Xscape Photonics Grabs $44 Million to Speed Up AI

In a move highlighting the growing demand for more efficient AI processing, optical component startup Xscape Photonics has closed a $44 million Series A funding round, bringing its total raised to $57 million. The company will use the funds, gleaned in part from Cisco and NVIDIA, to further develop the startup’s ChromX platform, a programmable laser module that the vendor says generates multiple wavelengths to substantially increase the performance of GPU “escape bandwidth.”

According to Xscape Photonics, escape bandwidth refers to the speed at which electrical signals can exit a GPU during processing—that is, leave the chip to go to memory or another chip. This type of connection has become a bottleneck, the vendor says, leaving GPUs waiting for data in a workload. The solution, says Xscape Photonics, is to hook each GPU to a bunch of wavelengths that can speed transmission.

Specifically, Xscape Photonics converts data coming out of the GPU from electrical to optical signals and transfers it to a pool of high-capacity memory capable of harnessing multiple wavelengths on a single optical fiber. This technique is based on the approach known as wave division multiplexing, meaning the process of splitting light into multiple “colors” or wavelengths, each capable of transmitting data over an optical link.

Speeding Up the Speed of Light

Today, most optical components used in datacenter transmission are limited to four to eight wavelengths per connection. Xscape Photonics is aiming to eventually increase that to 128 wavelengths from one device. To start with, the first ChromX product will support 4 to 16 programmable wavelengths. Spacing of wavelengths will be variable based on customer needs, the vendor says. This will not only shorten processing time, it will reduce the amount of power consumed by AI workloads—a major concern in hyperscaler datacenters.