Designed in partnership with semiconductor manufacturer Broadcom, OpenAI’s first chip is designed to work with models from the creator of ChatGPT and those from other companies. Tested internally, it would provide higher performance per Watt than current chips.
A new competitor for Nvidia? OpenAI, the parent company of ChatGPT, unveiled its first internally developed chip on Wednesday, named “Jalapeño” in reference to the chili pepper of Mexican origin. It is designed to make its artificial intelligence model work faster and cheaper.
Developed with semiconductor and infrastructure software maker Broadcom, the chip was not developed for the time-consuming and expensive process of training a model from scratch, but for so-called AI inference. This process allows an AI model to operate in a way that provides answers to users.
“OpenAI is currently conducting final tests, but the first results show that Jalapeño will offer much better performance per watt than current chips,” assured the group in a blog post. He doesn’t give figures, but plans to publish a full technical report about his chip in the coming months.
No technical details have been given yet, but we nevertheless know that Jalapeño is an ASIC chip, for Application-Specific Integrated Circuit, or Application-Specific Integrated Circuit, in other words an electronic chip designed to perform a task precise. Where a processor (CPU) is very versatile, a GPU capable of carrying out numerous calculations in parallel and an FPGA chip is programmable to adapt to a specific task, an ASIC chip only does one thing but in an extremely efficient manner.
“Increase the availability of computing power”
The company emphasized that its own AI models helped in the development of the chip, reducing the time needed to develop high-performance semiconductors. The final development would have only required nine months. The Jalapeño chip was designed to work with a wide choice of AI models and not just its own.
“Jalapeño is part of our long-term full-stack infrastructure strategy to increase the availability of computing power, which will result in faster, more reliable and more affordable AI for individuals and businesses,” said Greg Brockman, co-founder and president of OpenAI.
The young company believes that this will allow a greater number of people, particularly students, to use cutting-edge models on a daily basis and thus learn, create or solve complex problems. Its chip will be deployed in data centers operated by Microsoft and other partners starting this year.
Divest from Nvidia
But by having its own chip model, OpenAI seeks above all to reduce its dependence on external suppliers, starting with Nvidia. Chips are indeed the basis of how AI works and most companies in the sector depend on those developed by Jensen Huang’s company and a handful of other suppliers.
And OpenAI is not the only one who wants to get rid of it. Google, Amazon and Microsoft have also sought to develop their own chips to reduce costs and increase performance in recent years. Those of Nvidia, initially designed for video games, are particularly adapted for the significant computing power necessary for training AI models, which has currently made it the most valuable company in the world.
