Date/Time: Tuesday, January 30, 2024 | 8:00 – 9:00 AM PDT
Senior Director, Product Management
Silicon Photonics Business Line
Artificial Intelligence and Generative Models (GenAI) require the clustering of tens of thousands of Graphics Processing Units (GPU) into compute units. Accompanying these GPUs are banks of High Bandwidth Memory (HBM) chips. In addition to communicating with the HBM chips, GPUs must communicate with each other and to other clusters using low power, low latency, and high bandwidth connections.
Generative Models are expected to grow logarithmically in the number of parameters adding to the need for more memory and faster, low latency interconnects. For example, Generative Pre-trained Transformer (GPT) 4 will potentially have trillions of parameters as opposed to its predecessor GPT3’s 175 billion.
One of the several challenges of the increased demand for memory capacity and interconnection between GPUs and between clusters, is the trade-off of the GPU chip perimeter, also known as “beach front”, consumed by the input/output (IO) circuits for the memory interconnect vs. those consumed by the network interconnect.
Silicon Photonics has the ability to address the interconnect challenges.
What is Silicon Photonics?
Photonics has been around since the 1980s with the advent of optical fibers for data transmission. The technology was based on discrete active (using compound semiconductors such as indium phosphide and gallium arsenide) and passive components (such as lenses, filters) assembled into hermetically sealed packages.
Silicon Photonics is a technology that uses silicon and its associated materials such as silicon dioxide and silicon nitride to create photonic circuits. It leverages the existing silicon-based fabs and equipment infrastructure to produce photonic circuits at wafer level. This results in higher levels of circuit integration compared to the discrete components of the past.
Wave Division Multiplexing (WDM), the ability to transmit large amounts of data using discrete optical wavelengths on a single optical fiber, is now possible using a single silicon photonics chip.
This capability to efficiently transmit high speed data at very low energy consumption (<1 pJ/bit) through almost lossless optical fibers, is what the AI/GenAI industry is looking for. This is the niche of Silicon Photonics.
Join us for a webinar to learn more about Silicon Photonics, its application to Generative AI, and the ecosystem required to reach its full potential.