Company
Founded in 2017 by Mihai Raneti, CyberSwarm started with a question about materials rather than a question about architectures: if a component can store a value and multiply by it at the same time, what kind of machine can you build?
The premise
Every conventional processor keeps them apart and pays a toll on every operation to move data between them. Biology does not work that way. A synapse is memory and computation in one place, and it costs almost nothing to maintain.
A memristor is the closest electronic analogue we have. CyberSwarm started in Ploiești, in a lab rather than a slide deck, working out how to write reliable resistive states into an IGZO thin film and how to read them back without disturbing them. Everything since, the array, the compiler and the runtime, sits on top of that work.
The device physicists who filed the founding patents still work here. They are the inventors named on the core grants. The IP and the lab are the same effort.
We do not simulate a brain on top of a digital machine. The computation happens in the material. Ohm’s law and Kirchhoff’s law do the multiply-accumulate, for free.
The engine was designed for the constraint that matters in the physical world: a fixed power budget, no network, and a decision that has to land in milliseconds.
Device, array, compiler and runtime are developed together. That is why the portfolio starts at the material and ends at a deployable model.
Inference never leaves the device. No cloud dependency, no telemetry, no third party in the loop. That is a property of the architecture, not a policy setting.
Timeline
Drag the rail, or use the arrows. Nothing moves on its own.
The team
The multidisciplinary team spans materials science, analogue circuit design, and low-level software, bridging device physics directly to application deployment.
Founder and CEO
Founded CyberSwarm in 2017 on the thesis that memory and computation were never supposed to be separate, and has led it for the nine years since. Trained in mechanical engineering and psychology, with two companies behind him.
Chief Scientist
PhD in materials physics, University of Stuttgart. Five years at Honeywell, then a research career with more than 80 granted US patents. Nine years on this architecture, responsible for the device physics and the fabrication approach.
Specialists in memristive materials, thin-film deposition, and nanoscale characterisation who fabricate, model, and evaluate physical synapses directly in the laboratory.
Circuit engineers developing low-power interface electronics and interconnect routing to connect analogue memristive arrays with digital host systems.
Software engineers building the deployment stack: mapping trained network weights onto physical crossbars, and the export path a model takes to reach them.
Advisors
Former Vice President for Edge and Network at Intel Corporation
Former Vice President of the Network and Edge Group at Intel Corporation
Inside the lab
Grow the film, ask what it became, then measure it one device at a time.
01
Everything starts inside this. The chamber is pumped down before anything is deposited, because at ordinary pressure the layer would grow with the room in it. What comes out is a film a few tens of nanometres thick whose electrical behaviour is decided by how it was grown, not by anything done to it afterwards.
02
The regulators on the wall set what enters the chamber and how fast. Changing a device means changing these numbers and then measuring what came out, which is why fabrication and characterisation are one team here instead of two that send each other samples.
03
The same chamber during a run, seen through one of its viewports. The colour is heat, and the reason the vessel carries burn warnings on three sides. Temperature is one of the few settings that changes the material itself rather than the pattern drawn on it.
04
A diffractometer answers one question: how the atoms in the new layer are arranged. The output is not a picture of the film but a fingerprint of its structure, and it is the first honest check on whether a run did what the recipe said it would.
05
The array built on two different carriers. The transparent one is why the founding patents describe the synapse as transparent: the stack can be grown on glass, which matters for anything that has to sit on a surface without taking the surface over.
06
This is the object the rest of the company is built around. Nothing about it looks like a processor, because it is not one. There is no clock, no instruction and nothing that fetches. The grid is memory that also does the arithmetic.
07
The pads around the edge are where probes land. The regular field in the middle is the array. Every square in that field has to be written, read, and written again without the value drifting, and that requirement is what most of the last nine years went into.
08
Characterisation happens here. The stage can be taken down to low temperature, which separates behaviour belonging to the material from behaviour belonging to the room it was measured in. The table underneath is isolated, because at this scale a footstep is noise.
09
The enclosure is dark deliberately. These films respond to light as well as to voltage, so a measurement taken with the door open is not the same measurement. Loading a die and landing the probes is done by hand, one sample at a time.
10
Each needle is a wire to one line of the array. With the probes landed, a voltage can be put across a single device and the current read back, which is the whole of what a memristor measurement is: apply, read, repeat, and watch whether the number stays where it was put.
11
What the operator sees while landing a probe. The tip has to press hard enough to make contact and gently enough to leave the pad usable afterwards. The marks left by earlier measurements are visible on every pad that has already been used.
12
Most days look like this rather than like a render. One person drives the positioners while the other watches the read-out, and a single sweep can run long enough that it is worth having someone else confirm what is on the screen.
13
All of it. Nine years of this company came out of a space this size, which is the part that is easy to miss when the number in the headline is the funding round.
Fabrication and characterisation are shared with partner research institutes.
Locations
The company is incorporated in Delaware and registered in San Mateo. The engineering, from device fabrication and characterisation to the compiler, is in Bucharest, close enough together that a measurement in the lab reaches the toolchain team the same week.
California, USA
Headquarters & commercial
Romania
Engineering: devices, architecture & software
“Tim Draper was the 1,214th investor I spoke to, after roughly 16,000 emails sent over a period of eight years.”
For investors
Backed by Draper Associates since 2017. The market thesis, the commercial architecture, the traction behind it and an honest account of the roadmap ahead.
Read the investment caseIf you want to work on the layer under the layer, write to us at office@cyber-swarm.net.