Do or die – the domestic neuromorphic chip "Altai"
A board with 16 Altay processors. Source: motivnt.ru
Pull over
Russian microelectronics is currently mastering the production of chips using 350-nanometer topologies using domestic equipment. This applies to the Progress STP-350 photolithograph, developed in collaboration with the Belarusian company Planar. From the perspective of global microelectronics, Russia is expected to lag behind by 25-30 years. Taiwan, Japan, and South Korea currently use 3-5 nm topologies. Forecasts are very pessimistic: in the foreseeable future, the domestic industrial base will only be able to narrow the gap, at best by 10-15 years.
Now and in the near future, Russia will be primarily short of microchips for artificial intelligence systems. AI already plays a significant role in planning and conducting military operations, and its influence will soon become fundamental. For example, in strategic defense, when neural networks learn to predict and make decisions. Something needs to be done before it's too late.
The domestic neuromorphic microprocessor "Altai," presented to Mikhail Mishustin last September, was an attempt to overtake and surpass this. Let's be clear: this isn't even a pre-production prototype, but merely a technology demonstrator. And if all goes according to plan, Russia truly has a chance to create a sovereign microelectronics industry in no way inferior to market leaders, primarily focused on artificial intelligence.
First, a little theory and storiesIf you've ever heard that the data centers of large AI companies consume as much electricity as a small country, that's no exaggeration. There are even proposals to build nuclear power plants near them to provide access to affordable energy. Training a single large language model like GPT consumes hundreds of megawatt-hours.
The problem here isn't that the computation itself is complex. The problem is how modern processors are designed. All of them—from a standard laptop to a server-grade computing chip in a data center—are built according to a design proposed by mathematician John von Neumann back in 1945. The idea is simple: there's separate memory where data is stored, and a separate processor that processes it. Data constantly flows back and forth along electrical buses—and it's this endless data transfer that consumes most of the energy. Engineers call this the "von Neumann wall": the more powerful the processor becomes, the more acute this problem becomes.
Now let's look at the human brain. It consumes about 20 watts—about the same as a dim light bulb. Yet it performs tasks that no supercomputer on the planet could handle. The secret is simple: in the brain, there is no division between "memory" and "processor." Each neuron is both a computing unit and a local storage unit for information about its state. Data doesn't travel—it's processed where it's stored.
At the Microelectronics 2025 forum, the neuromorphic microprocessor was presented to the Prime Minister by Sergei Vlasov, Deputy Director of the Kurchatov Institute's Center for Advanced Microelectronics Development, and Gennady Krasnikov, President of the Russian Academy of Sciences. Source: motivnt.ru
This is precisely the idea embodied by the Altay neuromorphic chip (AltAI), a joint development between the Novosibirsk startup Motiv NT and Kaspersky Lab. Its mathematical foundation is spike neural networks (SNNs). Unlike conventional neural networks, where neurons constantly exchange numbers, Altay neurons are silent most of the time, transmitting only a short binary signal—a spike—when truly needed. The rest of the time, synaptic connections are inactive and consume no energy. This is a direct replica of how the human brain works.
The results of this approach are astounding: according to the developers' calculations, confirmed experimentally, neuromorphic solutions outperform similar systems on classic microchips in energy efficiency by more than 1000 times. This is a physically justified consequence of eliminating the main source of losses in classic architectures. The project began in 2015 in Novosibirsk's Akademgorodok, a place with a long tradition of fundamental science and a strong engineering school. The first prototype, "Altai-1," was created in 2020. In 2022, Kaspersky Lab joined the project as a strategic investor, taking the development to a new level and transforming it from an academic project into a commercial platform.
How does Altai work?
The architecture of the Altai neuromorphic processor mimics the cerebral cortex and is arranged like a chessboard of independent cores operating in parallel and asynchronously. Each core communicates with four neighbors, simultaneously computing and storing data, and consists of three blocks: a configurator (control), memory (neuron parameters and weights), and a router (spike signal transmission). In its full configuration, the chip contains 256 such cores, which together simulate 131,072 neurons and 67 million synaptic connections.
The Altai processor uses an artificial neuron model that precisely mimics the behavior of a living cell. It functions like a piggy bank: it collects incoming signals, but if new signals are not received for a long time, the old accumulations simply "evaporate." When a critical amount of charge accumulates, the neuron instantly fires, transmits the impulse further, and resets completely. Millions of these simultaneously operating "piggy banks" allow the processor to solve the most complex problems at a speed unimaginable for conventional computers.
What is the output?
