OpenAI Finalizes Custom AI Chip Design; TSMC Tape-Out Process Expected to Begin in H1 2025

OpenAI Finalizes Custom AI Chip Design; TSMC Tape-Out Process Expected to Begin in H1 2025

OpenAI is actively working to decrease its dependency on NVIDIA and its GPUs by developing a custom AI chip. According to the latest reports, progress on this initiative appears to be promising as the company moves forward with the silicon design, anticipated to be finalized in a few months. If the development goes smoothly, OpenAI aims to send this in-house chip design to TSMC for the tape-out phase in the first half of the upcoming year.

Initial Purpose of OpenAI’s Custom AI Chip

Despite the various challenges that may arise, OpenAI is resolutely committed to achieving this goal. As noted by CNBC, the tape-out process, expected to last six months, could incur significant costs. Nevertheless, if OpenAI opts to pay a premium to TSMC, the production of the AI chip might be expedited. However, it’s important to highlight that the first tape-out attempt could fail, necessitating a repeat process to identify any issues encountered.

Previously, reports indicated that OpenAI was leveraging TSMC’s A16 Angstrom process for its Sora video generator. However, it remains unclear whether the upcoming AI chip design involves this same process or if a different in-house solution is in development. The project is currently being spearheaded by Richard Ho at OpenAI, with the team expanding to 40 skilled individuals. Broadcom is also lending expertise to this in-house chip design, although the extent of their contribution has not been precisely disclosed.

As for the specific name of OpenAI’s custom AI chip, details are still under wraps. The chip is primarily intended for the training and operation of AI models, with its initial functionality being somewhat limited. Should everything proceed as planned, mass production is projected to commence in 2026. TSMC is expected to employ its advanced 3nm technology for the chip, integrating a systolic array architecture alongside High Bandwidth Memory (HBM), similar to the technology utilized in NVIDIA’s AI GPUs.

Source & Images

Leave a Reply

Your email address will not be published. Required fields are marked *