Revolutionizing AI: How Sageance AI Plans to Slash Power Consumption with Analog Technology

Summary:

Sageance AI aims to reduce generative AI power needs by 90% using analog circuits.
Claims to run Llama 2-70B at 10% power of traditional systems like Nvidia H100.
Focus on minimizing data movement and leveraging fundamental physics for computations.
Innovative solutions address variability in conductance cells and improve efficiency.
First product expected in 2025, targeting vision systems before expanding to generative AI.

Sageance AI is on a mission to significantly reduce the power needs of generative AI models. By utilizing flash memory cells and analog circuits, they claim their technology can run the Llama 2-70B language model at 10% of the power used by traditional systems like the Nvidia H100. This innovative approach not only promises energy savings but also reduces costs and space requirements by 90%.

The Vision Behind Sageance AI

Sageance's CEO, Vishal Sarin, founded the company in 2018 with the foresight that power consumption would be a critical barrier to widespread AI adoption. As generative AI models continue to grow, the urgency for efficient power solutions has never been more pressing.

How Analog AI Works

The efficiency of analog AI hinges on two main advantages:

It minimizes data movement, thus saving energy.
It employs fundamental physics principles to perform essential computations in machine learning, particularly in the multiply and accumulate operations.

Sageance leverages Ohm’s Law and Kirchhoff’s Current Law to streamline computations, embedding neural network parameters directly within computing circuits, which reduces the need for energy-intensive data transfers.

Analog AI Process Digital data is converted to analog voltages, processed, and then converted back to digital data.

Challenges and Innovations

Historically, analog AI has faced significant challenges, including variability in conductance cells and the need for analog-to-digital converters. Sageance addresses these issues with innovative calibration algorithms and low-power circuitry, enhancing the reliability and efficiency of the analog approach.

Future Developments

Sageance plans to launch its first product in 2025, targeting vision systems as a stepping stone before expanding into generative AI applications. Future systems will feature 3D-stacked analog chiplets designed for high performance and low power consumption, potentially transforming the landscape of AI technology.

Future systems will comprise vertically stacked analog chips linked through advanced interconnects.