Generative artificial intelligence company Recogni Inc. today announced a logarithmic number system for AI that delivers benefits that it says include low power, high computer density and low latency without compromising accuracy.
Called Pareto, the number system provides benefits for all AI chip design criteria by simplifying AI compute. It does show by turning multiplications into additions, resulting in AI chips that are smaller, faster and less energy-hungry.
The problem the system is seeking to address is one where the latest generative AI models demand multiplications and additions on the order of petaFLOPS, or quadrillions of operations per second, posing challenges in power consumption and computational speed. Pareto addresses the challenge by converting multiplications into additions, significantly reducing power usage and execution time without compromising accuracy.
Recogni says it’s the first to market with a logarithmic system that outperforms other quantized number systems for generative AI inference.
Pareto’s efficiency enables a more compact chip design and hence allows for significantly increased compute in data centers while reducing costs. The numbering system also reduces power consumption and outperforms traditional FP8 and FP16 formats — formats that define the precision levels for floating-point numbers in computing.
AI models using Pareto are also said to experience minimal accuracy loss, with less than 0.1% drop in 16-bit precision and under 1% in eight-bit precision, all without requiring retraining.
“By turning multiplications into additions, Pareto significantly reduces power consumption, latency and chip size, making it the optimal choice for modern AI chip design,” said Chief Executive Marc Bolitho. “Organizations running gen AI inference can now keep operating costs lower than any other technology and ensure uncompromised AI model quality for the widest variety of multimodal gen AI Inference applications and use cases.”
Pareto has undergone extensive testing on various AI models with some highly impressive results. Testing on Mixtral-8x22B, Llama3-70B, Falcon-180B, Stable Diffusion XL and Llama3.1-405B shows that Pareto achieves a relative accuracy of over 99.9% compared to the trained high-precision baseline model, while consuming significantly less power.
“With Pareto we came up with a number system that allows businesses to instantly deploy their models at high power efficiency with virtually no loss across all key performance and accuracy metrics,” said Gilles Backhus, founder and vice president of AI at Recogni. “While companies using standard math are spending considerable time converting models to lower precision to reduce the power and operational expenses, Pareto allows companies to bring new models to production faster and cheaper while maintaining high accuracy.”
Pareto is now available through a seven-nanometer chip manufactured by Taiwan Semiconductor Manufacturing Co. Ltd. Recogni also plans to announce a technology partnership that will make Pareto more widely available in the coming months.
The company was previously in the news in February, when it raised $102 million in additional venture capital funding. Investors include Celesta Capital, GreatPoint Ventures Management, Mayfield Fund, DNS Capital, BMW i Ventures GmbH and Tasaru Mobility Investments.