At yesterday’s GTC China 2019 conference, NVIDIA announced a new nuclear bomb product, drive AGX Orin, the next generation of self-driving platforms, with the new NVIDIA GPU and 12-core ARM CPU, with 200TFSLOP performing seven times as much as the previous generation of Xavier.
For NVIDIA’s next-generation system-level chip, orin chips consist of 17 billion transistors that integrate NVIDIA’s next-generation GPU architecture and the 12-core Arm Hercules – THE SPECIFIC s.of the CPU and GPU are not officially announced. The next-generation GPU should be ampere ampere after Turing, the CPU is the next generation after arm Cortex-A77, with the Super Energy Cortex-A78, mainly for 7nm and 5nm processes.
Because Hercules CPU is for 7nm or 5nm process, so we are looking forward to the DRIVE AGX Orin, there is a possibility of the first 5nm process, the most recent also has to be 7nm process, did not expect NVIDIA this time is really conservative to home – The Computerbase website reported that DRIVE AGX Orin was using samsung’s 8nm LPP process and did not have tSMC’s part-time.
DRIVE AGX Orin will not be available until 2022, three years later, and Samsung’s 8nm LPP process is actually an improved version of Samsung’s 10nm process. To know that by 2022, the 5nm process is not the first to enter, TSMC next year will be mass production of 5nm process.
But then again, NVIDIA in the process, although a little key, but compared xavier is still a lot of progress, after all, it is only TSMC 12nm process production, DRIVE AGX Orin with 8nm process has been a great progress.
There is no doubt that using Samsung’s proven 8nm LPP process can significantly reduce costs, but it is clear that energy efficiency cannot be compared to 7nm and 5nm. According to NVIDIA, the L5-level fully autonomous driving system can use two DRIVE AGX Orin plus 2 sets of GPUs, performance up to 2000 TFLOPS, also 2000 trillion times, this performance is more than a few.
Of course, the price is power consumption, such a system can reach 750W power consumption, compared to over-calculation will be very energy-efficient, Computerbase website made an energy efficiency calculation, you can see.