CPO or Co-Packaged Optics (Silicon Photonics) is the next-generation solution that reduces reliance on copper and harnesses light to transfer signals. These CPOs are packaged alongside hardware accelerators such as GPUs and will be a key solution for next-gen AI factories, offering improved interconnect latency and creating high-bandwidth connections between CPU and GPU.

Both AMD and NVIDIA will be leveraging these technologies for their next-gen AI GPU accelerators. Based on the latest information, AMD will be developing an MRM-based CPO solution for its next-gen Instinct MI500 accelerator. The PIC or Photonic Integrated Circuits will be handed to GlobalFoundries for manufacturing, while ASE will be responsible for the packaging. Last year, AMD acquired photonics experts, Enosemi, to accelerate its Co-Packaged Optics innovations.

Similarly, NVIDIA is also said to be working on its own CPO PIC for Vera Rubin accelerators, using TSMC for the PIC and packaging handled by SPIL. The chips themselves will be assembled at Foxconn Industrial Internet or Industrial Fulian, a subsidiary of Foxconn. For Rubin Ultra, the CPO solution will be prioritized over NPO (Near-Package Optics).

Moving forward, NVIDIA is going to fully adopt Co-Packaged Optics, removing the need for NPOs with its Feynman generation of AI accelerators.

As for AMD, the company has already confirmed that its MI500 series will be fabricated on an advanced 2nm process technology. The upcoming MI400 series makes use of a 2nm process technology too, but the one utilized by MI500 is more advanced, & offers various enhancements. The chips will be produced at TSMC.

The MI500 accelerators will leverage the new CDNA 6 architecture (CDNA 5 for MI400 series), & HBM4E memory, which will offer even higher speeds and memory bandwidth than 19.6TB/s on HBM4-powered MI400 accelerators. Also, unlike previous reports, it looks like AMD isn't going to change to the UDNA architecture naming for Instinct GPUs.

AMD has promised a big jump in AI performance with the Instinct MI500 series. The company is on the trajectory to achieve over 1000x AI performance in just four years. This is crucial to meet the growing AI demand and keep up with the competition, which is also accelerating at a rapid pace. MI500 launches in 2027.

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved