According to Google, the drive by companies to build tensor SoCs stems from the fact that innovation is faster than the chip industry can offer. This seems like an odd statement considering last year's Pixel 5 was powered by the Snapdragon 765G, a chip that's far more powerful than last year's flagship Snapdragon 865 processor from Qualcomm. This is important because in many of Google's performance claims about the Tensor SoC, Google compares it to the Pixel 5 and Pixel 4.
During the launch of the Pixel 6, Google said how important machine learning performance is to the enterprise, that no single metric is more important than a specific component. But that seems to run counter to Google's decision last year to ship the Pixel 5 with the Snapdragon 765G processor, which only has five AI performance charts. The previous Snapdragon 855 Pixel 4 versus Google's claimed 7 TOPS ML performance explains why some users (including Anshel) have seen the Pixel 5 slow down when performing some AI tasks like noise reduction and other image processing.
The new Tensor SoC has an unknown clock speed paired with two ARM Cortex X-1 cores, two ARM Cortex A76 cores and a small quad-core ARM-A55 core. As far as I know, Arm partners haven't done this setup yet. This is because Samsung and Qualcomm built their SoCs using Arm X1 cores, followed by 3 A78 cores and 4 A55 cores. In addition, Google's Tensor Arms SoC uses the same Mali-G78 GPU core as the Exynos 2100. However, Google appears to be using a 20-core GPU instead of a 14-core one.
Google manufactures Google's tensor SoCs on Samsung's 5nm technology node, as does Samsung's leading modern Exynos SoCs. However, this design looks more personal than the new Exynos 2100 chip, but at the same time looks like an almost finished shoulder design developed by Samsung using Exynos building blocks besides TPU and some security features.
0 $type={blogger}:
Post a Comment