15 Mar 2026
16:00 - 18:30
Room 515B
CPO has been an exotic communication paradigm for over a decade now, and now appears to be in the early stages of deployment by a few large companies and a few start-ups. The integration of CPO with ASICs offers a direct way for digital data to access the fiber without going through power-hungry and expensive electronic-to-optical conversion modules (i.e. pluggable optics). For AI, GPUs need to be connected together within a pod to create a large high bandwidth memory (HBM)-centric paradigm and to facilitate an efficient attention-centric execution of large parameter count models. As large language models (LLMs) grow in parameter count, the number of GPUs in a pod needs to increase in the scale up domain as increasing HBM is bounded by package size. To interconnect GPUs in the scale up (intra-pod) or scale out (for horizontal scalability) domain, there is a need for a high-bandwidth solution which would consume low power, integrate well with the ASIC/interposer, and support a plethora of network protocols while being low cost and reliable at the same time. In this quest, CPO turns out to be an apt solution for integration with both GPUs, CPUs, and switch ASICs in facilitating AI pipelines however there are many implementation issues that still need to be addressed.
The key questions to address in this workshop are:
Mass Adoption:
- Which is more important: power consumption or cost?
- What kind of beachfront densities are needed to support CPO for AI?
- Is CPO reliability within the bounds of AI infrastructure? (blast radius, impact of ELSFP, fiber shuffle, wavelength sparing, fiber attach)
Packaging:
- What CPO packaging technologies are most compatible with GPUs and switches?
- I.e. Soldered vs socketed and pigtailed vs detachable?
- What are ways to solve system-wide challenges (i.e. fiber cabling, link flapping, serviceability) while integrating CPO? At the pod, rack, data-center level?
Underlying Technology:
- From a SerDes perspective, what is a good client SerDes intercept – 224G or 448G?
- Does WDM make sense for CPO if it impacts the radix?
- Is there a benefit to using CPO with comb lasers (>16 λs)?
- For the optical modulator, which of the MZI and MRM technologies would be more efficient, cheaper and easier to implement?
The workshop will open with a keynote by an industry luminary, followed by two sessions of invited speakers and concludes with a Slido inspired panel with the attendees.
Organizers
-
Ashwin Gumaste
Microsoft, United States
-
Takako Hirokawa
GlobalFoundries, United States
-
Daniel Kuchta
NVIDIA, United States
-
Hari Potluri
Broadcom, United States