AI Accelerator Module: Disrupting AI

What is an Accelerator Module?
The AI Accelerator Module is a modular M.2 card that contains an AI processor. Cards can typically produce around 26 tera-operations per second while being highly energy efficient; which is due to the AI processor being used being optimized for AI. These cards can support PCIe Gen-3.0 2-lane interfaces and deliver enhanced performance to edge devices. Due to the small form factor any edge device that has a compatible M.2 socket can utilize AI processing in real time.

Through clever design these AI accelerator modules do not draw a lot of power while allowing software developers to quickly port their solutions to the AI module through robust SDK tools. Cards support standard AI workflows such as containers that can be rapidly integrated into the card. In addition to this, if the developer does not want to create an AI solution from scratch, they can use out of the box solutions that have been provided, with only minor changes needed.

A New Approach
These new AI modules are unique due to their small size and ability to work in real-time to process sensor inputs. Their novel modular slimline passively cooled solution is also an exciting feature in itself. Earlier AI solutions leveraged GPU processors but at the expense of cooling being required and larger PCI slots which many devices will never support. So, you can have the power of AI in small devices that do not need a constant internet connection. This makes them exciting for Internet of Things (IoT) devices where smarter devices with less connectivity are becoming a better solution; essentially mimicking social interaction where the device can act independently from the group yet interact when required, saving energy and reducing security risks.  

Edge Devices
Small edge devices are becoming useful in automated manufacturing plants where process control needs to be monitored with every product assessed for conformity. As you can imagine, a manufacturing plant can make hundreds of products per second so inline real-time solutions must be pragmatic to catch defects and issue inline maintenance activities quickly. Interestingly, such plants are so extensive that some edge devices used in these types of activities may need to be remote and work with a non-consistent internet or ethernet connection making conventional AI centralized systems difficult in such plants. The major benefit of AI Modules for edge devices is that a connection is not needed all the time.  
Another example for AI modules is in assisted driving car technology where a dedicated ECU is required that is connected to multiple sensor inputs and used to manage multiple AI functions. For example, assisted steering or reverse parking need pragmatic AI solutions and again require real-time performance. Here the compact M.2 design and passive cooling makes for a maintenance free ECU that is sealed from the elements and modular in nature. As you would not need a fan to cool the system you could use the same technology in military vehicles due to MIL-STD compliance requirements stating a sealed system is required along with zero maintenance.   

AI Modules and The Future
Due to the ease of software development workflows due to SDK tools, the modular nature of the cards and the passive cooling it is likely that we will see a lot more edge devices on the market soon. Companies will take advantage of low energy highly performant devices that have the intelligence to act independently to the next level of IoT alone. AI modules will also be mounted in automotive and manufacturing devices where decentralized AI is required to make a smarter offline product. Due to the modular design of the card, it also makes the overall product more sustainable and follows automotive industrial practices for cradle to incarnation practices. No matter your product, if you need AI that is easily added in terms of both hardware and software development practices, this is an easy way to achieve it.