Bellnix Co., LTD.

<p style=SUNNYVALE, Calif. — China’s ambitious drive to expand its semiconductor industry will fall far short of its targets, according to Bill McClean, president of IC Insights. In the short term, 2018 will be a good year for the global chip industry, despite the ups and downs of the DRAM market and capital spending, he said.

" style="width:80%;height:400px;">

Today, more than half the cost of a factory robot goes to systems integrators who configure it with sensors and train it, typically writing custom programs and generating proprietary data that stays on the factory floor.

Comparing real-world processors' efficiency (Source: Hailo)Click here for larger imageComparing real-world processors' efficiency (Source: Hailo)Click here for larger imageComparing real-world processors’ efficiency (Source: Hailo)

Based on each vendor’s published figures, Hailo compiled a list of AI processors — ranging from Nvidia’s Volta V100, Pascal P4, and Google’s TPU to GraphCore’s IPU and Wave Computing’s DPU — each with its own deep-learning tera multiply-accumulates per second (TMACs) and power. Hailo’s goal was to crunch out deep-learning TMACS per watt for each AI processor.

0550910874_Datasheet PDF

As it put together its own spreadsheet, Hailo observed that current AI processors achieve less than 0.1 TMACS per watt, or slightly above it when using batch processing methods.

Neural networks — baseline (Source: Hailo)Click here for larger imageNeural networks — baseline (Source: Hailo)Click here for larger imageNeural networks — baseline (Source: Hailo)

Let’s look at what it takes to process HD video,” said Hailo’s Danon. Assume that a vehicle receives full HD video at 30 frames per second and does deep learning using ResNet50 network. Typically required for processing full HD video stream is 5 TMACS per sensor. Depending on the autonomy level, a typical car is expected to feature four to 12 camera sensors.

0550910874_Datasheet PDF

The implication is that any current AI processor deployed in an automated vehicle is already consuming tens of watts per sensor, or a few hundreds of watts per car. In Danon’s view, this is too much. Alternatively, if OEMs decide that they can’t afford to squander so much power, they have no choice but to dramatically compromise on performance, he suspected.

In theory, a well-designed CNN accelerator should achieve orders-of-magnitude better performance per watt than traditional von Neumann processors,” maintained the Linley Group’s Gwenapp. But even this approach may not be enough to meet the challenging requirements for Level 4/5 self-driving cars.”

0550910874_Datasheet PDF

In other words, despite the plethora of AI processor chips teasing the market, none are close to achieving a performance standard that would make commercially viable, fully autonomous vehicles possible.

Gwennap remains optimistic. Fortunately, we are still early in the AI game. I expect hardware and software to deliver additional orders-of-magnitude improvement over the next decade.”

So already the designer of an IoT node that connects to an LPWAN is facing a tough cost problem. And if the cost of the radio is relatively high, then the BoM budget available to the remaining components in the circuit is even smaller. Of these remaining components, the memory chips are among the most expensive, and so offer the greatest scope for cost optimisation.

This article explores the choices of memory architecture available to designers of IoT end nodes containing a LPWAN radio, and explains the advantages of various external Flash memory solutions, depending on the system’s requirement for performance, memory capacity and system size.

Typical application requirementsThe configuration of internal and external memory capacity must be made in the light of the system’s functional requirements, which in turn depend on the application(s) it supports. The range of applications for LPWAN connectivity is of course huge. But examples of device types that might use LPWAN connections include:

In these examples, an SoC or microcontroller will be running application code as well as performing system management functions. This entails a requirement for memory: for code storage, for configuration data, and for user (application) data.


Related posts:

Copyright © 苏ICP备11090050号-1 tl431 datasheet All Rights Reserved. 版权投诉及建议邮箱