Bandwidth limitations have become a major bottleneck in AI and high-performance computing (HPC), as GPUs are underutilized due to bandwidth limitations, losing almost half of their computing power.
Nvidia is not expected to release optical connections for its NVLink protocol until the “Rubin Ultra” GPU computing engine is launched in 2027.
This slowdown has prompted hyperscalers and cloud builders to explore ways to leapfrog Nvidia’s technology by adopting optical connections earlier.
Meet ChromX
Xscape Photonicsan optical interconnect company born out of research at Columbia University, uses photonics to realize scalable, energy-sustainable, and cost-effective high-bandwidth solutions to power next-generation AI, ML, and simulation hardware.
This could help the AI industry save billions of dollars in wasted GPU capacity while providing a path to greener, more sustainable AI infrastructures.
The next platform recently took a closer look at Xscape Photonics and spoke to the team behind it, including CEO Vivek Raghunathan, a former MIT researcher and Intel engineer.
Highlighting the inefficiency of current GPU systems, Raghunathan explained that as scaling continues, the problem is shifting “from GPU device-level performance to a system-level network problem.”
This is where Xscape’s technology comes into play. By converting electrical signals to optical signals directly within the GPU, Xscape can dramatically increase bandwidth while reducing power consumption.
The startup’s solution, called the ‘ChromX’ platform, uses a laser that can transmit multiple wavelengths of light simultaneously through a single optical fiber – up to 128 different wavelengths (or ‘colours’). This enables a 32-fold increase in bandwidth compared to lasers that use only four wavelengths.
The ChromX platform also relies on simpler modulation schemes such as NRZ (Non-Return-to-Zero), which reduce latency compared to higher-order schemes such as PAM-4 used in other systems such as InfiniBand and Ethernet. The ChromX platform is programmable, allowing it to tailor the number of wavelengths to the specific needs of an AI workload, whether for training or inference tasks.
Raghunathan narrated The next platformTimothy Prickett Morgan: “The vision is to match the in-packet communications bandwidth with the escape bandwidth for out-of-packet communications. And we think if we use our multicolor approach, we can match that, so that giant data centers – or multiple data centers – behave like one big GPU.”
The potential impact of this technology is enormous. AI workloads consume enormous amounts of energy, and with demand for data centers expected to triple by 2035, power grids will struggle to keep up. Xscape Photonics’ innovations can provide an essential solution, allowing AI systems to function more efficiently and sustainably.