If Broadcom is collaborating with OpenAI to develop AI processors, here's a potential visual of the resulting hardware

OpenAI, the leading artificial intelligence research company, has teamed up with Broadcom to develop a custom AI accelerator. This partnership comes as OpenAI looks to enhance its internal AI capabilities and push the boundaries of AI technology.

The architecture of the OpenAI custom AI accelerator bears a resemblance to Broadcom's AMD MI300-series of accelerators. The accelerator will be built using Broadcom's 3.5D eXtreme Dimension System in Package technology (3.5D XDSiP), a technology that is expected to be a key component in the accelerator's design.

Broadcom's 3.5D XDSiP tech allows for the stacking of advanced compute tiles atop a base die containing the chip's lower-level logic and memory controllers. The largest 3.5D XDSiP design, which is slated to begin shipping next year, supports a pair of 3D stacks, two I/O, and up to 12 HBM stacks on a single 6,000 mm2 package.

For the compute architecture of the OpenAI AI accelerator, Broadcom will not be providing the matrix multiply-accumulate (MAC) units. Instead, NVIDIA will supply these essential components.

Package-to-package communication in the OpenAI custom AI accelerator will occur via discrete I/O dies. Broadcom's focus on Ethernet as the protocol of choice for networking paradigms means that OpenAI may not have to rely on Broadcom for its networking needs. Instead, Broadcom's Tomahawk 6 family of switches and co-packaged optical chiplets could be leveraged for scale-up and scale-out networking.

Meanwhile, Broadcom is serving four XPU customers, with one of them having already released production orders. The company's largest customer, worth over $10 billion, is OpenAI, as revealed by Broadcom CEO Hock Tan during Thursday's earnings call.

In a separate development, Apple has also committed to investing $500 billion and hiring 20,000 to bolster its domestic manufacturing capacity. This includes a Texas-based manufacturing plant producing AI-accelerated servers based on its own in-house silicon. Apple is also rumoured to be another company working with Broadcom on custom AI accelerators, with a chip due out in 2026 codenamed "Baltra."

The OpenAI custom AI accelerator, expected to debut next year, will primarily be used internally by OpenAI. This move is likely part of OpenAI's strategy to strengthen its AI capabilities and maintain a competitive edge in the rapidly evolving AI landscape.

If Broadcom is collaborating with OpenAI to develop AI processors, here's a potential visual of the resulting hardware