Nvidia is working on a new generation of hardware and software for autonomous driving and robotics – and has now launched the Drive AGX Orin platform. The centerpiece is the newly developed combination processor Orin (System-on-Chip / SoC), which consists of a total of 17 billion transistors. Nvidia told heise online that Orin is manufactured in an 8-nanometer manufacturing process and confirmed this when asked again. According to Nvidia, there will be first samples in the second quarter of 2021, the chip should be used from 2022.
Next generation GPU architecture
Orin contains twelve standardized ARM processor cores from the Hercules series – so Nvidia is turning its back on its previous custom chip strategy. The Hercules cores announced ARM for 2020 and promised a 15 percent higher computing power compared to Cortex-A76 cores – with a production of 7 or later 5 nm. , To what extent the Orin GPU will already contain bonds from Nvidia's next generation of professional GPUs in 2020 is unclear. Nvidia did not elaborate on the latter on request.
Together with additional accelerator cores, Orin should spark particularly high computing power in the application areas of machine learning (deep learning) and machine vision (computer vision). Nvidia speaks of a combined 200 trillion operations per second (200 TOPS INT8) per Orin SoC. However, it is unclear exactly how Nvidia determines the computing power. The maximum transfer rate is 200 GB per second, the video engines decode 8K at 30 fps and 4K at 60 fps; support the formats VP9, H.265 and H.264.
Like its predecessor Xavier (30 TOPS, Volta), Orin is also a fully scalable platform – this is important for implementing automated driving, for example. In the latter, a distinction is made between five levels, which describe the degree of relief for the driver through automated systems. The higher the level, the more computing and detection power is required. For level 3 (highly automated), Nvidia estimates two Orin SoCs with a total of 400 TOPS computing power and a power consumption of 130 watts; for level 5 (autonomous) two Orin SoCs and two GPUs with a total of 2000 TOPS at 750 watts. As part of the Orin announcement, Nvidia also announced that it would give interested automotive companies access to the Drive Neural Network trained with test car data.
Nvidia is under pressure in the field of autonomous driving – for example, the e-car manufacturer Tesla turned its back on the GPU specialist and developed its own module "FSD" for controlling its cars. This is manufactured by Samsung in the 14 nm FinFET process and is already installed in all current Tesla models ex works – and replaces the previously used Nvidia SoC Xavier. FSD delivers 144 TOPS at 72 watts.
. (tagsToTranslate) ARM (t) graphics chip (t) Nvidia (t) Orin (t) Tesla Motors (t) autonomous driving