Earlier this month, a little-known Israel-based startup, CogniFiber, closed a $6 million A-round with the lofty goal of reimagining “how modern computing is done”; no more no less. A series of emails with CogniFiber co-founder Dr. Eyal Cohen sheds more light (pun intended) on what may be the most fundamental shift in computing technology in decades.

At the center of the claims is Deeplight, the proprietary technology developed by CogniFiber that is at the forefront of fiber computing, the ability of fiber optic cables to “process complex algorithms in the fiber itself before the signal does not reach the Terminal”.

In other words, the fiber cloth itself does the heavy lifting with a healthy dose of electronics. With a proof of concept already available, CogniFiber plans to launch its full system prototype in April 2022, within weeks (at the time of writing) and will take center stage at the international CLEO conference in May 2022, which will focus on laser science and photonics applications.

None of these technologies will impact end users in the short term; don’t expect it to be used in laptops or smartphones anytime soon. Most of the benefits will be used by the AI ​​industry in data centers or in research.

“100x performance improvement”

Using the Nvidia DGX-A100 as a benchmark, Dr. Cohen told TechRadar Pro that they would be able to achieve 500 million tasks per second using the standard benchmark comparison, MLPerf, 100 times more than the current champ. performance from Nvidia.

And they’re just getting started; performance scaling could be achieved through multi-core fibers (up to 100,000 per fiber), using multiple wavelengths, using many processors per system (up to 1,800 per rack ) and adding more racks.

How long will it take for the competition to catch up? There are a number of silicon-based photonics companies such as lightelligence, lightmatter, celestial.ai and Luminous, but Dr Cohen is adamant that others will struggle to match the performance and efficiency. energy from CogniFiber. Moreover, the isolated intellectual property (11 patent applications) could deter others from using a similar approach.

But that’s not all; the trainable photonic auto-encoder neural network system is expected to draw only 500W, which is a fraction of what the rest of the competition uses. This is a few orders of magnitude improvement in the all-important TOPS per watt metric. By 2026, the company hopes to reach more than 100 Exa operations per second with an efficiency of one POP/Watt.

Latency, jitter and more

Does this system suffer from jitter? Dr. Cohen says that during the seed phase, “they developed an FPGA-based timing mechanism to minimize lag and jitter, and with a relatively slow clock (0.2-1 Ghz, compared to 10- 40 Ghz), a robust sampling of the output values ​​through the stable part of the cycle is performed.

What about line rates? Although they don’t refer to line rates, the task clock is the optical system clock or, as they put it, the optical system data injection clock, “our target alpha prototype 50-100Mhz (10x-20x speedup) and our beta and 0.5Ghz (100x speedup) products, higher speeds will be available later, Cohen added.

As for latency, it is divided into two phases: comm (receiving blocks of data from clients: 10G alpha, 100G beta, 400G for 2023 products) would take several msec, depending on data size and distance ( like any other service provider). Computation phase: up to 100 ns latency (FPGA + I/O + optics).

The beta phase is expected to be reached in the first quarter of 2023 as an AI solution as a service; the first commercial products are expected to arrive in late 2023 with complete systems selling for nearly US$1 million depending on the services included in the product package.