Paris, France-based VSORA, is delivering the first PetaFLOPS computational platform to accelerate L4 and L5 autonomous vehicles designs. The programmable solution is delivered as an IP block that combines DSP and ML acceleration for the autonomous driving industry. Its multicore DSP and AI architecture eliminates the need for DSP co-processors and hardware accelerators to provide a high degree of flexibility.
Elaborating on the VSORA PetaFLOPS computational platform, Khaled Maalej, CEO and founder of VSORA, said it offers significant processing power to implement AI and ‘traditional’ signal processing algorithms in the same chip simultaneously.
“The frontier between AI and DSP is fading away, and ADAS/AD (autonomous driving) is proof of that. In ADAS/AD applications, there is no limit in terms of required processing power, the more you have, the better algorithms you can design, to the benefit of higher reliability in your designs. This is key in automotive.”
It is not rocket science to implement a PetaFLOPS solution. The challenge is to design an efficient platform in terms of high processing power and low energy consumption to embed in a car.
For example, a PetaFLOPS platform can only use 10 percent of the resources because it cannot feed all the computational units with data to keep them constantly active is processing only 100 TeraFLOPS (10 times less). A major bottleneck rests with the external memory, and that is one of the main issues we have addressed in our innovative architecture. We can exceed 80% efficiency in most of the cases.
He added: “In addition to high computational power and low energy consumption, we also offer a high-level of abstraction development flow. We compile Matlab-like or TensorFlow-like code, or a combination of both, straight through to RTL to accelerate the development and allow the algorithmic engineers to focus on creating more advanced algorithms. In other words, we remove the implementation from designer tasks, and provide them with quick and accurate end results to enable “trail-and-errors” analysis or to experiment with different algorithms.”
Regarding the automotive mega trends, such as autonomy, electrification, and connectivity, positioned for 2021, he added that lately, lots of financial resources and human effort are spent in ADAS/AD, aiming at getting significant outcomes in 2023, and landmark changes in 2025.
Zero defect for zero accidents
Next, I wanted to know how VSORA has been enabling zero defect, a must to enable zero accidents for autonomous vehicles.
He said that the zero defect must be supported by several elements in the autonomous driving vehicle to reach the zero accidents. In hardware implementation, the most important elements today are algorithms and sensors. While algorithm redundancy is needed, also needed are several sensors in the car.
The data provided by all these sensors must be fused in order to build a reliable environment for driving the car. On a foggy day, for instance, the control system cannot rely on the cameras. Instead, it may have to use radars and/or lidars. The switch between the two types of sensors must happen smoothly and reliably. The VSORA device has been designed to ensure the above.
Next, there is a need to know how are automotive electronics changes, including those in the internal combustion engine (IGE), shaping up? He said car manufacturers are facing a significant challenge in implementing AD vehicles. The car is becoming the most complex system in the industrial world. Microsoft Office includes around 40-million lines of code. The software in the autonomous driving car is requiring around 100-million lines of code!
Maalej added, “Car makers have to build their own OS, and use a computational platform in the range of the PetaFLOPS to handle the complexity.”
Further, how is VSORA meeting the need for increased power density, integration of disparate technology? Maalej said a PetaFLOPS platform can consume significant energy that may prevent its integration into the car. We had to address this issue and we did so with two approaches:
First, the architecture is designed to reduce to the maximum data transfer from external memory using an embedded RAM for storing and transferring data between the AI and DSP sections of the platform. Unique to our approach is that the SRAM acts as a vast collection of registers. Second, we adjust the computation accuracy of the system on-the-fly as needed.
Achieving ADAS autonomy
Are we currently far from achieving autonomy for ADAS? He noted that while there are still some challenges to solve, in general, the development is advancing very rapidly. The Level 4 (L4) autonomous driving should be available in some high-end cars in 2025/2026.
The L5 autonomous driving, where the dashboard and the steering wheel will disappear, will take longer and not because of intricate technical issues. Rather, because of legal issues (who is responsible when an accident happens?). Basically, L4 is L5 without a dashboard and steering wheel. It has the same level of technology.
Next, how is the multicore DSP and AI architecture eliminating the need for DSP co-processors and hardware accelerators? He said that in virtually, all the existing DSP implementations, an important portion of processing power is off-loaded to a dedicated hardware. This is called co-processors. They are hard-wired and not programmable. If you need/want to change your algorithm, the above prevents you from achieving your objective.
He added: “In our solution, we do not need/use co-processors. Everything is programmable. We have a different architecture and we implemented a new DSP approach, driven by 5G and 6G applications.”
Finally, who all are using VSORA solutions so far? Without disclosing names, a major European car manufacturer already taped out the platform in 7nm process and confirmed the validity of VSORA’s claims.