Photo by Shiva Smyth on Pexels

Amid ongoing US semiconductor restrictions, China is turning to chip stacking as a potential solution to enhance domestic chip performance. This technique involves combining multiple, older chips produced within China to emulate the capabilities of more advanced GPUs that are now inaccessible. While promising architectural innovation, significant hurdles remain in bridging the performance gap with industry leaders like Nvidia. Industry sources suggest that Nvidia’s A100 GPU can still deliver up to 2.5 times the performance of systems built with the stacked chip approach.

By focusing on system architecture and software optimization, chip stacking offers a potential avenue to bypass the CUDA dependency that locks many players into Nvidia’s ecosystem. Key challenges persist, including managing thermal dissipation and developing a robust software ecosystem to support the new architecture. Nonetheless, the approach shows promise for specific, memory-intensive applications such as AI inference. This pursuit of architectural alternatives could ultimately reshape the definition of an “AI chip” itself.