- Recogni Inc. has launched Pareto, a logarithmic number system enhancing AI chip efficiency.
- Pareto simplifies AI computations by converting multiplications into additions, reducing power consumption and execution time.
- The system enables smaller, faster, and more energy-efficient AI chips, which are crucial for handling petaFLOPS-scale operations.
- Pareto outperforms traditional quantized number systems, delivering minimal accuracy loss without model retraining.
- The system has demonstrated over 99.9% relative accuracy in various AI models, consuming significantly less power.
- Available through a seven-nanometer chip from Taiwan Semiconductor Manufacturing Co. Ltd., with broader availability expected soon.
- This launch follows Recogni’s recent $102 million venture capital funding round.
Main AI News:
Recogni Inc., a leader in generative artificial intelligence, has unveiled Pareto, an innovative logarithmic number system designed to enhance AI chip efficiency. Pareto simplifies AI computations by converting multiplications into additions, resulting in smaller, faster, and more energy-efficient chips. This system addresses the growing demand for petaFLOPS-scale operations in generative AI models, significantly reducing power consumption and execution time without compromising accuracy.
As the first logarithmic system to outperform traditional quantized number systems in generative AI inference, Pareto enables more compact chip designs, increases computational capacity in data centers, and reduces costs. It also outperforms standard FP8 and FP16 formats, delivering minimal accuracy loss—less than 0.1% in 16-bit precision and under 1% in eight-bit precision—without requiring model retraining.
Tested on various AI models like Mixtral-8x22B, Llama3-70B, and Falcon-180B, Pareto has demonstrated over 99.9% relative accuracy compared to high-precision baseline models while consuming significantly less power. This efficiency positions Pareto as an optimal choice for modern AI chip design, enabling faster, cost-effective deployment of new models with high power efficiency and accuracy.
Currently available through a seven-nanometer chip from Taiwan Semiconductor Manufacturing Co. Ltd., Pareto’s market presence is expected to grow with an upcoming technology partnership. This launch follows Recogni’s recent $102 million venture capital funding round, supported by Celesta Capital, GreatPoint Ventures Management, and BMW i Ventures GmbH.
Conclusion:Â
Recogni Inc.’s introduction of Pareto represents a significant advancement in AI chip design, offering a powerful solution to the challenges of high computational demand and energy consumption. By enabling more efficient AI computations, Pareto positions Recogni as a key player in the AI hardware market, potentially setting new standards for performance and cost-effectiveness. This development will likely intensify competition among AI chip manufacturers, driving innovation and potentially reshaping the landscape of data center technology and AI inference applications.
