OpenAI Unveils First Custom AI Chip 'Halving' to Cut Nvidia Reliance

OpenAI, the creator of ChatGPT, has officially revealed its first custom-designed chip for artificial intelligence inference, a move aimed at curbing its dependence on Nvidia's dominant graphics processing units (GPUs). Codenamed 'Halving,' the chip was developed in partnership with Broadcom.

The company stated that Halving was designed from the ground up, leveraging OpenAI's deep understanding of large language model (LLM) principles. The chip's architecture focuses on minimizing data transfer bottlenecks and balancing computation, memory, and networking to achieve near-theoretical performance utilization. While still in final testing stages, initial results suggest superior power efficiency compared to current high-end products from competitors, according to OpenAI.

Hawk Tan, CEO of Broadcom, told Reuters that Halving is comparable to Nvidia's Blackwell and Google's Tensor Processing Units (TPUs). The collaboration between OpenAI and Broadcom significantly accelerated the chip development process. The entire cycle, from initial design to the "tape-out" stage where the design is sent for manufacturing, took a remarkably short nine months, setting a new benchmark for high-performance custom chip development.

This initiative represents OpenAI's first tangible step towards addressing the soaring costs of AI inference and alleviating supply chain constraints for GPUs. While Nvidia GPUs will likely remain crucial for AI model training, OpenAI aims to reduce operational expenses and improve supply chain stability for the repetitive, large-scale inference computations through its in-house chip. Google, another major AI player, is also challenging Nvidia's market dominance by offering its custom TPUs beyond its own data centers and cloud services.

“

Halving is on par with Nvidia's Blackwell or Google's Tensor Processing Units.