As Trillion dollar companies spend hundreds of billions on AI (Artificial Intelligence) infrastructure, the hardware requirements are becoming ever more problematic.

Patrick Kennedy was able to photograph the XAI Colossus Data Center [1]. The sheer volume of optical fiber required highlights both the importance of East-West communications, but also the limits of existing fiber communication. Figure 1 shows the backside of the racks where multiple 400 Gb/s electrical signals are converted to 3.6Tb/s of optical bandwidth per GPU compute server. This approach is clearly not scalable.

NVIDIA has posted a YouTube video [3] highlighting the various elements of their next generation optical switch. Clearly, more than 100 Terabytes/sec is an impressive result, but consider that bandwidth management requires 400 billion transistors. The lasers, now moved to separate packages are only 8 lasers per package. Presumably this is limited by low yield from the process of flip-chipping laser die while maintaining optical as well as electrical alignment.

Nanorods have some unique features that make them desirable candidates for AI. Nanorod LEDs [5], vertically emitting lasers [6] and cleaved edge emitting geometries [7] have all been fabricated. Their small size makes them less susceptible to dislocations with dimensions.

References

1. https://www.supermicro.org.cn/CaseStudies/Success_Story_xAI_Colossus_Cluster.pdf
2. https://developer.nvidia.com/blog/scaling-ai-factories-with-co-packaged-optics-for-better-power-efficiency/
3. https://www.youtube.com/watch?v=kS8r7UcexJU
4. Y. Wu, Y. et al, Light, Sci. Appl., vol. 11, no. 1, p. 294, 2022.
5. S. Arafin, et al Nanophotonics, vol. 7, no. 1, 2013, Art. no. 074599.
6. E. Stark, et al, Proc. IEEE Photon. Conf., Oct. 2014, pp. 591–592.
7. S. Deshpande, et al, ACS Photon., vol. 4, no. 3, pp. 695–702, Mar. 2017.
8. H. Sekiguchi, et al, Appl. Phys. Lett., vol. 96, no. 23, 2010, Art. no. 231104.
9. D. R. Dykaar, et al, IEEE EDL, vol. 45, no. 5, pp. 789-792, 2024.