The American company Avicena presented the LightBundle high-parallel optical interconnect technology designed for communication between microcircuits at distances up to 10 meters. According to the developers, this technology can find application in distributed systems, means of disaggregating processors and memory, and other advanced solutions for high performance computing.
At the heart of the LightBundle are High Speed GaN Microemitter Arrays (CROME), which are built using the microLED display manufacturing ecosystem. The technology is fully compatible with high-performance silicon chips.
Interconnections are becoming a key bottleneck in computing and networking systems. Highly variable workloads drive the development of tightly interconnected, heterogeneous software-defined clusters of central processing units, GPUs, data processing units, and shared memory units. Artificial intelligence and machine learning are examples of new applications driving the growing demand for extremely high density, low power and low latency interconnects.
The Avicena LightBundle is said to outperform competing technologies in energy efficiency and bandwidth density by 10 to 100 times. More specifically, energy efficiency reaches 0.1 pJ / bit, bandwidth is 10 Tbit / s / mm2… It is also important that optical channels operate reliably at temperatures from –40 ° C to + 150 ° C, and are characterized by high reliability.
The manufacturer demonstrated the operation of an array of 200 CROME devices installed at 30 μm pitch, connected to an array of photonic converters using a multicore optical fiber. The data transfer rate in each line was 10 Gbps, that is, the total bandwidth of the channel reached 2 Tbps.
The parallel nature of LightBundle technology works well with parallel chiplet interfaces such as AIB, HBI, and BoW, and can also be used to expand the capabilities of the widespread PCIe, NVLink, DDR and GDDR interfaces, providing low latency communications.