Fiber optics as an alternative to DRAM for AI?

John Carmack recently raised an interesting question: the use of a fiber optic line as an L2 cache for streaming AI data. The idea, seemingly futuristic, is based on the physical properties of optical fiber and its potential to offer a combination of high bandwidth and low latency.

The use of fast memories is a known bottleneck in the field of artificial intelligence, especially when it comes to managing large models and continuous data streams. DRAM, while a consolidated technology, has limitations in terms of capacity and speed.

Carmack's idea is still conceptual, but it opens up interesting scenarios on the evolution of hardware architectures for AI. It remains to be seen whether this idea will translate into practical and implementable solutions, but the debate is open.