A reconfigurable processor that makes AI faster

Deep learning is a key computing method that is driving technology forward—processing vast amounts of data and discovering subtle patterns that humans would never be able to discern on their own. But for optimal performance, deep learning algorithms need to be supported by the right combination of software compilers and hardware. In particular, reconfigurable processors that allow flexible use of hardware resources for computing as needed are key.

In a recent study, researchers in Hong Kong report that a new reconfigurable processor called ReAAP outperforms several computing platforms commonly used to support deep neural networks (DNN), a type of deep learning. A useful form that often involves data layers for large datasets that are computationally intensive. They describe it in a paper published Oct. 10 in IEEE Transactions on Computers.

In their study, the researchers compared their proposed software compiler in ReAAP with three other baseline software compilers on Nvidia GPUs and ARM CPUs. The results show that its execution speed is 1.6 to 3.3 times faster than running the same software compiler on a GPU and 1.6 to 3.3 times faster than running the same software compiler on a CPU.

In addition, Zheng noted that ReAAP enables sustained high utilization of hardware resources for a variety of different compute-intensive layers.

While ReAAP is good at handling DNNs with typical data-intensive workloads, it is currently not well suited to supporting DNNs when data is sparse. Zheng said his team hopes to address this issue in the future. What's more, the researchers hope to build on ReAAP so that it can better handle quantitative data (processed data in a way that significantly reduces the memory requirements and computational costs of neural networks).

“After the extension [of ReAAP to better handle quantitative data] is completed and evaluated, we will consider commercializing it along with several other AI computing acceleration solutions,” Zheng said, noting that this would make ReAAP more efficient in terms of resources More efficient for constrained platforms, such as various Internet of Things (IoT) devices.

While regular processors typically allow data to be processed using specific hardware paths, reconfigurable processors offer a more adaptable option: reconfiguring the most efficient hardware resources to process data as needed.

"Reconfigurable processors combine the advantages of software flexibility and hardware parallelism," explained Zheng Jianwei, a postdoctoral researcher in the Department of Electrical and Computer Engineering at the Hong Kong University of Science and Technology who participated in the research.

These advantages led his team to create ReAAP, an integrated hardware and software system. Its software compiler is responsible for evaluating and optimizing various deep learning workloads. Once it determines the best solution for processing data in parallel, it sends instructions to reconfigure the hardware coprocessor to allocate appropriate hardware resources for parallel computation. "As an end-to-end system, ReAAP can be deployed to accelerate a variety of deep learning applications by simply customizing a Python script for each application in [the] software," Zheng explained.

★ Click [Read the original text] at the end of the article to view the original text link of this article!

*Disclaimer: This article is original by the author. The content of the article is the personal opinion of the author. The reprinting by Semiconductor Industry Watch is only to convey a different point of view. It does not mean that Semiconductor Industry Watch agrees or supports the view. If you have any objections, please contact Semiconductor Industry Watch.

Today is the 3245th content shared by "Semiconductor Industry Observation" with you. Welcome to pay attention.