth. Computer Hardware Architecture 3 Manfred Liebmann November 10, 2015 Cache Coherence Figure 3: A shared memory system with two cores and two caches What happens if the same data element z1 is manipulated in two different caches? The hardware enforces cache coherence, i.e. consistency between the caches. Expensive! Computer Hardware Architecture 4 Manfred Liebmann November 10, 2015 False Sharing The cache coherence protocol works on the granularity of a cache line. If two threads manipulate different element within a single cache line, the cache coherency protocol is activated to ensure consistency, even if every thread is only manipulating its own data.
    [Show full text]
  • Superscalar Fall 2020
    CS232 Superscalar Fall 2020 Superscalar What is superscalar - A superscalar processor has more than one set of functional units and executes multiple independent instructions during a clock cycle by simultaneously dispatching multiple instructions to different functional units in the processor. - You can think of a superscalar processor as there are more than one washer, dryer, and person who can fold. So, it allows more throughput. - The order of instruction execution is usually assisted by the compiler. The hardware and the compiler assure that parallel execution does not violate the intent of the program. - Example: • Ordinary pipeline: four stages (Fetch, Decode, Execute, Write back), one clock cycle per stage. Executing 6 instructions take 9 clock cycles. I0: F D E W I1: F D E W I2: F D E W I3: F D E W I4: F D E W I5: F D E W cc: 1 2 3 4 5 6 7 8 9 • 2-degree superscalar: attempts to process 2 instructions simultaneously. Executing 6 instructions take 6 clock cycles. I0: F D E W I1: F D E W I2: F D E W I3: F D E W I4: F D E W I5: F D E W cc: 1 2 3 4 5 6 Limitations of Superscalar - The above example assumes that the instructions are independent of each other. So, it’s easily to push them into the pipeline and superscalar. However, instructions are usually relevant to each other. Just like the hazards in pipeline, superscalar has limitations too. - There are several fundamental limitations the system must cope, which are true data dependency, procedural dependency, resource conflict, output dependency, and anti- dependency.
    [Show full text]
  • The Multiscalar Architecture
    THE MULTISCALAR ARCHITECTURE by MANOJ FRANKLIN A thesis submitted in partial ful®llment of the requirements for the degree of DOCTOR OF PHILOSOPHY (Computer Sciences) at the UNIVERSITY OF WISCONSIN Ð MADISON 1993 THE MULTISCALAR ARCHITECTURE Manoj Franklin Under the supervision of Associate Professor Gurindar S. Sohi at the University of Wisconsin-Madison ABSTRACT The centerpiece of this thesis is a new processing paradigm for exploiting instruction level parallelism. This paradigm, called the multiscalar paradigm, splits the program into many smaller tasks, and exploits ®ne-grain parallelism by executing multiple, possibly (control and/or data) depen- dent tasks in parallel using multiple processing elements. Splitting the instruction stream at statically determined boundaries allows the compiler to pass substantial information about the tasks to the hardware. The processing paradigm can be viewed as extensions of the superscalar and multiprocess- ing paradigms, and shares a number of properties of the sequential processing model and the data¯ow processing model. The multiscalar paradigm is easily realizable, and we describe an implementation of the multis- calar paradigm, called the multiscalar processor. The central idea here is to connect multiple sequen- tial processors, in a decoupled and decentralized manner, to achieve overall multiple issue. The mul- tiscalar processor supports speculative execution, allows arbitrary dynamic code motion (facilitated by an ef®cient hardware memory disambiguation mechanism), exploits communication localities, and does all of these with hardware that is fairly straightforward to build. Other desirable aspects of the implementation include decentralization of the critical resources, absence of wide associative searches, and absence of wide interconnection/data paths.
    [Show full text]
  • EECS 570 Lecture 5 GPU Winter 2021 Prof
    EECS 570 Lecture 5 GPU Winter 2021 Prof. Satish Narayanasamy http://www.eecs.umich.edu/courses/eecs570/ Slides developed in part by Profs. Adve, Falsafi, Martin, Roth, Nowatzyk, and Wenisch of EPFL, CMU, UPenn, U-M, UIUC. 1 EECS 570 • Slides developed in part by Profs. Adve, Falsafi, Martin, Roth, Nowatzyk, and Wenisch of EPFL, CMU, UPenn, U-M, UIUC. Discussion this week • Term project • Form teams and start to work on identifying a problem you want to work on Readings For next Monday (Lecture 6): • Michael Scott. Shared-Memory Synchronization. Morgan & Claypool Synthesis Lectures on Computer Architecture (Ch. 1, 4.0-4.3.3, 5.0-5.2.5). • Alain Kagi, Doug Burger, and Jim Goodman. Efficient Synchronization: Let Them Eat QOLB, Proc. 24th International Symposium on Computer Architecture (ISCA 24), June, 1997. For next Wednesday (Lecture 7): • Michael Scott. Shared-Memory Synchronization. Morgan & Claypool Synthesis Lectures on Computer Architecture (Ch. 8-8.3). • M. Herlihy, Wait-Free Synchronization, ACM Trans. Program. Lang. Syst. 13(1): 124-149 (1991). Today’s GPU’s “SIMT” Model CIS 501 (Martin): Vectors 5 Graphics Processing Units (GPU) • Killer app for parallelism: graphics (3D games) Tesla S870 What is Behind such an Evolution? • The GPU is specialized for compute-intensive, highly data parallel computation (exactly what graphics rendering is about) ❒ So, more transistors can be devoted to data processing rather than data caching and flow control ALU ALU Control ALU ALU CPU GPU Cache DRAM DRAM • The fast-growing video game industry
