NVIDIA announced a collaboration with AyarLabs at Computex, which is very important for the company’s future graphics architectures. That is, the chips they will launch in the future. Beyond the recently introduced H100 for servers and the upcoming RTX 40. Why are light-based optical interfaces of interest to NVIDIA and what might this mean for the brand’s graphics cards in the future?
Splitting one chip into several with the same functionality combined with each other is the future trend. We’ve talked to you at length about so-called chiplets and it’s more than proven to be the bet for the future, at least from AMD and Intel. But what about NVIDIA? It was speculated that Hopper would follow this trend, but in the end it didn’t. The general feeling, then, is that Jen Hsun Huang’s company lags behind its rivals in this regard. Let’s not forget that it was necessary to create new types of interconnections to overcome a common problem. And it is that in semiconductor materials as in metals, the consumption increases with the length of the wiring. Well, one solution to that is the use of optical interfaces and that seems to be NVIDIA’s gamble.
AyarLabs and NVIDIA agree on the use of optical interfaces on graphics cards
Through a press release, NVIDIA announces collaboration with AyarLabs to create luminous interconnections for future artificial intelligence architectures. A statement that, without the proper context, may seem confusing, but it is very revealing. For some time now, the green company has been naming its graphics chips optimized for high-performance computing, AI chips. So we’re talking about the design of what we call Blackwell and that it will replace the recently introduced H100 chip with the Hopper architecture.
What’s the idea? Well, in the fact that instead of two chips communicating with each other through copper interconnects, optical interconnects. For this it is used a part that converts electrical signals into light as a transmitter. And another that performs the opposite function as a receiver. The advantage of this is that the resistance created by the distance of the wiring disappears completely
In its current form, AyarLabs technology, known as Tera Phyis composed of 8 optical ports with a transfer capacity of 2 TB/s and a power consumption per transmitted bit of 5 pJ. This is much higher than current processor-level interconnects. However, it does give us a clue about NVIDIA’s future plans and application of this technology.
Why does NVIDIA want this technology?
Although the advantage is not really in the reduction of energy consumption. Something that has already been done with the TSMC CoWoS and Intel Foveros packages, except that consumption does not increase with distance. This is critical in HPC systems that use multiple GPUs at the same time. In other words, thanks to this, NVIDIA can create a large graphics card with dozens of chips networked and interconnected with each other via optical interfaces for intercommunication. What do not forget that they can emit hundreds of meters without increasing consumption.
In any case, this should not take us by surprise. Because Nvidia has already expressed its interest in the use of optical interfaces at the level of supercomputers and data centers made up of dozens of interconnected graphics cards. In any case, at the level of a chiplet graphics chip. We believe that just as AMD has based its designs on technology from TSMC, so will NVIDIA.
Finally, optical interfaces are also essential in the development of future video memories. Let’s not forget that your access is what consumes the most in a graphics card. A type of VRAM where its chips use optical ports to communicate would mean a huge step forward in alleviating one of today’s biggest headaches. Something that only NVIDIA has to fix and which also affects its rivals. In any case, we are talking about things in the future and we do not know if this technology will reach the domestic level.