AI startup Recogni unveils new computing method to slash costs, power requirements

AI startup Recogni unveils new computing method to slash costs, power requirements

Technology

It develops specialized chips and software to enable AI inferencing

Follow on
Follow us on Google News
 

(Reuters) - AI chip and software startup Recogni unveiled a novel computing method on Tuesday that could make its chips used to train and run artificial intelligence systems smaller, faster and less expensive to operate.

Backed by BMW (BMWG.DE) Bosch and venture capital firm Mayfield, Recogni develops specialized chips and software to enable AI inferencing - the process of trained AI models making predictions or decisions on new, unseen data.

The company said the new patented system, called Pareto, utilizes a logarithmic approach that outperforms existing methods when running large AI models.

"It is a huge leap in all of the KPIs (key performance indicators) that influence silicon hardware system design when it comes to AI computing," Recogni's co-founder and VP of AI, Gilles Backhus told Reuters.

Current AI models, such as OpenAI's GPT-4 and Google's Gemini, require hundreds of thousands of power-hungry mathematical operations for the simple of prompts on chatbots like ChatGPT.

Recogni said that its new system converts these multiplication operations into additions, significantly reducing power consumption while maintaining accuracy.

The startup said it has already tested Pareto on AI models developed by Meta Platforms, Stability AI, and others.

Recogni, whose first chip was designed, manufactured Taiwan Semiconductor Manufacturing Co's (2330.TW) seven nanometer process, said it was working with an unnamed partner to make Pareto more widely available and will announce the partnership in the coming months.

"We are speaking to companies that are putting hardware in data centers and offering it to the world to whoever wants to basically rent it ... that's definitely one of the deployment routes that we're considering," Backhus added.