OpenAI wants to reduce its dependence on Nvidia GPUs, which control 80% of the market for chips used for artificial intelligence. That is why, since last year, the start-up has been working on developing its own chip completely adapted to its needs with the help of Broadcom. The small team of 40 people, led by Richard Ho, a Google veteran, has moved quickly: the chip will be sent in the coming months to TSMC to prepare for production, according to Reuters.
Still a lot of work ahead
So be careful, this does not mean that the first OpenAI-branded chips will come off the lines of the Taiwanese giant at the end of this year. The current process called "tape-out" refers to the process of sending an initial chip design to a manufacturing plant. In other words, it is the final stage of silicon design: the design is frozen and sent to production. And OpenAI is still a few months away from finalizing this famous design.
A standard “tape-out” takes about six months to produce a functional chip (unless OpenAI pays to speed up the pace). There is no guarantee that it will work the first time; in the event of a defective chip, the process will have to be started again, which can delay the availability of the first functional units.
OpenAI is also aiming for mass production at TSMC in 2026, after testing and optimization phases. This first chip will first be tested on a small scale, mainly to run AI models. In short, OpenAI is still very far from being able to compete with Nvidia, but also Microsoft, Google or Amazon on this front.
In any case, this internal development will allow OpenAI to more easily negotiate computing capacity with other chip suppliers. The company is also at the helm of the famous $500 billion “Stargate” project, which aims to make AI server data centers flourish in the United States. And why not servers with OpenAI chips…
Source: Reuters
0 Comments