Embedded Systems Architecture. Computer Architectures

Transcription

1 Embedded Systems Architecture Computer Architectures M. Eng. Mariusz Rudnicki 1/18

2 A taxonomy of computer architectures There are many different types of architectures, and it is worth considering some way to classify them. Flynn s taxonomy is a famous taxonomy of various architectures - it is well known, though it is rough and not precise. Name Description Instruction streams Data streams SISD Single Instruction Single Data 1 1 SIMD Single Instruction Multiplie Data 1 Multiple MISD Multiplie Instruction Single Data Multiple 1 MIMD Multiplie Instruction Multiplie Data Multiple Multiple 2/18

3 Architectures taxonomy Flynn s taxonomy two basic concepts: parallelism in instruction streams; parallelism in data streams. A n C system has n program counters, so there are n instruction streams that can be executed in parallel. A data stream can be thought of as a sequence of data. In a stream, each data is processed in the sequence it belongs to. There can be multiple independent streams. Data and instruction streams are orthogonal, and there exist 4 possible, different combinations: 3/18

4 Flynn s taxonomy SISD SISD CU Control Unit PROGRAM C RAM Instructions Processing Unit Data Memory Module Single Instruction Single Data classical uniprocessor architecture instructions are executed one at a time on a single data stream: variables in the program being. This view is fairly inaccurate (consider the operation of a single pipeline); currently correct only for simple processors, such as Intel Intel 8051 family processors have no pipeline and cache, they issue and execute instructions in order. Classical Von Neumann machine. 4/18

5 Flynn s taxonomy MISD MISD PROGRAM CU1 CU2 CUn Control Units Instructions C1 C2 RAM Data Cn Processing Units Memory Module Multiple Instruction Single Data this class have no hardware implementations. Many authors think, that it doesn t make any sense. However, same argue that considering modern processors an instance of SISD architectures is less precise than ascribing them to the MISD class: data to be processed flow from one instruction to the next as the instructions stream within the stage of the pipeline. 5/18

6 SIMD PROGRAM CU Control Unit C1 RAM1 C2 RAM2 Instructions Processing Units Data Memory Modules Cn RAMn Single Instruction Multiple Data this model was adopted in one of the first models of parallel architectures ever proposed Illiac IV. Illiac IV was designed in mid 60 ; was built in 1972; had computational power of same 50 MFLOPS (initial target was 1 GFLOP). Illiac IV is the most famous case of an architectural model currently almost disappeared, known as array processor: huge number of identical processors that execute the same sequence of instructions on different data streams. 6/18

7 GENERAL RPOSE PROCESSOR ARRAY PROCESSOR CU PROGRAM I/O System RAM Block diagram of Array Processor 7/18

8 SIMD PROGRAM CU C1 C2 RAM1 RAM2 Instructions Processing Units Data Memory Modules Control Unit Cn RAMn Vector processors conceptually very similar to array processors. This is a second type of SIMD architecture. Vector processors work on array of data, single processor is in charge of carrying out the operation on all elements in the array. CRAY models, DCD Cyber 205, Fujitsu VPX 2000, Hitachi S3600 adopted this architectures. The SIMD paradigm has been used in the microarchitecture of general purpose processors called multimedia extensions. But the most notable exploitation of this paradigm is in graphic processing domain. 8/18

9 SIMD PROGRAM CU Control Unit Instructions C1 RAM1 C2 RAM2 Processing Units Data Memory Modules Cn RAMn In Pentium processors the MMX Multi Media Extensions and SSE Streaming SIMD Extensions technologies were implemented for the first time. MMX technology fixed-point operations on 64-bit vectors. SSE technology floating point operations on vectors containing four 32-bit floatingpoint numbers. 9/18

10 Nvidia GeForce GTX TITAN contains 2688 CUDA cores; CUDA Compute Unified Device Architecture: 5 GPCs Graphics Processor Clusters; 14 SMx Streaming Multiprocessors; Stream Processors; 10/18

11 Nvidia GeForce KEPLER 11/18

12 Intel Broadwell G 12/18

13 Intel Broadwell G 13/18

14 Intel Broadwell G 14/18

15 Intel Broadwell G 15/18

16 Flynn s taxonomy MIMD MIMD PROGRAM CU1 C1 RAM1 CU2 C2 RAM2 CU1 Instructions Data CUn Cn RAMn Control Units Processing Units Memory Modules MIMD all multiprocessors and multicomputers architectures ( from dual core processors, to small UMA systems, up to clusters) belong to this class. It shows that Flynn s taxonomy is really coarse and we can meet architectures that do not fit well to this scheme: multithreaded processors: SISD or MIMD? general purpose processors with SIMD extensions: where do they belong? 16/18

17 Extended Flynn s taxonomy Tanenbaum Tanenbaum Fig /18

18 References 1. Andrew S. Tanenbaum - Structured computer organization 5 TH Edition, Prentice Hall, David Padus Encyclopedia of Parallel Computing, Springer US, 2011, p kground/parallelhardware.htm 5. h.pdf 6. MF%20part%20II.pdf Tanenbaum Fig /18