Greetings, dear readers!
Concluding our series of articles on cutting-edge neuromorphic processor architectures, we decided to focus on domestic developments that often remain outside the public spotlight. In this article, we will discuss the “Altai” neuromorphic processor.
In the world of neuromorphic computing, we have already explored several remarkable developments: the architecture of IBM’s TrueNorth — one of the first commercial neuromorphic processors; a deep dive into SpiNNaker, a massively parallel platform capable of emulating millions of biological neurons in real time; and Tianjic — a hybrid Chinese chip that combines support for both conventional and spiking neural networks. Today, we will turn our attention to a Russian solution for energy-efficient computing — the “Altai” neuromorphic processor, designed for autonomous real-time processing of video, audio, and sensor data.
Altai (ALTAI) is a Russian digital spiking neuromorphic processor developed by the company “Motiv NT” for high-performance yet energy-efficient processing of large volumes of streaming data in real time. The project aims to ensure technological independence and autonomy in the field of artificial intelligence, where traditional CPU-, GPU-, and FPGA-based solutions often fail to meet stringent requirements for power consumption, size, and scalability.
At the core of Altai lies a principle of biologically inspired computing: the processor models the behavior of spiking neural networks (SNN), enabling a high degree of parallelism and significantly reducing power consumption. This approach is particularly important for robotics, IoT devices, unmanned systems, and other domains requiring a compact, fast, and low-power hardware platform. The emergence of such technology in Russia reflects not only the high expertise of domestic developers but also a strong potential for the ongoing growth of neuromorphic systems.
History of the Altai Neuromorphic Processor
In the mid-2010s, artificial intelligence began developing rapidly, but its hardware implementation was still tied to traditional architectures. Despite substantial computing capabilities, these solutions were hampered by drawbacks such as high power consumption, limited scalability, and latency. Around this time, the idea arose to create a fundamentally new computing device inspired by biological neural networks. Thus, in 2015, within the framework of SNIL VT (Special Research Laboratory of Computing Technologies) on the basis of NSTU, the basic principles of the Altai processor architecture were formulated.
Practical work on the project began in 2017 when “Motiv Neuromorphic Technologies” (now “Motiv NT”) received grant support from the Foundation for Assistance to Small Innovative Enterprises (FASIE) and the Skolkovo Innovation Center. In 2018, the team received a Skolkovo mini-grant and focused on developing the first hardware prototypes. Collaboration with Kaspersky Lab, which became a strategic partner of the project, played a crucial role.
By 2020, the first prototype, Altai-1, was created, featuring 16 neuromorphic cores, each modeling 512 neurons with 512 input lines. Overall computing capacity allowed for modeling up to 8,192 neurons and 4 million synapses. The processor showed high energy efficiency with low power consumption (around 300 mW), surpassing traditional solutions. That same year, a research article describing the chip architecture and test results on the MNIST (98% accuracy) and CIFAR-10 (83.2%) datasets was published. Notably, energy consumption proved to be orders of magnitude lower than that of conventional graphics processors performing analogous tasks.
In 2022, Kaspersky Lab became an investor in “Motiv NT,” allowing the team to expand research and ready the platform for commercial use. That year also saw a research collaboration agreement with the National Research Center “Kurchatov Institute,” and the company joined the Microelectronics Accelerator program. By 2023, an updated Altai-1 (v.2) prototype was presented, becoming part of demonstration solutions in industrial metering, cybersecurity, and computer vision.
In parallel, development began on the Kaspersky Neuromorphic Platform (KNP) — a software platform for designing, training, and testing neuromorphic algorithms. KNP became a fundamental software solution ensuring that neuromorphic computing would be accessible to the research community and developers alike.
According to information presented at the Kaspersky Neuromorphic Conference 2024, a pilot batch of Altai-3 processors with a small number of cores is expected in 2025, followed by the engineering launch of the first commercial version of Altai-3 in 2026. This version will target energy-efficient embedded systems, “AI-on-chip” devices, and industrial solutions.
Altai Processor Architecture
The Altai processor is a neuromorphic Very Large Scale Integration (VLSI) circuit aimed at energy-efficient real-time execution of Spiking Neural Networks (SNN). The architecture is designed to handle the large volumes of noisy data typical in machine learning, the Internet of Things, robotics, and autonomous intelligent devices.
Key architectural goals include:
- Energy efficiency: Minimizing power consumption by localizing computation and data storage.
- Scalability: Enabling expansion of the number of cores without changing physical interfaces or packaging, ensuring consistency during integration into larger systems.
- Real-time performance: Processing neural signals in real time with a synchronization frequency of up to 2,000 Hz.
Altai is implemented using modern monolithic silicon CMOS technology at a 28 nm process node, ensuring stable operation and high performance. The 9×9 mm chip package allows integration into various classes of devices, from IoT systems to data centers. All computing logic is digitally implemented, offering reliability, precision, and immunity to analog noise.
