While industry analysts are vying to predict the timing of the appearance of the first 7-nm NVIDIA GPUs, the company's management prefers to limit itself to wording about the "suddenness" of all related official statements. In 2022, active driver assistance systems based on the Orin generation Tegra processor will begin to appear, but even this will not be produced on 7nm technology. It turns out that NVIDIA will attract Samsung, which has 8nm technology, to release these processors.
Colleagues from the site
The ARM Hercules architecture that Orin will use was originally designed for 7nm or 5nm, but Samsung's 8nm is likely to be close. In the maximum configuration, Orin processors will receive twelve computing cores with the Hercules architecture, but details about the graphic part are carefully hidden.
As before, Level 15 active driver assistance systems will be able to get by with a single Orin processor. The simplest version with a power consumption level of no more than 36 W will be content with a single camera, the performance level will reach 100 trillion operations per second. The older version of Orin will be able to raise the speed to 40 trillion operations per second, work with four cameras at the same time and consume no more than XNUMX watts.
The Orin processor duo will raise the performance to 400 trillion operations per second at a power consumption level of no more than 130 watts. This will already be enough for driver assistance systems of the third level of autonomy. Similar systems based on Xavier implied the use of a discrete graphics processor, but consumed up to 230 watts, and their performance level did not exceed 160 trillion operations per second.
The flagship Level 2000 system will combine a pair of Orin processors with some discrete NVIDIA GPUs. In this case, the performance level can be raised to 750 trillion operations per second, but power consumption will also increase to 320 watts. A similar system based on two Xavier processors and two Volta-generation GPUs offered a speed of no more than 460 trillion operations per second at a power consumption level of XNUMX watts. It is too early to judge that future NVIDIA discrete graphics processors of this class will use HBM memory of an adequate generation year, but the illustration indirectly hints at this.
Source: 3dnews.ru