Dr. Shaoshan Liu is CEO and founder of PerceptIn, an smart robotics company.
In the past couple of decades, the semiconductor sector has been a vital contributor to our economic climate. The most recent stats from the Semiconductor Sector Association (SIA) show that world-wide semiconductor sector profits totaled $555.9 billion in 2021.
In addition to the direct influence on our financial system, the semiconductor industry is the critical motor that powers all information and facts technology fields, such as particular computing, cellular computing, cloud computing and so on. For that reason, the semiconductor industry has a significantly further and broader effect on our fashionable overall economy.
Searching back at details technological know-how record, personal computing has been generally run by reasonably priced microprocessors created and made by companies these as IBM, Apple and Intel. Cellular computing has been largely powered by method-on-chips (down load expected) built by Apple, Qualcomm, Broadcom and Samsung. Cloud computing has been generally driven by server processors made by Intel, AMD, Nvidia and, much more lately, Google’s TPU.
Just after doing the job in the robotic industry for the past 10 years, I would argue that the next huge trend in info know-how is autonomous machine computing. Considering that autonomous device computing is the core know-how stack that empowers numerous sorts of robots, such as Mars or Lunar explorers, smart autos, autonomous drones, shipping and delivery robots, dwelling provider robots, agriculture robots, sector robots and numerous additional that we have however to visualize. These robots are going to absolutely revolutionize our overall economy, as I reveal in my former publish.
Very similar to other facts technology stacks, the autonomous equipment computing know-how stack is made up of components, programs computer software and software application. Sitting in the middle of this technology stack is computer system architecture, which defines the main abstraction in between hardware and software package.
The existence of this abstraction layer makes it possible for software builders to emphasis on optimizing the software package to absolutely make use of the fundamental hardware to produce superior applications as properly as to reach larger general performance and increased electricity efficiency. This abstraction layer also lets components builders to focus on developing a lot quicker, much more affordable, additional strength-economical hardware that can unlock the imagination of application builders. In a way, computer system architecture is the contract involving hardware and software package developers to obtain best division of labor and consequently optimal enhancement effectiveness.
Hence, computer architecture is necessary to information technologies. For occasion, in the personal computing era, x86 has become the dominant laptop architecture thanks to its exceptional general performance. In the cell computing era, ARM has come to be the dominant personal computer architecture due to its outstanding vitality effectiveness. Much more not too long ago, RISC-V could come to be a big laptop architecture in Web of Issues (IoT) computing because of to its openness.
Nevertheless, we however will need a appropriate computer architecture for the robotic age. In the earlier 5 a long time, several laptop or computer architecture proposals have emerged, and some have arrived at the early stage of commercialization.
1. Dataflow Architecture
Dataflow architecture was a preferred study topic in the 1980s as it is potentially far more effective when compared to Von Neumann architecture, on which the preferred x86 architecture is centered. Even so, dataflow architecture has been tested not suited for personalized computing workloads and failed to come to be a mainstream pc architecture.
Nonetheless, autonomous machine computing exhibits solid dataflow patterns as the modules in just autonomous machines properly sort dataflow graphs. Scientific studies, like a research exactly where I served as a researcher, have demonstrated that making use of dataflow graphs as abstractions of autonomous machine computing delivers a great deal increased performance as opposed to present laptop architectures. For instance, startup companies these types of as SambaNova, GraphCore and Cerebras are optimizing robotic workloads working with dataflow-like computing paradigms.
2. Factor Graphs
A variable graph is a graph representing the factorization of a likelihood distribution operate and has been utilized in lots of autonomous device computing features, these types of as localization, tracking, setting up and command. At the moment, research labs all over the environment are building algorithms to use element graph as a typical abstraction for most if not all autonomous device computing features. If prosperous, this will give a quite straightforward interface for mapping autonomous equipment functions to the underlying compute components.
3. Close-To-Conclusion Deep Discovering Designs
Stop-to-stop deep studying styles have also turn out to be a new exploration trend. For occasion, the computing of an autonomous car or truck can be modeled as two transformer designs, one particular handles the notion and localization features, whilst the second handles the arranging and management capabilities. Despite the fact that finish-to-end deep finding out models continue to put up with from lots of complications, I consider it has immense prospective. For occasion, Tesla’s comprehensive self-driving (FSD) is optimized for deep understanding workloads, and it has obtained substantial good results in powering Tesla’s robotic products. The FSD architecture could progressively evolve to help close-to-conclude deep understanding executions.
To conclude, autonomous machine computing is probably by significantly the major prospect for the semiconductor marketplace. As semiconductor firms rush to build several autonomous machine computing items in the up coming decade, we should witness an explosion of pc architectures for autonomous devices.
By the demo-and-error technique, the industry will ultimately recognize the dominant laptop or computer architecture for autonomous device computing. The robotic age is certainly a golden age for laptop or computer architects.