Earlier this month, 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 change 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 on the fiber itself before the signal fails to reach to 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 release its full system prototype in April 2022, a few weeks from now (at the time of writing) and will take center stage at the CLEO international conference in May 2022, which will focus on laser science and applications of photonics.
None of these technologies will affect end users in the short term; don't expect it to be used in laptops or smartphones anytime soon. The AI industry will use most of the benefits in data centers or in research.
"100x performance improvement"
Using the Nvidia DGX-A100 as a benchmark, Dr. Cohen told TechRadar Pro that they could achieve 500 million tasks per second using the standard benchmark comparison, MLPerf, 100 times more than Nvidia's current performance.
And they are just beginning; Performance scaling can be achieved through multicore fibers (up to 100 per fiber), using multiple wavelengths, using many processors per system (up to 000 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, including lightelligence, lightmatter, celestial.ai and Luminous, but Dr. Cohen insists that others will struggle to match CogniFiber's performance and power efficiency. Furthermore, isolated intellectual property (11 patent applications) could deter others from using a similar approach.
But that is not all; The trainable photonic autocoder neural network system is expected to consume only 500 W, which is a fraction of what the rest of the competition uses. This is a few orders of magnitude improvement on the important TOPS per watt metric. By 2026, the company expects 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 early phase, they "developed an FPGA-based timing mechanism to minimize delay and jitter, and with a relatively slow clock (0,2-1 Ghz, compared to 10-40 Ghz). , a robust sample of the values output through the stable part of the loop.
What about the line fees? Although they are not referring to line speeds, the task clock is the optical system clock or, as they call it, the optical system data injection clock, "our target alpha prototype 50-100Mhz (10x speed increase- 20x) and our beta prototype and 0.5Ghz products (100x acceleration), higher speeds will be available later, Cohen added.
Regarding latency, it is divided into two phases: comm (receiving data blocks from clients: 10G alpha, 100G beta, 400G for 2023 products) would take several msec, depending on the size and distance of the data (like any other provider of services). Compute 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 almost US$1 million, depending on the services included in the product package.