Home PC News Photonics startup Lightmatter details its AI optical accelerator chip

Photonics startup Lightmatter details its AI optical accelerator chip

Ahead of the Hot Chips convention this week, photonics chip startup Lightmatter revealed the primary technical particulars about its upcoming take a look at chip, which is on monitor for a fall 2021 launch. Unlike standard processors and graphics playing cards, the take a look at chip makes use of gentle to ship indicators, promising orders of magnitude larger efficiency and effectivity.

The expertise underpinning the take a look at chip — photonic built-in circuits — stems from a 2017 paper coauthored by Lightmatter CEO and MIT alumnus Nicholas Harris that described a novel technique to carry out machine studying workloads utilizing optical interference. Such chips require solely a restricted quantity of power as a result of gentle produces much less warmth than electrical energy. They additionally profit from diminished latency and are much less vulnerable to modifications in temperature, electromagnetic fields, and noise.

Lightmatter makes outstanding claims in regards to the take a look at chip, asserting its over 1 billion transistors can course of greater than 2TB of information per second. The firm additionally says — albeit with out benchmarks — that the take a look at chip outperforms Nvidia graphics playing cards, Intel and AMD processors, and even special-purpose {hardware} like Google’s tensor processing items (TPUs) on state-of-the-art AI fashions. (Lightmatter says a take a look at chip ran the ResNet-50 object classification mannequin on the open supply ImageInternet knowledge set with 99% of single-precision floating point accuracy.)

“Lightmatter’s chip and communications platform could power the entirety of Google Search on just one rack,” the corporate wrote in a press launch.

Hyperbole apart, Lightmatter says its communications platform — venture Wormhole, named after the system of move management in laptop networking known as “wormhole switching” — permits roughly 50 take a look at chip processors to change knowledge at charges exceeding 100Tbps with out optical fibers. Communicating by way of a “wafer-scale” photonic platform, clusters of chips behave as if they’re one huge system, sending knowledge far throughout the array.

Lightmatter P1 chip

Above: A Lightmatter take a look at chip on a board.

Image Credit: Lightmatter

Like many-colored beams of sunshine refracting by means of a prism, a person take a look at chip can carry out calculations utilizing completely different wavelengths of sunshine concurrently. It encodes knowledge and sends it by means of optics by modulating the brightness in wire-like elements known as waveguides. Test chips talk with different chips and the surface world very like normal electronics chips (i.e., by sending a sequence {of electrical} indicators). When they end performing calculations on gentle beams, the P1 chips leverage a photodetector akin to a photo voltaic panel to transform the sunshine indicators again into electrical indicators that may be saved or learn.

So how does that profit machine studying? When an object detection algorithm processes a picture, it divides every pixel into three channels — crimson, inexperienced, and blue — and converts the picture line by line into a group of values (a vector). Separate crimson, inexperienced, and blue vectors are handed by means of a processor, which executes an algorithm on the vectors to establish in-image objects.

Digital processors go the vectors by means of arrays known as multiply-accumulate items (MACs) to carry out the execution. A silicon processor has a small variety of MACs, whereas a GPU has a complete array of MACs, making the latter extra performant. But optical chips like Lightmatter’s take a look at chip are capable of execute algorithms by passing entries of vectors by means of a spread of digital-to-analog converters, the place they’re transformed from digital sequences into proportionally sized electrical indicators.

An optical modulator inside the take a look at chip converts the indicators into optical indicators carried by beams of lasers inside waveguides fabricated from silicon. The vectors encoded in gentle and guided by waveguides shine by means of a 2D array of optical gadgets that carry out the identical operations as a MAC. But in distinction to digital processor MACs, the place every layer has to attend for the earlier layer to complete, calculations within the take a look at chip happen whereas the beam of sunshine is “flying” (sometimes in 80 picoseconds).


Above: A Lightmatter chip.

Image Credit: Lightmatter

Lightmatter’s {hardware}, which is designed to be plugged into a normal server or workstation, isn’t resistant to the constraints of optical processing. Speedy photonic circuits require speedy reminiscence, after which there’s the matter of packaging each element — together with lasers, modulators, and optical combiners — onto a tiny chip wafer. Plus, questions stay about what sorts of nonlinear operations — fundamental constructing blocks of fashions that allow them to make predictions — could be executed within the optical area.

That could also be why firms like Intel and LightOn are pursuing hybrid approaches that mix silicon and optical circuits on the identical die, such that components of the mannequin run optically and components of it run electronically. These firms aren’t alone — startup Lightelligence has thus far demonstrated the MNIST benchmark machine studying mannequin, which makes use of laptop imaginative and prescient to acknowledge handwritten digits, on its accelerator. And LightOn, Optalysis, and Fathom Computing, all vying for a slice of the budding optical chip market, have raised tens of hundreds of thousands in enterprise capital for their very own chips. Not to be outdone, Boston-based Lightmatter has raised a complete of $33 million from GV (Google’s AI-focused enterprise arm), Spark Capital, and Matrix Partners, amongst different traders. Lightmatter says its present focus past {hardware} is guaranteeing the take a look at chip works with in style AI software program, together with Google’s TensorFlow machine studying framework.

Most Popular

Recent Comments