Artificial intelligence is one of the disciplines that moves the most data when it comes to running on the hardware of our PCs, this involves enormous bandwidths, which are closely linked to the energy consumption of data transfer. In an age when information movement is more expensive than information processing, this has led to the world’s highest bandwidth.
However, looking to the future, we may have to opt for different technologies than usual in order to reduce energy consumption per transmitted bit. Not only advanced intercom systems based on advanced packaging based on 2.5DIC and 3DIC, as well as the use of photonic interfaces. It is precisely the use of photonics that promises the most advances in terms of artificial intelligence, especially in a future where integrated circuits will not be in a single chip, but in several intercommunicating with each other.
Lightmatter, AI processor more efficient than an NVIDIA GPU
There is no question that NVIDIA GPUs are the most powerful type of processor for artificial intelligence out there right now. It is precisely this discipline that has dictated the advances of the latest NVIDIA architectures. However, a startup named Lightmatter has developed a processor that can be considered a generational leap compared to NVIDIA GPUs. The reason? The chip of this company, called Lightmatter uses photonics for the communication of information.
The big problem regarding the development of future processors for artificial intelligence is that within the same processor where data shipments are very local, consumption is very low, but if we move away then consumption increases. This problem does not occur with intercommunications based on photons instead of electrons, since the energy consumption in the optical interfaces does not increase with the distance from which the data is located. And it is in photonics that the people of Lightmatter have bet to create their processors.
NVIDIA is not going to be left behind
As GPUs and other chips become more complex and make the leap to the chiplet paradigm, the use of different types of interfaces to mitigate the increase in energy consumption becomes a reality. Something that the people of Lightmatter want to solve before anyone else. However, NVIDIA and other companies are not idle and have been developing communication interfaces based on the transmission of data through light for years. Which we’ll see in the server market first, but for now the people at Lightmatter have gotten ahead of NVIDIA in developing a GPU that communicates using optical interfaces.
In any case, NVIDIA has not been left behind, since we know that they have developed a version of their NVLink interface that transmits information using light. Which they presented a few months ago. We’ll see how all this plays out, if at all. Will we see it on our gaming PCs? Who knows, but we will still have to wait a few years to see it.
