This laser technology could save the planet and help AI industry claw back more than $100 billion — half of GPUs are 'wasted' because of limited bandwidth, and this US startup wants to change that

Trending 1 week ago
Xscape Photonics ChromX
(Image credit: Xscape Photonics)

Bandwidth limitations person go a important bottleneck successful AI and high-performance computing (HPC), arsenic GPUs are underutilized owed to bandwidth constraints, pinch astir half of their computational powerfulness going to waste.

Nvidia is not expected to merchandise optical interconnects for its NVLink protocol until nan "Rubin Ultra" GPU compute motor launches successful 2027.

This hold has led hyperscalers and unreality builders to research ways to leapfrog Nvidia’s exertion by adopting optical interconnects earlier.

Introducing ChromX

Xscape Photonics, an optical interconnect institution spun retired of investigation astatine Columbia University, is utilizing photonics to recognize scalable, high-bandwidth, energy-sustainable, and cost-effective solutions to alteration nan adjacent procreation of AI, ML, and simulation hardware.

This could thief nan AI manufacture prevention billions of dollars successful wasted GPU capacity while besides offering a way to greener, much sustainable AI infrastructures.

The Next Platform precocious took a person look astatine Xscape Photonics and said pinch nan squad down it, including CEO Vivek Raghunathan, a erstwhile MIT interrogator and Intel engineer.

Raghunathan highlighted nan inefficiencies of existent GPU systems, explaining that arsenic scaling continues, nan problem shifts "from GPU device-level capacity to a system-level networking problem."

Sign up to nan TechRadar Pro newsletter to get each nan apical news, opinion, features and guidance your business needs to succeed!

This is wherever Xscape’s exertion comes into play. By converting electrical signals into optical ones straight wrong nan GPU, Xscape tin dramatically summation bandwidth while simultaneously reducing powerfulness consumption.

The startup’s solution, called nan "ChromX" platform, uses a laser that tin transmit aggregate wavelengths of ray simultaneously done a azygous optical fibre - up to 128 different wavelengths (or "colors"). This enables a 32-fold summation successful bandwidth compared to lasers that usage only 4 wavelengths.

The ChromX level besides relies connected simpler modulation schemes for illustration NRZ (Non-Return-to-Zero), which trim latency compared to higher-order schemes for illustration PAM-4 utilized successful different systems specified arsenic InfiniBand and Ethernet. The ChromX level is programmable, allowing it to set nan number of wavelengths to lucifer nan circumstantial needs of an AI workload, whether for training aliases conclusion tasks.

Raghunathan told The Next Platform’s Timothy Prickett Morgan, “The imagination is to lucifer in-package connection bandwidth to off-package connection flight bandwidth. And we deliberation erstwhile we usage our multicolor approach, we tin lucifer that truthful that elephantine datacenters - aliases aggregate datacenters - behave arsenic 1 large GPU.”

The imaginable effect of this exertion is enormous. AI workloads devour immense amounts of energy, and pinch information halfway request projected to triple by 2035, powerfulness grids whitethorn struggle to support up. Xscape Photonics’ innovations could connection a captious solution, enabling AI systems to run much efficiently and sustainably.

More from TechRadar Pro

  • Researchers creation tech that could 'potentially switch star cells'
  • Intel debuts important optical tech that will boost AI performance
  • Humble integrative could go cardinal worldly successful title to conquer AI information halfway market

Wayne Williams is simply a freelancer penning news for TechRadar Pro. He has been penning astir computers, technology, and nan web for 30 years. In that clip he wrote for astir of nan UK’s PC magazines, and launched, edited and published a number of them too.

More
Source Technology
Technology