The video showcases Lightmatter’s Envise, an optical photonic processor that uses light to perform AI computations with low power consumption, demonstrated by playing an Atari game through a neural network running on the chip. While promising for future ultra-fast, energy-efficient computing, the technology is still in early stages with limitations in scale and performance compared to traditional and quantum computing.
The video explores the concept of computing using light instead of traditional electricity-based methods. Historically, computing has relied on electrical signals and transistors to perform operations like addition and multiplication, which consume significant power and require complex timing and circuitry. Optical computing, which uses light to perform computations, promises near-instantaneous processing speeds because light interacts at the speed of light. However, achieving practical optical computing is challenging because the physical structures needed to manipulate light for computation are much larger than the tiny transistors used in modern chips, making the chips themselves quite large.
Lightmatter, a startup initially focused on optical computing before pivoting to optical networking, has continued developing optical computing technology and introduced a product called Envise. Envise is an optical chiplet that uses optical interposer technology to combine light for computational tasks like addition and multiplication at a reasonable scale. The chip is relatively large due to the size requirements of optical components but offers very low power consumption, primarily limited by the power needed for the lasers. The system involves converting electrical data into optical signals and back, which adds complexity to the design and implementation.
At Supercomputing 2024, the presenter was shown a demo of Envise in action. The setup included a PCIe card with the Envise chip and lasers, connected to a server running the system. The demo featured an Atari game being played, not by running the game on the photonic processor itself, but by using a neural network model running on Envise to interpret the game’s video output and decide the next moves. The game ran on a regular CPU, while the photonic processor acted as an AI player, processing frames and controlling the game through the neural network.
The neural network used in the demo was a basic convolutional neural network (CNN) designed to process 84x84 pixel images with 4-bit color depth. It performed standard operations like convolution and ReLU activation to output one of nine possible actions, such as moving left or right, firing, or combinations of these. While the photonic processor was able to play the game, its performance was modest and far from human-level skill, highlighting that this technology is still in its early stages. The demo demonstrated the potential of photonic computing for AI tasks but also underscored the current limitations in scale and capability.
In conclusion, photonic computing represents an exciting but nascent field with the potential to revolutionize AI and computing by offering ultra-low latency and power consumption. However, significant challenges remain in scaling the technology and integrating it into practical systems. The presenter expressed some skepticism about the near-term impact of photonic computing compared to quantum computing, which currently has more established roadmaps and industry backing. Nonetheless, the Envise demo marks an important milestone in the development of optical processors, and future advancements could lead to more powerful and compact photonic computing solutions.