    [Show full text]
  • Computer Architecture Out-Of-Order Execution
    Computer Architecture Out-of-order Execution By Yoav Etsion With acknowledgement to Dan Tsafrir, Avi Mendelson, Lihu Rappoport, and Adi Yoaz 1 Computer Architecture 2013– Out-of-Order Execution The need for speed: Superscalar • Remember our goal: minimize CPU Time CPU Time = duration of clock cycle × CPI × IC • So far we have learned that in order to Minimize clock cycle ⇒ add more pipe stages Minimize CPI ⇒ utilize pipeline Minimize IC ⇒ change/improve the architecture • Why not make the pipeline deeper and deeper? Beyond some point, adding more pipe stages doesn’t help, because Control/data hazards increase, and become costlier • (Recall that in a pipelined CPU, CPI=1 only w/o hazards) • So what can we do next? Reduce the CPI by utilizing ILP (instruction level parallelism) We will need to duplicate HW for this purpose… 2 Computer Architecture 2013– Out-of-Order Execution A simple superscalar CPU • Duplicates the pipeline to accommodate ILP (IPC > 1) ILP=instruction-level parallelism • Note that duplicating HW in just one pipe stage doesn’t help e.g., when having 2 ALUs, the bottleneck moves to other stages IF ID EXE MEM WB • Conclusion: Getting IPC > 1 requires to fetch/decode/exe/retire >1 instruction per clock: IF ID EXE MEM WB 3 Computer Architecture 2013– Out-of-Order Execution Example: Pentium Processor • Pentium fetches & decodes 2 instructions per cycle • Before register file read, decide on pairing Can the two instructions be executed in parallel? (yes/no) u-pipe IF ID v-pipe • Pairing decision is based… On data
    [Show full text]
  • 14 Superscalar Processors
    UNIT- 5 Chapter 14 Instruction Level Parallelism and Superscalar Processors What is Superscalar? • Use multiple independent instruction pipeline. • Instruction level parallelism • Identify dependencies in program • Eliminate Unnecessary dependencies • Use additional registers & renaming of register references in original code. • Branch prediction improves efficiency What is Superscalar? • RISC-like instructions, one per word • Multiple parallel execution units • Reorders and optimizes instruction stream at run time • Branch prediction with speculative execution of one path • Loads data from memory only when needed, and tries to find the data in the caches first Why Superscalar? • In 1987, scalar instruction • Most operations are on scalar quantities • Improving these operations to get an overall improvement • High performance microprocessors. General Superscalar Organization Superscalar V/S Superpipelined • Ordinary:- • One instruction/clock cycle & can perform one pipeline stage per clock cycle. • Superpipelined:- • 2 pipeline stage per cycle • Execution time:-half a clock cycle • Double internal clock speed gets two tasks per external clock cycle • Superscalar :- • allows parallel fetch execute • Start of the program & at branch target it performs better than above said. Superscalar v Superpipeline Limitations • Instruction level parallelism • Compiler based optimisation • Hardware techniques • Limited by —True data dependency —Procedural dependency —Resource conflicts —Output dependency —Antidependency True Data Dependency • ADD r1,
    [Show full text]
  • Concepts Introduced in Chapter 3 Instruction Level Parallelism (ILP)
    Intro Compiler Techs Branch Pred Dynamic Sched Speculation Multi Issue ILP Limits MT Intro Compiler Techs Branch Pred Dynamic Sched Speculation Multi Issue ILP Limits MT Concepts Introduced in Chapter 3 Instruction Level Parallelism (ILP) introduction Pipelining is the overlapping of dierent portions of instruction compiler techniques for ILP execution. advanced branch prediction Instruction level parallelism is the parallel execution of a dynamic scheduling sequence of instructions associated with a single thread of speculation execution. multiple issue scheduling approaches to exploit ILP static (compiler) limits of ILP dynamic (hardware) multithreading Intro Compiler Techs Branch Pred Dynamic Sched Speculation Multi Issue ILP Limits MT Intro Compiler Techs Branch Pred Dynamic Sched Speculation Multi Issue ILP Limits MT Data Dependences Control Dependences An instruction is control dependent on a branch instruction if the instruction will only be executed when the branch has a specic result. An instruction that is control dependent on a branch cannot be moved before the branch so that its execution is not controlled by the branch. Instructions must be independent to be executed in parallel. instruction types of data dependences branch branch True dependences can lead to RAW hazards. name dependences instruction Antidependences can lead to WAR hazards. before transformation after transformation Output dependences can lead to WAW hazards. An instruction that is not control dependent on a branch cannot be moved after the branch so that its execution is controlled by the branch. instruction branch branch instruction before transformation after transformation Intro Compiler Techs Branch Pred Dynamic Sched Speculation Multi Issue ILP Limits MT Intro Compiler Techs Branch Pred Dynamic Sched Speculation Multi Issue ILP Limits MT Static Scheduling Latencies of FP Operations Used in This Chapter The latency here indicates the number of intervening independent instructions that are needed to avoid a stall.
    [Show full text]