The new AI-purposed numerical system offers the unparalleled benefits of low power consumption, high computational density and low latency without compromising accuracy.
San Jose, California, August 20, 2024 SAN FRANCISCO, Oct. 13, 2020 /PRNewswire/ — Generative AI inference company Recogni Inc. today announced Pareto, a patented logarithm system for AI that delivers benefits to all major AI chip design criteria without compromise. Designed to change the way GenAI operates, Pareto dramatically simplifies AI computing by converting multiplications into additions, making Recogni’s chips smaller, faster and less power-hungry.
We help hyperscalers, computing providers and enterprises get the most out of compute, space and energy.
Modern GenAI models require multiplications and additions on the order of petaflops, creating challenges in power consumption and computation speed. Pareto addresses these challenges by effectively converting multiplications into additions, significantly reducing power usage and execution time without compromising accuracy. Recogni was the first to market to introduce a logarithm system that outperforms other quantized number systems for GenAI inference. Years of research have culminated in a solution that provides the following capabilities:
Smaller chip size: Pareto efficiency enables more compact chip designs, significantly increasing computational power and reducing costs in data centers. Lower power consumption: Pareto reduces the power requirements of AI models and outperforms traditional FP8 and FP16 formats, enabling sustainable AI computing. High accuracy: AI models using Pareto exhibit less than 0.1% accuracy degradation at 16-bit precision and less than 1% accuracy at 8-bit precision without the need for retraining.
“With Pareto, we are accelerating the world’s AI ambitions.” Mark Bolitho“Pareto’s logarithmic system delivers the lowest average error and best performance for AI models. By converting multiplications to additions, Pareto dramatically reduces power consumption, latency and chip size, making it the best choice for modern AI chip designs,” said Dr. Tim Allen, CEO of Recogni. “Organizations running GenAI inference can now achieve lower operational costs than any other technology and ensure uncompromising AI model quality for the most diverse, multi-modal GenAI inference applications and use cases.”
Helping developers produce new models in less time and with greater accuracy
Extensive testing on various AI models including Mixtral-8x22B and Llama370BFalcon-180B, Stable Diffusion XL, and Llama 3.1-405B Pareto shows that compared to the trained high-precision baseline model, it achieves a relative accuracy of over 99.9% with significantly less power consumption.
Pareto offers FP16 accuracy at lower power consumption than competitors’ FP8, enabling developers to quickly and efficiently deploy trained models. Pareto can be converted instantly with little loss of accuracy, eliminating the need for time-consuming retraining processes.
“Our goal is, and always has been, to directly address the needs of both enterprises and machine learning developers,” he said. Jill Buckhus“With Pareto, we’ve devised a numerical system that allows companies to instantly deploy models with high power efficiency, with virtually no loss in all key performance and accuracy metrics,” said , founder and vice president of AI at Recogni. “Companies using standard math spend significant time converting models to lower precision to reduce power and operational costs, but with Pareto, companies can put new models into production faster and cheaper while maintaining high accuracy.”
Seven years of development to change the way GenAI operates
Recogni has demonstrated the Pareto advantage: Their first chip, designed and manufactured using 7nm TSMC with their first generation logarithmic math, exceeded performance expectations and proved every hypothesis.
In the coming months, Recogni plans to announce technology partnerships that will make the power of Pareto even more widely available and accelerate a new wave of advancements.
Join Recogni at NeurIPS 2024 to learn more about Pareto charts.
About Recogni
Go global North America and EuropeRecogni is building the most compute-dense, energy-efficient generative AI inference system to accelerate the world’s AI ambitions. Recogni is achieving its mission with backing from GreatPoint Ventures, Celesta Capital, Mayfield and DNS Capital. To learn more, visit www.recogni.com.
SOURCE Recogni Inc.
