The new generation of heterogeneous systems enables artificial intelligence to unleash greater potential

Translated from eecatalog

Abstract: This article introduces the obvious advantages and convenience of heterogeneous computing for language programming. When AI or high computing power is required, heterogeneous computing can further improve CPU efficiency. Therefore, heterogeneous computing is an indispensable combat capability for the future development of AI.

A little science - What is heterogeneous computing?

The computing power is ultimately provided by the underlying chip. According to the composition of computing chips, they can be divided into:

Homogeneous computing: A computing method that uses computing units of the same type of instruction set and architecture to form a system. Heterogeneous computing
: A computing method that uses computing units of different types of instruction sets and architectures to form a system. Common computing unit categories include CPU, GPU, ASIC, FPGA, etc.

Engineers have invented a computing method that combines computing units with different types of instruction sets and architectures, such as CPU+GPU. The CPU is used to control the system, decompose tasks, and schedule, while the GPU easily has thousands of computing units, each of which can perform calculations in parallel at the same time. Coupled with powerful floating-point and vector computing capabilities, it can be used to complete heavier computing tasks. This is heterogeneous computing.

Open computing platforms can allow artificial intelligence to skip the training phase and bring benefits to the automotive, medical, industrial and other fields.

Artificial intelligence (AI) has three elements: algorithms, computing power, and data.

Today's AI applications will touch every aspect of our lives, including transportation, finance, retail, healthcare, smart manufacturing, education, and the service industry. AI technology will be at the forefront of digitally connected cars, smart manufacturing, and medical image recognition. The question we have to ask ourselves is, how can we harness the power of AI and take advantage of today's diverse systems and protocols? The answer lies in an emerging ecosystem that is designed to combine today's many different "computing capabilities."

Injecting abstract ideas into heterogeneous platforms

As heterogeneous processors become widely available, new platforms are expected to take advantage of the super computing power, including acceleration units (GPUs, DSPs, and FPGAs). Artificial intelligence, machine learning, and neural networks are at the forefront of this new computing paradigm. In addition, new architectures are needed to handle the large-scale computing power enhanced by CPU cluster-based computers. It is a challenge to transplant this approach into the mainstream, mainly because heterogeneous programming models have not been standardized and lack portability.

Heterogeneous Systems Architecture (HSA) Alliance ^[1]

The challenge facing many industries is that existing architectures cannot meet today's artificial intelligence and big data workloads. The open computing platform of heterogeneous system architecture (HSA) provides a viable solution. This new architecture will lead to a whole new opportunity, and the biggest beneficiaries are autonomous driving, the need for higher computing power, and powerful data centers. System designers will finally have an efficient new ecosystem, an ecosystem designed specifically for today's emerging computer architectures and protocols.

Easier programming of heterogeneous devices

The HSA Alliance is an alliance of semiconductor companies, tool/IP vendors, software vendors, and academic institutions dedicated to developing architectural specifications to fully utilize the performance and energy efficiency of parallel computing engines in modern devices. This makes it very easy to write programs for heterogeneous computing devices. It reduces the complexity of heterogeneous systems through a new ecosystem; it specifies parameters such as runtime and system architecture APIs that host cache-related shared virtual memory hardware and no longer require time-consuming operating system calls. Using a single-source shortest path algorithm, both control and computer code reside in the same file or project. Programmers do not need to interpret the tool chain of multiple processors for personal access.

Standard language programming

Another boon to AI application programmers is that the HSA platform can be used with a variety of different programming languages. Compilation tools are available from both proprietary and open source projects (LLVM and GCC), and HSA compilers are available for languages ​​such as C/C++, OpenCL, OPENMP, C++ AMP, PYTHON, etc. This flexibility greatly expands the capabilities and scope of AI applications currently on many drawing boards.

Improve developer efficiency

As a productivity engine, HSA fully exploits the power and potential of heterogeneous computing and eliminates many barriers to traditional heterogeneous programming. Developers can finally focus on their algorithms without having to manage system resources in a comprehensive manner. Their goal is to fund applications that seamlessly integrate scalar processing with high-performance computing such as CPUs, GPUs, DSPs, image signal processors, VLIWs, neural network processors, FPGAs, etc.

There is no doubt that AI will gradually affect the way we live, work and play. AI technology will serve as the forefront of digital transportation, intelligent manufacturing and medical technology. The power and flexibility of heterogeneous computing will make breakthroughs in AI technology possible and change the face of our world.

[1] The Heterogeneous System Architecture (HSA) Alliance has four goals: 1) achieve energy savings and improve performance; 2) improve the programmability of heterogeneous processors; 3) increase code portability between processors and platforms; and 4) increase the ubiquity of heterogeneous solutions across the industry.