3D chip tech key as Meta AR prototype chips get huge performance boosts

CheckCitation/SourcePlease click:spectrum

(XR Navigation Network February 23, 2024）增强现实系统的设计存在诸多限制，尤其是眼镜形态，AR 眼镜身形狭小，所以搭载的芯片在能效和性能方面均有高度要求。根据IEEE日前发布的博文，MetaResearch scientist Tony Wu introduced the AR prototype chip the team is building. According to the presentation.3DChip technology became the key to the breakthrough, and the AR prototype chip gained a huge performance boost through wafer-to-wafer bonding.

Tony Wu所在的团队致力于开发一种名为Aria的增强现实系统。他告诉工程师，解决方案的一个重要组成是3D芯片集成技术。在ISSCC大会，Meta介绍了他们的AR原型芯片是如何使用3D技术在相同的区域以相同或更少的能量做更多的事情。

The prototype chip consists of two integrated circuits of the same size - 4.1 × 3.7 millimeters. They are joined together by a bonding process called face-to-face wafer-to-wafer hybrid bonding. In simple terms, it involves flipping two fully machined wafers so that they face each other and bonding them so that they are straight together. hybrid bonding means that it is a direct copper-to-copper connection that does not require solder.

Using TSMC technology, this means that two wafers can form vertical connections about every 2 microns. Of course, the prototype doesn't take full advantage of this density: about 33,000 signal connections and 6 million power connections are needed between the two wafers. Additionally, the chip on the bottom uses TSV silicon vias for power in and signal out.

3D stacking means the team can increase the chip's computational power without increasing the chip's size. The chip's machine learning unit has four compute cores and 1 megabyte of local memory on the bottom chip, but the top chip adds 3 megabytes of memory, which can be accessed through 27,000 vertical data channels at the same speed and with the same amount of energy (0.15pJ/byte).

The team tested the chip around a machine learning task critical to augmented reality (hand tracking). Among other things, the 3D chip was able to track two hands simultaneously, saving 401 TP3T of energy over a single chip tracking a single hand. What's more, it was 401 TP3T faster.

In addition to machine learning, this chip can also perform image processing tasks, and again 3D plays a very big role here. While the 2D version is limited to compressed images, the 3D chip can achieve full HD with the same power consumption.