The journey to reduce OpenAI’s reliance on NVIDIA and its GPUs starts with designing its custom AI chip, and the company behind ChatGPT has made some decent progress in this regard, assuming the latest report is to be believed. The artificial intelligence startup is said to be in the middle of finalizing the design of the silicon, with the completion said to take a few months. If this route has a minimum amount of potholes, OpenAI could send its in-house chip to TSMC for tape-out in the first half of the year.
The custom AI chip from OpenAI will have a limited role at first and will primarily be tasked with running AI models
While there can be more than a few obstacles standing in the way of OpenAI’s plans, the company seems determined to see its goal to fruition. As reported by CNBC, the tape-out process will take six months to complete and will cost millions and millions of dollars. However, if OpenAI is willing to pay TSMC a premium, the latter may produce the AI chip faster. Sadly, there is no guarantee that the first tape-out will succeed, as a failure will mean another tape-out is required to repeat the process and isolate the problem.
OpenAI was previously reported to leverage TSMC’s A16 Angstrom process for its Sora video generator, but it is unconfirmed if this is the same AI chip whose design is set to be finalized in the coming months or a different in-house solution. This specific division is currently being led by OpenAI’s Richard Ho, with the headcount increased to 40 talented individuals. The in-house chip design process is being assisted by Broadcom, though the exact capacity of the company’s contribution is unknown.
The name of OpenAI’s custom AI chip has yet to be revealed, but its functionality will revolve around training and running artificial intelligence models. The silicon’s capabilities will initially have a limited role, which can increase depending on how many units OpenAI intends to deploy in the future. It is mentioned that if everything goes according to plan, then mass production is estimated to start from 2026, with TSMC utilizing its 3nm technology for the chip, along with systolic array architecture paired with High Bandwidth Memory (HBM) that NVIDIA uses for its own AI GPUs.
Follow Wccftech on Google to get more of our news coverage in your feeds.
