Although it is a recurring theme, we cannot forget a reality when talking about future processors. What they do is process data, but said data must move at a sufficient speed so that it can be processed at an adequate speed. The current problem? Moving data has become much more expensive to process.
Why is photonics necessary for communication between chiplets?
The problem in data transmission becomes clear if we take into account the distance of the memory or cache that stores data from the unit that has to process it. The more distance from the cable, the greater the energy consumption when transmitting this data and the use of new nodes does not serve to alleviate this.
If we take into account the configurations based on chiplets and we think about the enormous bandwidth that is necessary to communicate not only several CPUs but also several GPUs then the consumption problem is triggered in the face of communication and more if we take into account that there will be an increase in the distance in the cabling, which translates into an increase in energy consumption only in the transfer of data.
Therefore, it is necessary to develop solutions to this, especially with regard to solutions such as graphics and artificial intelligence where a large volume of data is used. Market that by the way is the main one of NVIDIA. That is why, for example, they have in development a version of their NVLink that consumes half per transmitted bit based on photonics. Its main use? It is clear that it is for your future NVIDIA Hopper, to intercommunicate the chiplets of the same one and that is that the advantage of the interfaces based on photonics is that they do not increase their consumption with distance.
While Intel and AMD have developed 3DIC solutions for the intercommunication of chiplets. NVIDIA’s solution seems to go through the use of optical transceivers and therefore photonics to communicate their chiplets. Something that for the moment they have implemented in none of their current GPUs.
The market for photonics transceivers will grow
The main advantage of photonic transceivers is that despite being a new type of communication interface, it is not as expensive to implement as advanced packaging systems that are dependent on certain foundries and are very expensive to manufacture. Since its complexity makes manufacturing much slower, it reduces production capacity and potentially increases the number of failures that can occur in manufacturing.
This will make the market for integrated optical transceivers grow considerably in the coming years as it is one of the pieces that will be used for the communication of chiplets in graphics processors and for AI. NVIDIA seems to be the best known exponent, but the problem of energy consumption when communicating chiplets is not unique to NVIDIA and is shared by other designers and manufacturers.