PerfView 1
An Performance View For Heterogeneous Computing System - Perfview Project
Heterogeneous Computing and Heterogeneous System
Heterogeneous Computing
Heterogeneous computing is a type of computing that involves the use of different types of processors or specialized hardware components to perform different types of tasks. This approach is used to improve performance and efficiency, as each component can be optimized for specific tasks.
Some common examples of heterogeneous computing include combining central processing units (CPUs) with graphics processing units (GPUs) to accelerate graphics and video processing, or using field-programmable gate arrays (FPGAs) for specialized tasks such as data encryption.
Heterogeneous computing is becoming increasingly important in modern computing systems, as it allows for more efficient use of resources and improved performance for specific tasks. It also enables the use of specialized hardware for emerging technologies such as artificial intelligence and machine learning.
Heterogeneous System
A heterogeneous system refers to a computer system that uses a combination of different types of processors or hardware components to perform various tasks. These components may have different architectures, instruction sets, and capabilities.
Examples of heterogeneous systems include supercomputers that combine CPUs with GPUs or custom accelerators, mobile devices that combine CPUs with specialized chips for graphics or image processing, and servers that use a mix of CPUs and FPGAs for high-performance computing.
Heterogeneous systems offer several advantages over homogeneous systems, including improved performance, energy efficiency, and cost-effectiveness. However, they also present challenges in terms of software development and system integration due to the need to manage different architectures and programming models.
Why Heterogeneous?
Every computer science student should be familiar with the importance of the GPU. Its significance extends beyond its widespread use in machine learning and deep learning. Furthermore, its innovative design presents a fresh approach to challenging the limitations of Moore’s Law.
Dave Patterson, a prominent computer architecture researcher, delivered a lecture several years ago. During his presentation, he highlighted the increasing importance of data movement in modern applications. Additionally, he emphasized that different applications often possess unique characteristics, leading to certain computation tasks dominating the overall execution time in specific scenarios. And then there comes DSL(Domain Specific Language) and DSA(Domain Specific Application)
Talk link: (20) “A New Golden Age for Computer Architecture” with Dave Patterson
Why performance view?
In order to attain highly specific optimization for these applications, mere hardware development is insufficient. It is evident that programs running on x86 or ARM architecture often fail to fully utilize the capabilities of new hardware. For instance, GPUs possess CUDA cores and tensor cores for optimized linear algebra, which are lacking in most x86 chips.
We must remain focused on our primary goal of developing a new architecture: to achieve accelerated computations across a wide range of tasks. Additionally, it is essential to evaluate our usage and identify any bottlenecks that may be present in our new hardware. A subsequent presentation provided a comprehensive overview of the integration of hardware design and software design.
ASPLOS Keynote: The Golden Age of Compiler Design in an Era of HW/SW Co-design by Dr. Chris Lattner