1. ELEC 6200, Fall 07, Oct 29 McPherson: Vector Processors1 Vector Processors Ryan McPherson ELEC 6200 Fall 2007

2. ELEC 6200, Fall 07, Oct 29 McPherson: Vector Processors2 Overview History Description Advantages Disadvantages Applications Conclusions

3. ELEC 6200, Fall 07, Oct 29 McPherson: Vector Processors3 What is a Vector Processor? Also called an Array Processor. Runs multiple mathematical operations on multiple data elements simultaneously. Common in supercomputers of the 1970’s 80’s and 90’s. Today most CPU designs contains at least some vector processing instructions, typically referred to as SIMD. Typically operate on a few vectors elements per clock cycle in a pipeline v. SIMD which will operate on all at once.

4. ELEC 6200, Fall 07, Oct 29 McPherson: Vector Processors4 History 1962 University of Illinois Illiac IV - completed 1972 with 64 ALUs 100-150 MFlops (massively parallel computer) (1973) TI’s Advance Scientific Computer (ASC) 20-80 MFlops (1975) Cray-1 first to have vector registers instead of keeping data in memory (8 registers with 64 64-bit words in each) Cray-1 had separate pipelines for different instruction types allowing vector chaining. 80-240 MFlops

5. ELEC 6200, Fall 07, Oct 29 McPherson: Vector Processors5 How It Works Typical Vector Processor (Cell Processor) 3

6. ELEC 6200, Fall 07, Oct 29 McPherson: Vector Processors6 Cell Processor Result of a partnership between IBM, SCEI/SONY, and TOSHIBA Parallelism at all levels Thread - multicore design (8 processors) Instruction - statically scheduled and power aware Data - data parallel instructions Contains a data processor instead of a control system Statistics:  Observed clock speed: > 4 GHz  Peak performance (single precision): > 256 GFlops  Peak performance (double precision): >26 GFlops  Local storage size per SPU: 256KB  Total number of transistors: 234M

7. ELEC 6200, Fall 07, Oct 29 McPherson: Vector Processors7 How It Works (con’t) VRF is Dynamic - 128 entry 128b wide (128x1 64x2 32x4 16x8 8x16 1x128) Stores Scalar and Vector data Computes all answers, then sorts them to reduce latency. Accesses memory in blocks. Operates on low-latency SRAM 3

8. ELEC 6200, Fall 07, Oct 29 McPherson: Vector Processors8 Advantages Each result is independent of previous results - allowing deep pipelines and high clock rates. A single vector instruction performs a great deal of work - meaning less fetches and ewer branches (and in turn fewer mispredictions). Vector instructions access memory a block at a time which allows memory latency to be amortized over many elements. Vector instructions access memory with known patterns, which allows multiple memory banks to simultaneously supply operands. Less memory access = faster processing time.

9. ELEC 6200, Fall 07, Oct 29 McPherson: Vector Processors9 Disadvantages Not as fast with scalar instructions Complexity of the multi-ported VRF Difficulties implementing precise exceptions High price of on-chip vector memory systems Increased code complexity

10. ELEC 6200, Fall 07, Oct 29 McPherson: Vector Processors10 Applications Servers Home Cinema Super Computing Cluster Computing Mainframes “Astrophysicist Replaces Supercomputer With 8 PS3’s” 2

11. ELEC 6200, Fall 07, Oct 29 McPherson: Vector Processors11 References 1. Overcoming the limitations of conventional vector processors Kozyrakis, C.; Patterson, D.; Computer Architecture, 2003. Proceedings. 30th Annual International Symposium on 9-11 June 2003 Page(s):399 - 409 2. Astrophysicist Replaces Supercomputer with Eight PlayStation 3. Wired Magazine October 17, 2007. 3. A novel SIMD architecture for the Cell heterogeneous chip-multiprocessor. Hot Chips 17 (Aug 15, 2005). 4. Computer Organization and Design. Patterson, David and Hennessy, John. Chapter 9.11 p. 48-51. Morgan Kaufmann Publishers. 2005. 5. ©2006 IBM CorporationChip Multiprocessing and the Cell Broadband EngineM. Gschwind, Computing Frontiers 2006