NVIDIA have unveiled a new System-on-Chip designed for autonomous or self driving vehicles, codenamed "Parker." Parker is the chipset that's behind NVIDIA's Drive PX 2 platform, which the company has described as a mini "supercomputer for cars." The company's Andi Skende, an engineer, talked about Parker at the Hot Chips conference, California.
In the technical detail, the Parker chipset is based around a six core system that belongs to the NVIDIA Tegra chipset family but uses NVIDIA's Project Denver technology. Project Denver is a five year plan designed to put desktop class performance into mobile devices and each Parker chipset has a dual core 64-bit Denver 2.0 core, related to the application processor cores found in the HTC Nexus 9. Under testing, the Nexus 9's chipset has exceptionally high single thread performance but in synthetic multi-core benchmarking, it usually does not outperform its competitors, which use higher numbers of less powerful cores. For the Parker chipset, NVIDIA have supplemented this core Denver arrangement with a quad-core, 64-bit, ARM Cortex-A57 tier. All six processor cores work together in what NVIDIA call a "fully coherent heterogeneous multi-processor configuration." Interestingly enough, the Tegra X1 chipset, which has made it into a small number of Android devices (a modified variant is used in the Google Pixel C, which has the less powerful processors disabled) does not use the Denver architecture. NVIDIA have also said that each Parker unit delivers between 50% to 100% higher multi-core CPU performance than other mobile chipsets. According to NVIDIA, Parker delivers up to 1.5 teraflops of performance, which is used for deep learning technologies. One more thing: the Drive PX 2 platform uses two Parker units, plus two Pascal graphics processor units.
This performance is necessary as the deep learning required in order to control a vehicle, recognizing and avoiding hazards, and understanding the appropriate and necessary avoiding action requires considerable computing resources. NVIDIA explains that the Drive PX 2 delivers 24 trillion deep learning operations a second, which is used for deep learning technologies. The Drive PX 2 unit has already been successful and over eighty car manufacturers, universities and suppliers have started using the technology. Volvo is planning to road test cars containing the Drive PX 2 system in 2017. Meanwhile, we are unlikely to see the Parker chipset in our mobile devices: although the Drive PX 2 is significantly more than a single Parker chipset, the cost of this high performance is both heat output and power consumption. The Drive PX 2 unit is liquid cooled and consumes 250W of power. Much of the power consumption and heat output is from the discrete Pascal GPU boards rather than the CPU cores, but these will still generate a lot of heat.