Design method for high efficiency super computing system based on task data flow drive

Abstract

The invention relates to a design method for high efficiency super computing system based on task data flow drive; the method uses reconfigurable internet to interconnect n processing units, memory array and input / output interfaces, each functional parts are reconfigurable and programmable according to task data flow, to realize coordination operation of the whole system, the task data flow is stored in instruction pool of the memory array in instruction mode, the processing unit may take out the instruction to be executed via the reconfigurable internet to store in a cache, then type analysisstatistics is executed, the processing unit executes the computation of structure reconstruction according to the result of the type analysis statistics; the invention provides a design method for high efficiency super computing system based on task data flow drive, the method may realize self-adaptive computation through the processing unit may configure programming, programmable internet, and other programmable software and hardware, so as to realize optimum performance of a computing system.

Description

(1) Technical field: [0001] The invention relates to a design method of a high-efficiency supercomputing system, in particular to a design method of a high-efficiency supercomputing system driven by task data flow. (two), background technology: [0002] At present, there are two main features of the top supercomputer: first, the information flow inside the computer is driven by instructions, and the order of instruction execution is determined by the instruction counter (PC); second, the application of the computer is mainly for numerical calculation and data processing. To make high-performance computers have super sustained and stable computing power and achieve a high cost-effectiveness ratio, that is, to make high-performance computers not only have unparalleled peak computing speed in cutting-edge scientific computing, but also have stable and efficient continuous computing performance. It is well known that using the instruction stream mechanism, it is difficult to ach...

AI Technical Summary

Problems solved by technology

The common way to improve the computing power of a single processing node is to increase the clock frequency or increase the number of cores inside a single processor. Calculation speed is a problem, but it still cannot solve the problem of high computing efficiency

On the other hand, linear expansion of the number of processors cannot achieve a linear growth acceleration ratio. On the contrary, increasingly severe bottlenecks such as storage walls, scale walls, high energy consumption, high cost, and low utilization have restricted the development of high-performance computers.

[0004] It is difficult to improve the cost-effectiveness of the mechanism based on traditional command control data information flow. Can we get rid of this high-performance computer research and development idea and study a high-performance computing system to truly achieve high equivalent peak computing speed and high continuous processing speed? 1. The "double high and one low" high-efficiency computing system with low computing energy consumption is the main goal of the present invention

Images