With a power consumption of less than 0,5 watts, the chip is capable of:
- process video streams at speeds up to 2200 frames per second;
- perform up to 67 billion computational operations per unit of time;
- All this is in a 9 x 9 mm case, suitable for installation in any device, from a simple sensor to an on-board system drone.
Source: motivnt.ru
To put this into perspective, a typical NVIDIA Jetson AI accelerator used in robotics and unmanned systems consumes 15–60 watts. Altai solves comparable problems with 0,3–0,8 watts—a 50–200-fold difference. When thousands of such chips are deployed in a hypothetical data center, this difference translates into megawatt-hours of savings per year. When it comes to the onboard system of a small drone with a 50 Wh battery - the difference means a flight time of a few minutes or a few hours.
Neuromorphic chips aren't a Russian innovation. Experiments with this new architecture are also underway abroad. A comparison with global competitors shows that Altai is in a strong position. IBM TrueNorth, the most energy-efficient commercial neuromorphic chip available today, processes 1738 frames per second at 0,2 watts. Intel Loihi runs 296 frames per second at 0,11 watts. Altai runs 1000–2200 frames per second at 0,3 watts. In absolute performance, the Russian chip is on par with TrueNorth and ahead of Loihi; in terms of energy efficiency, IBM maintains its advantage. Context is important: the neuromorphic market hasn't yet matured, and none of the competitors have become the definitive standard—and this in itself means the field is still open to everyone. The software for the neuromorphic chip is being developed by Kaspersky programmers. The product is called the Kaspersky Neuromorphic Platform.
Altai for war
When discussing the topic of Russia's sovereign microchip, we can't ignore its military potential. And it's practically boundless.
The Altai's main advantage is its extremely low power consumption. For small UAVs, this means a fundamentally new balance of capabilities: the aircraft no longer has to choose between range and intelligent onboard electronics. Computer vision, object recognition, and autonomous navigation can be used onboard with virtually no loss of flight time or payload.
Altai is especially important when paired with a video camera, which captures only changes in the scene with microsecond resolution, rather than a full frame. This allows information to be processed as it emerges, without the delays of storing and analyzing the entire image. As a result, reaction time is reduced to sub-millisecond levels—faster than traditional systems, and even faster than a human operator. In short, Altai is ideal for robots—it enables machines to make decisions and navigate faster.
This opens up equally important opportunities for reconnaissance sensor networks. The chip makes it possible to create sensors that operate for weeks on a small solar panel, continuously analyzing acoustic, seismic, and thermal conditions, and transmitting ready-made conclusions rather than raw data: for example, "column movement" or "sound" tank engine." This dramatically reduces radio traffic and makes such a system significantly less visible to enemy radio reconnaissance.
The Altai also holds particular promise for guided munitions and autonomous weapons. In conditions of GPS jamming and a disrupted guidance channel, a processor is needed that simultaneously meets weight, power consumption, and resistance to severe mechanical stress. A neuromorphic chip meets these constraints far better than most traditional solutions. Furthermore, neuromorphic networks remain a less studied target for adversaries in terms of countermeasures, optical decoys, and camouflage patterns.
In the field of electronic warfare, the advantages are also obvious. Altai can process radio signals as a stream of events, quickly classify modulations and communication protocols, and adapt to changes in enemy operations faster than traditional techniques and methods. In a wearable format, this opens the way to individual means. EW- protection that can operate for days on one battery.
The ALTAI neuromorphic microprocessor on display at the Kurchatov Institute. The neuromorphic accelerator for the first-generation processor is located in the center. Source: motivnt.ru
Altai's use in cybersecurity also shows promise. Instead of searching for matches against a database of known attacks, the system can analyze device behavior and identify anomalies in real time, isolating potentially compromised nodes without accessing the central security system. This is the main advantage of neuromorphic microprocessors—they don't require large amounts of memory or cloud storage.
But it wasn't without its challenges. First, the neuromorphic chip is still in the development stage. It's unknown how long it will take to bring it to production, or even if it will be produced at all. Second, the Altai is designed for a 28nm process. Russia doesn't have its own lithography equipment for such tasks—neither domestic nor imported. Manufacturing will have to be done in China. And that's a significant dependency. The Altai could, of course, be redesigned for the 65–90nm processes available at Mikron, but then the chip's unique energy efficiency would be lost. Third, training models for neuromorphic AI chips will have to be developed anew. Standard models like those used in ChatGPT or Grok won't work.
Without a doubt, the development of the Altai microchip is a technological breakthrough. Not a repurposed Chinese development, but a truly world-class, independent product. A rare, and therefore valuable, phenomenon these days.
Information