At the core of the architecture lies a digital neuron model that reproduces key characteristics of biological neurons: the activation threshold, signal integration, and other parameters of the digital model. Neuron parameters (e.g., threshold, time delay, refractory period) are set programmatically, enabling adaptation to different tasks without hardware changes.
Altai follows a GALS (Globally Asynchronous, Locally Synchronous) approach. It consists of multiple independent neural cores arranged into a unified scalable network. Within each core, synchronous logic ensures stability and predictability of calculations, while asynchronous inter-core communication eliminates the need for a global clock signal and reduces power consumption. This design mimics the asynchronous interactions of biological neurons in the brain.
To efficiently exchange data among the neural cores, Altai uses special asynchronous communication blocks that enable signal transmission without waiting for a global clock pulse — crucial for real-time event-based processing.
The architecture also supports scalability via modular core clustering and interfaces for inter-chip integration. Processors can be interconnected into practically unlimited matrices to increase computing power, facilitating systems from small embedded devices to large computing clusters.
Another key feature is fault tolerance: in tasks relying on imprecise algorithms, the failure of a single core does not substantially distort the results. The system remains operational even if multiple cores malfunction.
Technical Specifications
The Altai-1 neuromorphic processor has the following technical specifications:
- Performance: Up to 100 billion operations per second
- Video processing: Up to 1,000 frames per second (can be boosted to 2,200 fps)
- Power consumption: 70 mW
- Number of cores: 16 neuromorphic cores
- Number of neurons: Up to 8,192 modeled neurons
- Number of synapses: Up to 4 million
- Maximum neural network synchronization frequency: 2,000 Hz
- Process technology: 28 nm
- Package dimensions: 9×9 mm
Kaspersky Neuromorphic Platform: Software for the Altai Chip
The Kaspersky Neuromorphic Platform (KNP) provides an extensive suite of software solutions for designing, training, and executing neural networks. Developed with the hardware features of the Altai neuromorphic processor in mind, these tools allow engineers to create energy-efficient and high-performance applications.
The KNP SDK offers:
- Execution of SNNs (Spiking Neural Networks) on a CPU or Altai-1 processor
- Conversion of ANN (traditional Artificial Neural Networks) to SNN with support for ternary layers
- C++ and Python APIs for interacting with neurons, synapses, and their populations
- Multiple backends for computations, including multithreading and Altai hardware acceleration
The software package also includes:
- A compiler for systems programming:optimizes code for neuromorphic architecture.
- A neural network preprocessor: prepares neural models for spiking architecture.
- A neural network compiler: converts models into a format optimized for the Altai’s execution environment.
- A collection of examples: demonstrates core functionalities using examples of deep neural networks.
The KNP platform also supports integration with machine learning libraries, making Altai processors a versatile solution for tasks demanding high energy efficiency. According to Kaspersky Lab’s official website, the source code of KNP is now available under the Apache 2.0 license, allowing developers not only to use the ready-made solutions but also actively contribute to its further development, tailoring it to specialized tasks or experimenting with new algorithms.
Pilot Projects
The Altai neuromorphic processor is used in various pilot projects that showcase its real-time data processing capabilities. Below are some key demonstration projects:
Physical Process Characterization
Combining a neuromorphic camera and the Altai processor enables energy-efficient characterization of dynamic physical processes. In one experiment, a variable-speed fan was used as the model. A spiking network running on Altai solved a regression task: predicting the fan’s speed and rotation direction from DVS-camera data. Compared to a GPU solution (GTX 1650 Ti), Altai’s power consumption was one to two orders of magnitude lower (0.3 W versus 13 W), while maintaining high accuracy and data processing speed.
Acoustic Data Processing
In collaboration with the Kurchatov Institute, a demonstrator was developed for classifying urban environment sounds. A classifier running on the Altai neuromorphic chip divided audio signals into 10 classes. The methodology involved converting an ANN to an SNN while preserving accuracy, as well as frequency coding of amplitude. Thirteen cores were involved, covering 5,100 neurons. Power consumption totaled less than 70 mW, whereas a similar process on an Nvidia GPU required 27.59 W — demonstrating a substantial advantage in energy efficiency.
Fast Process Handling
A demonstration by Kaspersky Lab showcased the ability to handle high-speed signals. The task was to determine, via regression, the blinking frequency of an LED (light-emitting diode) from DVS-camera data, with the LED frequency varying sinusoidally from 0.25 to 1.25 kHz. The Altai-based system could evaluate the blinking frequency in an average of 1.4 ms, processing data streams of up to 2,500 frames per second. This illustrates the potential of neuromorphic computing for high-speed data processing.
Attack Detector
Another pilot project demonstrated how spiking neural detectors could identify attacks on a system. In this experiment, the spiking neural network detector running on Altai displayed enhanced resilience to adversarial attacks on facial recognition systems (Face ID), a crucial feature for information security applications.
These pilot projects highlight the practical applicability of the Altai neuromorphic processor to a range of real-time data processing tasks where energy efficiency and speed are critical.
