This transcript discusses the concept and current state of photonic computing, contrasting it with traditional electrical computing. The speaker highlights the theoretical advantages of light-based computation, such as near-instantaneous speed, but also addresses significant challenges, including the physical size of optical components and the complexities of converting data between electrical and optical formats. The video then focuses on Lightmatter's "Envise" optical chiplet, showcasing a demonstration from Supercomputing 2024 where the photonic processor, leveraging a basic convolutional neural network, acts as an AI player for an Atari game. While acknowledging this as an impressive early step, the speaker expresses skepticism about the immediate scalability and market readiness of photonic computing compared to quantum computing, suggesting it remains in its "baby steps" phase despite its potential for ultra-low latency and power AI.
Summarizing Photonic Computing: Lightmatter's Envise Chip and Atari AI
0:00 - The Evolution of Computing: Traditionally, computing has relied on electricity and transistors for operations. Photonic computing, using light, offers the potential for near-instantaneous calculations due to light-speed interactions.
0:55 - Lightmatter's Envise Chip: Lightmatter, initially focused on optical compute before pivoting to networking, developed the "Envise" optical chiplet. This chip uses optical interposer technology to combine light for calculations like addition and multiplication.
1:52 - Challenges of Optical Compute: Unlike tiny nanometer-scale transistors, optical computing structures require larger sizes (microns or hundreds of nanometers). While power consumption per operation is very low (primarily laser power), challenges remain in manufacturing at scale and efficiently converting data between electrical and optical forms.
2:51 - Supercomputing 2024 Demo Revealed: The speaker, having photographed the demo at Supercomputing 2024 (November), is now permitted to share details following a related paper's publication in Nature.
3:19 - The Envise Package: The Envise chip itself resembles a standard BGA processor but features fiber optic connections. It's designed to be integrated as a PCIe accelerator card, with on-card lasers providing the necessary light input.
4:26 - The "Atari" Demo: Lightmatter's demo showcased the Envise card playing an Atari game. Crucially, the photonic hardware isn't running the game itself but acts as an AI player. The game's video output is fed into a neural network model on the Envise chip, which then determines and sends back the optimal action (e.g., joystick movements) for the next frame.
7:02 - Neural Network Architecture: The AI player uses a basic convolutional neural network (CNN) processing 84x84 pixel input at 4-bit color depth. This network generates one of nine possible outputs, corresponding to various in-game actions like left, right, or fire.
8:29 - Early Stages and Future Outlook: The speaker emphasizes that photonic computing is still in its "early baby steps." While it holds promise for ultra-low latency and ultra-low power AI, the speaker expresses more immediate optimism for quantum computing due to more developed roadmaps and established companies in that field.
9:08 - Speculation on Future Form Factors: The speaker concludes by humorously pondering a future "optical cube of compute" on their desk.