Using light to connect an AI processor’s cores

Lightelligence has unveiled its optical network-on-chip designed to scale multiprocessor designs.

The start-up’s first product showcasing the technology is the Hummingbird, a system-in-package that combines Lightelligence’s 64-core artificial intelligence (AI) processor and a silicon photonics chip linking the processor’s cores.

A key issue impeding the scaling of computing resources is the ‘memory wall’ which refers to the growing gap between processor and memory speeds, causing processors to be idle as they wait for data to crunch.

“The memory wall is a genuine problem,” says Maurice Steinman, vice president of engineering at Lightelligence. “Even with the generational scaling of computing, the input-output (I/O) and access to off-chip communications are struggling to keep up.”

Lightelligence has developed Hummingbird to show how its optical networking approach could be used to scale multi-chip designs.


Lightelligence, an MIT spin-out, is a fabless chip company founded in 2017. The start-up has 200 engineers and has raised over $200 million in funding. Now, after five years, the company is beginning to generate revenues.

The company started by using nanophotonics to tackle such computations as matrix multiplications and linear algebra.

After two years, the focus broadened to include communications. There is no point in developing a low-latency analogue compute engine only to then encounter I/O and scaling issues, says Steinman.

Optical network-on-chip

The start-up is pursuing two communication tracks. The first is developing board-to-board or rack-to-rack communications based on optical transceivers and fibre. The second is communications at a smaller, system-in-package or chip-to-wafer scale, with Hummingbird and the optical network-on-chip being the first example.

“How do we address the challenges of purely electronic solutions today?” says Steinman. “How can we use photonics and an optical waveguide-based solution to get back some of the limitations there?”

The issue is that while chips can now have transistor counts in the tens of billions, a dimensions of the die are limited to some 800mm2, dictated by the reticule size. A multi-chip module or a chiplet approach is needed if additional computation is required.

“Ideally, you would want the performance to scale with the sum of silicon area; if I’ve got N chips, I want pure linear scaling,” says Steinman.

The issue with multiple chips is that they need interfaces which introduce communication and power consumption issues. And with an array of chips, a scheduler must oversee workload assignment to the processors as they become available.

“You’ve got all these interfaces, you’ve got queuing delay, you’ve got contention, you have multiple messages trying to contend for the same path,” says Steinman. “It is hard to see all that in multiple dimensions.”

Lightelligence wants to use optical networking to enable routing topologies linking compute resources that are impractical if attempted electronically.

“That’s the breakthrough we’re trying to bring to the world,” says Steinman. “What we are calling optical network-on-chip.”

The first implementation uses a network to link cores of a single-chip parallel processor, while the goal is to extend the networking beyond the chip scale, he says.

Lightelligence believes the technology will be attractive to silicon vendors facing similar scale-out issues. And by having a functioning device, the start-up has credibility when approaching potential customers.

“We want to work with them to design a purpose-built semi-custom solution because I’m sure every scale-out solution has different topology needs,” says Steinman.


The Hummingbird device uses programmable cores that implement scalar, vector, and matrix operations, including 2D convolution. Such computations are accelerated using the optical network-on-chip with convolution used to implement a convolutional neural network for AI.

The chip includes a central instruction unit to implement a single-instruction, multiple-data (SIMD) architecture; each core performing the same operation on part of the data set.

To aid the computation, each core has an optical broadcast transmitter. Every core can send and also receive data from every other core using the silicon photonics chip’s splitters and optical waveguides.

The ratio used by the optical-network-on-chip is 64 optical transmitters and 512 optical receivers rather than 64×64 optical receivers. This simplifies the optical design’s complexity, with electronics being used for the final stage to get data to particular cores in eight-core clusters.

“It is an all-to-all broadcast, an unusual topology,” says Steinman. “But in doing that, we have put a lot of transmitters and receivers on our companion photonic die that goes with the electronic [AI processor] die in Hummingbird.”

Lightelligence says the silicon photonics chip could implement other topologies, such as a 2D torus, for example.


The motivation for the Lightelligence design is to achieve linear scaling, beyond what Hummingbird is showing for the single chip design. That said, the design already shows that many optical transmitters and receivers can be integrated into a dense space.

Lightelligence’s approach is also pragmatic. It has taken several years to develop Hummingbird and the start-up didn’t want to wait before developing a wafer-scale solution.

“There’s a little bit of constraining the problem to a single die, which is not the optimal proof point for it,” says Steinman. “But now we have a tangible working thing that gives us credibility for those higher-scale conversations.”

Steinman says the limit using the optical-network-on-chip technology is a wafer-sized design. Scaling beyond a wafer will require conventional optical interface technology (oNET, see diagram below) which Lightelligence is also developing.


The start-up has developed a PCI Express (PCIe) plug-in card that hosts the Hummingbird system-in-package. “It is a programmable machine; it does have an instruction set, compiler, and toolchain,” says Steinman.

Samples are with early adopter customers with AI inference tasks and Lightelligence is awaiting their feedback. “That next level of feedback is really important for our commercial objectives,” says Steinman.

Lightelligence also wants to partner with chip companies to start working on what Steinman calls ‘semi-custom’ engagements. “What are their problems that need to be solved, and how can we help?” he says.