We first saw the news last July that OpenAI was looking to develop its own AI chip, and that the company was in talks with Broadcom regarding the matter. Those talks seem to have gone well, as more recent reports have indicated that the ChatGPT company is indeed going ahead with its plans for its own custom chips. The first custom chip design will be finalised within the year.
Reuters reports that said in-house chip design will be finalised “in the next few months”, and that TSMC, previously listed as a potential partner, will be the one handling the fabrication of said chips. These will be made using the 3nm process, and if things go according to plan, OpenAI will be able to test them as an alternative to NVIDIA’s.
On one hand, the report notes that OpenAI’s custom chip is capable of both training and running AI models. That being said, sources cited claim that it “will initially be deployed on a more limited scale, and primarily for running AI models”. Similarly, it will also “have a limited role within the company’s infrastructure”.
Which is probably a short-term thing, as the sources cited in the report also claims that a new chip design such has this could cost US$500 million (~RM2.23 billion). Making the necessary software and peripherals for it could double said cost.
(Source: Reuters)
