OpenAI, the developer of ChatGPT, was reported to have been pushing to secure alternative chips out of dissatisfaction with the performance of Nvidia's latest artificial intelligence (AI) chips, but CEO Sam Altman strongly denied it.
Reuters reported on the 2nd, citing multiple sources, that OpenAI has been searching since last year for alternatives to Nvidia graphics processing units (GPUs) for inference. According to the sources, OpenAI is said to be dissatisfied with the response speed of ChatGPT running on Nvidia chips in certain areas such as coding and other software development, and communication between AI and software.
Accordingly, OpenAI is reportedly reviewing a plan to meet about 10% of its overall future inference computing demand with alternative products. In particular, it sees latency occurring in communication between GPUs and external high-bandwidth memory (HBM) and is considering as an alternative a design that integrates SRAM on the chip to boost memory access speed.
In fact, OpenAI on Jan. 14 signed a $10 billion chip supply deal with semiconductor startup Cerebras, which adopts this approach. It also held talks with Groq, a chipmaker developing SRAM-based chips, but negotiations were reportedly halted after Nvidia signed a $20 billion licensing deal with Groq in December last year. Separately, OpenAI is said to be preparing to develop its own chips in collaboration with Broadcom to reduce its reliance on Nvidia.
In response, Nvidia said, "The reason customers choose Nvidia for inference is that we deliver the best performance and total cost of ownership at scale."
CEO Sam Altman also dismissed the report. "We love working with Nvidia, and Nvidia makes the best AI chips in the world," he said. "We hope to remain a huge customer of Nvidia for a very long time." Referring to recent reports of repeated friction between the two companies, he added, "I don't know where all this craziness started."
Earlier, The Wall Street Journal reported that Nvidia CEO Jensen Huang criticized OpenAI's business approach as lacking discipline and put on hold a $100 billion investment in OpenAI announced in September last year. However, Huang denied it the next day and said he would participate in OpenAI's current fundraising round.