How ASIC chips beat CPU and GPU in AI

How ASIC chips beat CPU and GPU in AI
Needless to say, AI has become an increasingly widespread concept. Whereas in the past the backend server was tasked with running complex AI algorithms, more and more this capability has moved to the edge, which now has enough computing power to do the job. End users, then, stand to benefit from this AI edge processing.
 
Increasingly, AI is gaining traction among users across verticals due to its various advantages. AI- or deep learning-based facial recognition, for example, can detect whether someone belongs to a building or whether more men or women are visiting a particular aisle at a retail shop. The end user, then, gains situational awareness and business intelligence over their operations.
 
The processing of AI data can be done in the cloud or on the edge. More and more, the latter has become popular and widespread. “The cloud-based approach has some issues. For example, data traveling back and forth between endpoints and the cloud results in long latency, and to process such tremendous data requires high power consumption,” said Nobuhiko Aneha, President of Socionext Taiwan. “So we want to focus on the edge side.”
 
AI edge computing applications are wide and varied and can include the following: video computing, whereby images can be immediately processed to allow security and business intelligence; sensing, whereby smart home doorbells integrated with radars, for example, can detect the heartbeat and possible emotional status of a visitor; storage, whereby the AI decides which videos/images should be sent to which storage destination; and automotive, where a vehicle’s infotainment system can immediately play a song upon the driver’s voice command, or detect if the drive shows signs of fatigue and issue an alert accordingly.
 
Indeed, with AI edge computing applications becoming more prevalent, more and more AI-based analytics and algorithms have emerged. The need to differentiate, then, becomes critical in an increasingly competitive market. Unfortunately, standardized solutions such as CPUs and GPUs make such differentiation difficult. Application-specific integrated circuits (ASICs), then, become a viable solution that can be customized to any specific application and optimize a unique AI algorithm.
 
 

Comprehensive AI edge solutions by Socionext

 
This is where Socionext’s value comes in. Socionext’s AI edge solutions range over arm core-based, digital signal processors (DSPs) and ASICs, depending on the user’s requirements. For the customer, the end result is AI edge products/devices that can better differentiate in the market with low-power, low-latency and high-miniaturization.
 
“For example we helped develop an ASIC/AI accelerator with four chip consuming 3.2 watts of power yet achieving 22 TOPS performance. This is pretty much the same level of performance as a 250W GPU,” said Allen Chang, Business Development Director at Socionext Taiwan. “Our ability to create low-power, low-latency and high-miniature AI edge solutions is how we can enable differentiation for our customers, which can include ODMs/OEMs, AI developers and even competitors. We firmly believe that with new technologies, new supply chains will be created, and this is what’s happening now.”
 

A robust partnership ecosystem

 
Socionext’s success lies on its vast ecosystem of robust business partners, including Quanta, who's actively working with Socionext on their AI camera.
 
“Quanta, as a leading electronic hardware manufacturer who's targeting to expand the IoT market with AI technology, shares the same vision with Socionext on developing AI-based smart city, smart homes, smart transportation, and medical applications in the future,” Aneha stated.
 
“Quanta is one of the top Taiwanese companies investing heavily in artificial intelligence. In addition to consistently providing the market with high-quality products, we work closely with our excellent partners,” Quanta states. “With our vision of AI-enabled in various fields, we’re proud to work with Socionext, whose advanced chips helped elevate Quanta’s AI products to the next level. Quanta continues to leverage the power of edge computing to create a high-density and instant AI image-processing applications. We look forward to finding future collaboration opportunities with Socionext along with its expertise in low-power, high performance customized chips.”

For further information about Socionext please contact here.


Product Adopted:
Other


Share to:
Comments ( 0 )
security 50