Camera/Vision-Centric AI, Using Highly Specialized Hardware, will be Essential for Full Level 5 Autonomy
Mountain View, Calif., October 4, 2017 – AImotive, the full-stack autonomous vehicle technology supplier, today announced the availability of its FPGA prototyping platform, designed to demonstrate the hardware acceleration potential of aiWare, using the company’s aiDrive software stack that targets Level 5 fully autonomous vehicles. aiDrive software uses a wide range of CNNs (Convolutional Neural Networks) to implement a compete solution, resulting in a uniquely rich set of data that AImotive uses to drive its aiWare roadmap.
While many hardware solutions are now available on the market, they are designed to implement today’s relatively simple Level 2 and Level 3 applications. AImotive’s experience has shown these hardware solutions are not adequate to implement a realistic, safe Level 5 implementation for several reasons, including:
“Image processing is a well-understood, mature area that has been developing for decades,” said Laszlo Kishonti, founder and CEO of AImotive. “The research has shown that CNNs and Deep Neural Networks (DNNs) are the preferred method for analyzing images, so we can now focus on creating more highly optimized hardware to implement them as most algorithms being developed rely on CNN performance.”
Kishonti went on to say, “If we are to realize mass production of truly autonomous level 5 vehicles, we need to optimize the hardware wherever we can to maximize performance while minimizing cost and power consumption - just as video did over the past decade. We need dedicated hardware to hit the huge performance targets necessary for on-board embedded inference engines, which is why we created aiWare.”
AImotive is also helping drive open industry standards such as Khronos Group’s NNEF (Neural Network Exchange Format). The SDK (Software Development Kit) for the new aiWare hardware will provide full support for the emerging standard, providing an easy way for application developers to move their existing implementations into the aiWare environment. This is part of AImotive’s roadmap,making it as easy as a possible for OEMs to develop their autonomous driving solutions, upgrade their hardware and software quickly and easily, and take full advantage of the benefits of aiWare hardware acceleration.
“We find that many hardware platforms fail to achieve a fraction of their claimed performance due to a lack of understanding of the complete system,” said Marton Feher, head of hardware at AImotive. “We have our own cars running our own software on roads in the U.S. and Europe every day, allowing us to optimize the aiWare hardware architecture to deliver levels of performance never believed possible in an embedded, mainstream in-car inference engine.”
Many customers have discussed with AImotive the benchmarks they’re using to make significant decisions on what hardware platforms they use. “The problem is that these benchmarks are written only for today’s relatively simple use cases,” continued Feher. “They’re simply not good enough. Since many of our engineers come from Kishonti, a globally respected benchmarking company, we understand the importance of having the right benchmarks to drive the industry forward. We believe the current benchmarks being used are quite dangerous, as they will encourage people to choose hardware platforms that will not scale to the levels needed for Level 5 autonomy.”
Feher will deliver a paper on the challenges of designing hardware accelerators for Level 5 autonomy at the Linley Processor Conference in Santa Clara, California, October 4-5, 2017. The Linley Group recently published an article analyzing the aiWare hardware acceleration technology in their well-known Microprocessor Report. While at the conference, AImotive will demonstrate aiWare FPGA systems, running real-world examples using aiDrive software, and key executives will be available to discuss the future of embedded AI and autonomous vehicles.
Download the full press release here.
Read our publicly available benchmark results here.