Good morning


Interconnects in all computer systems are increasingly critical. In particular, in interposer systems, in 2D or 3D packing and chip-to-chip interconnects, are essential for achieving seamless coupling of all processing and memory elements.

NVIDIA, IBM & Intel have mastered to connect optically 16 silicon GPU chips in a Multi-Chip Module (state-of-the-art & commercially available).
Laser used are VCSEL’s meaning direct-modulated (switch on/off as Vivek & Suresh have explained)

Ok - so what could a higher integrated photonics system do for Energy Efficiency (pJ/bit, lower numbers win) and higher bits rates per second (Gbps higher numbers win) ?

A case study by University of Chicago & University of Illinois at Urbana–Champaign answers some of this questions - please

Table 1: High-Speed Link Technologies (measurement results from industry designs and research prototypes).

 *TL (transistor lasers)

In this case study, we apply a new « Rome architectures » to Multi-Chip Module GPU (MCM-GPU) systems to evaluate network performance and power. Our results show that « Rome » can efficiently scale MCM-GPUs with up to 1024 streaming multiprocessors (SMs), which is >16×larger than state-of-the-art commercial GPUs (e.g., Nvidia Pascal) and 4×larger than existing research proposals. 
We also quantitatively compare « Rome » with different interconnect technologies and architectures. Our results show that, compared to state-of-the-art electrical 2D mesh, « Rome » reduces network power by up to 62% (and up to 4× with more advanced laser technologies) while improving application performance by up to 143%. Compared to architectures designed for silicon photonics, « Rome » reduces network power by >2.5× while achieving comparable application performance.

Amazing metrics, next question can it be produced at scale, well you may know that answer, …. mhhh

Cheers