How Computers and Artificial Intelligence Evolve Together

Co-design, or designing software and hardware at the same time, is one way to attempt to meet the computing power needs of today’s artificial intelligence applications. A compiler that converts instructions from one representation to another is an important piece of the puzzle. A group of researchers from the Chinese Academy of Sciences summarized existing compiler techniques in deep learning collaborative design and proposed a unique framework, Buddy Compiler.

A group review paper was published in a journal intelligent computing.

Optimization, hardware architecture, co-design approaches, and compilation techniques have been summarized by others, but no one has discussed deep learning systems in terms of compilation techniques for co-design. Researchers study deep learning from this angle because they believe that “compilation techniques can increase opportunities for co-design and, as a result, better achieve the performance and power requirements of deep learning systems.” bottom.

Five topics are covered in this review.

• The history of deep learning and co-design
• Deep learning and co-design today
• Compilation Techniques for Deep Learning Collaborative Design
• Current Problems and Future Trends
• buddy compiler

The history of deep learning and co-design

Since the 1950s, neural networks have undergone many ups and downs, leading to today’s explosive growth in deep learning applications and research. Collaborative design began in his 1990s and has since been adopted in a variety of disciplines, progressing from manual to computer-aided design and finally to a complex process involving modeling, simulation, optimization, synthesis and testing. became.

Since 2020, a network model called Transformers has achieved great success. ChatGPT is a chatbot built using a “Generative Pretrained Transformer”. Current AI applications like ChatGPT are reaching new performance bottlenecks that again require hardware and software co-design.

Deep learning and co-design today

Deep learning breakthroughs come from the use of large numbers of layers and huge numbers of parameters, which greatly increase the computational demands for training and inference. As a result, it becomes difficult to achieve reasonable execution times when relying solely on software-level optimizations. To address this, both industry and academia have developed domain-specific We’ve been looking at hardware solutions.

Comprehensive systems have recently emerged, consisting of deep learning frameworks, high-performance libraries, domain-specific compilers, programming models, hardware toolflows, and co-design techniques. Together, these components contribute to improving the efficiency and effectiveness of deep learning systems.

Compilation Techniques for Deep Learning Collaborative Design

There are two popular ecosystems used to build compilers for deep learning. The Tensor Virtual Machine, known as TVM, and the Multilevel Intermediate Representation, known as MLIR. These ecosystems employ different strategies, with TVM acting as the end-to-end deep learning compiler and MLIR acting as the compiler infrastructure. Meanwhile, in the area of ​​hardware architectures customized for deep learning workloads, there are his two main types: streaming architectures and compute engine architectures.

The hardware design tool flows associated with these architectures also employ new compilation techniques to drive progress and innovation. The combination of deep learning compilers and hardware compilation techniques opens up new opportunities for deep learning co-design.

Current Problems and Future Trends

Effective co-design is critical because performance requirements are growing too fast for processor development to keep pace. The problem with co-design is that there is no single way to do it, no unified co-design framework or abstraction. It becomes less efficient when multiple layers of abstraction are required. It takes a lot of effort to customize the compiler for a specific domain. A unified ecosystem is forming, but the root cause of fragmentation remains. The solution to these problems is a modular, extensible integration framework.

buddy compiler

The Buddy Compiler project contributors “are committed to building a scalable and flexible hardware-software co-design ecosystem.” Ecosystem modules include Compiler Frameworks, Compiler Platform as a Service, Benchmarking Frameworks, Domain-Specific Architectural Frameworks, and Co-Designed Modules. The latter two modules are still in progress.

The authors foresee continued development of a compilation ecosystem that will help consolidate the work being done in the rapidly evolving somewhat fragmented field of deep learning.