Key Application Areas
Use of the Altai neuromorphic processor spans a wide range of tasks in various sectors:
IoT Devices and Robotics
Altai can locally process large data streams from sensors and cameras in real time. Its low power consumption and compactness make it ideal for robotics, wearables, smart sensors, and autonomous gadgets in the Internet of Things. Local, on-device intelligent data processing in a small form factor opens new possibilities for building autonomous robots and IoT systems that can operate for extended periods on a single power source.
In industrial IoT, Altai can be used for predictive equipment maintenance, analyzing data from vibration, temperature, and other sensors on-site without transferring large amounts of data to the cloud — particularly relevant for remote installations with limited network access.
Data Centers
The neuromorphic architecture of the processor delivers energy-efficient big data processing. Significant reductions in electricity costs are achieved thanks to Altai’s low power consumption. In server solutions, this means lower spending on power and cooling as well as increased throughput for machine learning and analytics tasks.
According to the developers, using neuromorphic processors in data centers can reduce power consumption for data processing by up to 90% compared to conventional GPU solutions. This is particularly important in light of the ever-increasing volume of data handled by modern data centers.
New Devices
Altai can serve as the foundation for passive 3D sensors and computer vision systems in robotics and unmanned vehicles. For instance, it supports the operation of passive 3D sensors for machine vision, enabling on-device image recognition and decision-making. This is essential for autonomous transportation, industrial process control, and other systems where the combination of high processing speed and low power consumption is critical.
With the ability to process several thousand frames per second, Altai is indispensable in high-speed video analytics scenarios — for example, on industrial assembly lines or in scientific research focusing on rapidly occurring processes.
Cybersecurity
Neuromorphic networks powered by Altai can detect anomalies in computer and industrial networks in real time. Spiking neural networks exhibit strong resilience to various types of attacks and can be used to build robust monitoring and protection systems. Anomaly detection happens faster and with less energy compared to traditional intrusion detection systems.
Kaspersky Lab is actively researching how to employ Altai for protection against new attack types, including advanced persistent threats (APT) and zero-day exploits. By enabling online learning, neuromorphic systems can adapt to new threats without the constant need to update signature databases.
These application areas demonstrate the broad potential of neuromorphic processors in tasks where energy efficiency, compactness, and high performance are essential.
Conclusion
The creation of the Altai neuromorphic processor marks a significant milestone in the development of domestic computing technologies. The chip demonstrates the capabilities of Russian solutions in energy-efficient artificial intelligence. Its architecture, inspired by biological neural networks, opens up new avenues for building autonomous systems, embedded AI, and event-driven data processing. By integrating Altai with the Kaspersky Neuromorphic Platform, researchers and engineers gain direct access to neuromorphic solutions, simplifying the development and testing of AI applications.
High performance, complemented by low power consumption and compact dimensions, makes Altai a universal solution for IoT and embedded systems, as well as for data centers. The processor architecture is focused on event stream processing and real-time operation, which is especially important in conditions of limited computing and energy resources. Thus, Altai demonstrates not only superiority in energy-efficient computing, but also broad opportunities for implementation in a wide range of industries, forming the basis for further development of domestic solutions in the global market of artificial intelligence.
With this article, we conclude our block dedicated to modern neuromorphic processor architectures and advanced approaches in the field of spiking neural networks. The world of neuromorphic computing is rich and varied. We have not covered many other architectures, such as Intel’s Loihi, BrainScaleS, Akida, Neurogrid, DYNAP-SE, SenseEdge, and many others. Each one offers intriguing technical solutions, and we will certainly revisit them in the future. For now, our priority is highlighting domestic achievements and emerging hardware alternatives. The next block in our article series will focus on chip design technologies and microelectronic solutions that underpin the future of computing platforms. In our following article, we will share our team’s achievements and project results. Stay tuned — there’s much more to come!
We value your feedback: which topics in neuromorphic computing are you most interested in? What would you like to discuss with us? Let us know in the comments — we’re always open to dialogue.
Thank you for being with us!
Best regards, the MemriLab Team!
References:
- Neuromorphic Processor “[Altai” for Energy-Efficient Computing]"(https://www.researchgate.net/publication/342929519_NEJROMORFNYJ_PROCESSOR_ALTAJ_DLA_ENERGOEFFEKTIVNYH_VYCISLENIJ), 2020. GRISHANOV, ZVEREV, IPATOV, KANGLER, KATOMIN, KODENKO, KOSTRITSYN, MAKAROV, MAMYCHEV, PAVLOV, PANCHENKO, POLSTYANKIN.
- Development of a Custom Neuromorphic Very Large Scale Integrated Circuit. Poster. GRISHANOV, ZVEREV, IPATOV, MAKAROV, MAMYCHEV, POLSTYANKIN.
- Motiv
- Kaspersky Neuromorphic Platform
- Kaspersky Neuromorphic AI 2023
- Kaspersky Neuromorphic AI Conference 2024
- Presentations: Neuromorphic Chips in Robotics 4.0*