From Self-Driving Cars to Smart Homes: The Tech Revolution Powered by Spatial-Temporal Mapping

image-16.jpg

The next generation of smart devices, from autonomous vehicles to sophisticated home assistants, hinges on the ability to process complex tasks with unprecedented speed and efficiency. A recent breakthrough by researchers Song Wang, Yiyuan Gao, Bingfeng Seng, Jing Pei, Yuan Zhang, and Jianqiang Huang offers a significant leap forward. Their spatial-temporal density mapping technique optimizes many-core systems, enabling them to handle large neural networks with remarkable resource efficiency, unlocking new possibilities for real-time AI applications.

Many core systems are popular in modern technology because they can handle multiple tasks at once while keeping power use low. These systems are made up of many small cores (processing units that act like mini-computers), each with its own memory and computing abilities. In these chips, memory space and computing power need to work together to run neural networks (software models that help machines learn and make decisions). However, neural networks often have uneven demands, with some parts needing more memory and others more computing power. This imbalance can slow things down or waste a significant amount of resources.

Standard methods to organize tasks on these chips fall into two categories. Temporal mapping splits tasks over time, letting each core work on its piece one after another, but this often leads to extra data copying, which fills up memory quickly. Spatial mapping divides tasks across different cores based on space, reducing data copying but causing delays because some cores finish faster than others. A common approach, called layer-wise mapping, combines both methods but still struggles with uneven resource use, leading to delays when switching between different parts of the neural network.

The researchers propose a new method called spatial-temporal density mapping to fix this resource inefficiency. It balances memory and computing power by carefully organizing how tasks are assigned to cores. They define a concept called spatial-temporal density (a measure of how much computing work a core does per unit of memory space) to ensure each core is used evenly. On their test chip, TianjicX, they showed that this method makes all cores work at the same level, eliminating delays from uneven task distribution. It also reduces the need for data shuffling between layers of the neural network, saving both memory and time.

To improve upon this method, the team introduced two techniques. The first, Negative Sequence Memory Management (NSM, a way to reuse memory space more efficiently), stores data in a reverse order so that memory from one task can be reused for the next task right away. This approach cuts down on wasted memory space compared to the older method, Positive Sequence Memory Management (PSM, which stores data in a forward order and delays reuse). Tests on TianjicX using a neural network called ResNet-50 showed that NSM boosts memory efficiency by 3.05 times, increasing available memory from 204 kilobytes to 631.5 kilobytes across 17 test blocks. In one specific block, NSM left 45 kilobytes of free memory, a 3.75-fold improvement over PSM’s 12 kilobytes, preventing memory overflow that PSM couldn’t avoid.

The second technique, Many-core Parallel Synchronous (MPS, a method to sync up computing tasks across cores), focuses on speeding up calculations. Neural networks often break down tasks into smaller pieces called partial sums (Psums, intermediate results that need to be combined for the final answer). Older methods, like the Pipelined Manner (PM, a way to pass Psums through cores one by one), leave some cores idle, slowing things down. MPS ensures all cores work together at the same time, dividing Psums evenly and combining them faster. Simulations on TianjicX with 16 cores showed that MPS cuts computing time by 6.7 percent compared to PM, making it the fastest among other methods like Step-by-Step (SS, a slower sequential approach) and Dichotomy Step-by-Step (DSS, a method using a tree-like structure for calculations).

The TianjicX chip itself is extremely powerful, built with a 28-nanometer process (a manufacturing technique for making tiny circuits) and featuring 160 cores. Each core can handle complex math operations like vector multiplications and accumulations, thanks to its six modules working together. These include a dendrite module (a part with 128 parallel multipliers for fast calculations) and a dual port memory (a storage system with two 64-kilobyte banks for quick data access). The chip uses a fully digital design, avoiding crossbar structures (a type of memory layout common in other neuromorphic chips), and supports flexible task mapping. The research team tested it with a setup including a host computer, an Intel Arria 10 FPGA (a programmable chip for running simulations), and an oscilloscope (a tool to measure timing), ensuring accurate results through a dual verification simulation process.

To test their mapping method, the researchers used two neural networks on TianjicX. One was a part of ResNet-50, a standard model for benchmarking chips, and the other was a custom network without residual connections (links that help data flow between layers). They compared their method to the layer wise approach, varying how many layers were coupled together to adjust the spatial-temporal density. Results showed that coupling six layers together reduced the time per layer from 259 microseconds to 140 microseconds, a 1.85-fold speed increase over the layer-wise method. This setup also eliminated delays between cores, as data stayed local and didn’t need reshaping between layers. However, they noted that coupling too many layers can backfire, as it complicates task division and slows down calculations, highlighting a trade-off that depends on the chip’s design.

This technology could improve many devices we use daily. In a self-driving car, it could help the car’s system process road data faster, spotting obstacles in real time to avoid accidents. For a smart home camera, it could mean quicker detection of unusual activity, sending alerts without delay. In gaming, it could make virtual characters react faster, creating a smoother experience. Even in edge robotics, like a delivery drone navigating a busy street, it could ensure quick decisions without draining the battery. By making chips more efficient, this method paves the way for smarter, faster technology that fits into our lives seamlessly.

The researchers believe their work sets the stage for a general framework to optimize many core systems. By fine tuning how tasks are mapped, they’ve shown that balancing memory and computing power can lead to large gains in speed and efficiency. This could help future chips handle even larger neural networks, opening doors for more advanced applications in robotics, artificial intelligence, and beyond.

Citation

Wang, S., Gao, Y., Seng, B., Pei, J., Zhang, Y., & Huang, J. (n.d.). Fine spatial-temporal density mapping with optimized approaches for many-core system. Frontiers in Neuroscience. Available at: https://www.frontiersin.org/articles/10.3389/fnins.2024.1418735/full

NeuromorphicCore

Related Articles

PHP Code Snippets Powered By : XYZScripts.com
★

Track This Company

Create a free account to add companies to your watchlist and get notified when patents, funding, or key developments change.