BrainChip Akida
Last reviewed
May 21, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 4,750 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 21, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 4,750 words
Add missing citations, update stale details, or suggest a clearer explanation.
Akida is a family of fully digital, event-based neuromorphic processor intellectual property (IP) cores and reference chips developed by BrainChip Holdings Ltd, an Australia-based fabless semiconductor company (ASX:BRN, OTCQX:BRCHF).[^1][^2] First announced in September 2018 and commercially released in late 2021 as the AKD1000 reference AI accelerator, Akida implements spiking neural network (SNN) inference and on-chip incremental learning for ultra-low-power edge AI workloads, targeting smart sensors, automotive cabin systems, industrial IoT, robotics and defense applications.[^3][^4][^5] BrainChip positions itself as "the world's first commercial producer of neuromorphic IP," and its second-generation Akida 2.0 IP, announced in March 2023, added Vision Transformer acceleration, 8-bit weight support and Temporal Event-Based Neural Nets (TENNs) for streaming sensor data.[^6][^7] Publicly disclosed deployments and license agreements span Mercedes-Benz (Vision EQXX concept car), Renesas Electronics, MegaChips, Edge Impulse, NASA's Ames Research Center and the U.S. Air Force Research Laboratory.[^8][^9][^10][^11][^12][^13]
| Item | Detail |
|---|---|
| Developer | BrainChip Holdings Ltd (ASX:BRN) |
| Headquarters | Sydney, New South Wales, Australia; engineering offices in California, Toulouse and Hyderabad |
| Founder | Peter van der Made (2004) |
| First architecture announcement | 10 September 2018 (Akida NSoC architecture) |
| First reference chip | AKD1000 (TSMC 28 nm CMOS) |
| AKD1000 production testing complete | 8 November 2021 |
| Full AKD1000 commercialization | 17 January 2022 |
| Second-generation IP announcement | 6 March 2023 (Akida 2.0) |
| Second reference chip | AKD1500 (GlobalFoundries 22 nm FD-SOI) |
| AKD1500 tape-out announced | 29 January 2023 |
| Neural fabric (AKD1000) | 80 Neural Processing Units; 1.2 million neurons; 10 billion synapses |
| Weight precision | 1, 2, 4 bits (Akida 1.0); 8 bits added in Akida 2.0 |
| Software stack | MetaTF (akida, akida-models, quantizeml, cnn2snn) |
| On-chip learning | Last-layer continual learning (fully connected) |
| Stock listing | Reinstated on Australian Securities Exchange via reverse merger of Aziana, September 2015 |
BrainChip's neuromorphic technology traces back to research begun in 2004 by Dutch-born Australian engineer Peter van der Made, who had previously founded the graphics company PolyGraphic Systems and served as CTO of IBM Internet Security Systems from 2002 to 2004.[^14] According to BrainChip's company history, van der Made "designed the first generations of digital neuromorphic devices on which the Akida chip is based" between 2004 and 2008, filing the original patents on the architecture during that period.[^1][^14] Van der Made later authored a popular-science book on brain-inspired computing, Higher Intelligence, summarizing the underlying motivation for the company's spike-based design.[^14]
(The user prompt referred to the predecessor name "Aiconic Systems"; publicly available sources do not corroborate that name. Wikipedia and Australian small-cap reporting instead trace the modern listed entity to BrainChip Inc, founded by van der Made in 2004, and to the Australian Securities Exchange shell company Aziana Limited, which acquired BrainChip in March 2015 and was subsequently relisted as BrainChip Holdings in September 2015.[^2][^14] To avoid fabrication, this article uses only the names confirmed in cited sources.)
In March 2015, the dormant ASX-listed mining company Aziana Limited acquired BrainChip Inc; following a reverse merger, BrainChip Holdings Ltd was reinstated on the Australian Securities Exchange in September 2015 under the ticker BRN.[^2][^14] In 2016, former Exar Corporation chief executive Louis DiNardo succeeded van der Made as CEO, and van der Made took the title of chief technology officer.[^14] BrainChip's headquarters is registered at Level 12, 225 George St, Sydney, with a Research Institute office in Perth and engineering staff in California (where U.S. operations are managed from Laguna Hills), Toulouse and Hyderabad.[^1][^15]
On 10 September 2018, BrainChip publicly disclosed the Akida Neuromorphic System-on-Chip (NSoC) architecture, describing a fully digital, event-based spiking neural network device intended for edge AI.[^16] At the same time, the company opened access to an Akida Development Environment for early customers, with first silicon sampling targeted for the third quarter of 2019.[^16] The initial press materials described a programmable mesh of neural processing nodes that could scale by chip-to-chip linking, with a single Akida NSoC reporting an effective capacity of approximately 1.2 million neurons and 10 billion synapses; up to 1,024 chips could in principle be tiled to extend the mesh.[^16][^17]
The first production reference chip, AKD1000, was fabricated by TSMC on a 28 nm digital CMOS process.[^17][^18] On 8 November 2021, BrainChip announced that "functionality and performance testing of the AKD1000 production chips has been completed, which showed better performance than the original engineering samples," marking the transition from R&D to full production of the chip and its IP.[^19] CEO Peter van der Made characterised the milestone as the company's move into commercialization. The same year, BrainChip began shipping Akida AI Processor Development Kits, and in October 2021 it opened orders for x86 (Shuttle PC) and Arm-based (Raspberry Pi) development kits.[^14][^20]
On 17 January 2022, BrainChip announced full commercial availability of the AKD1000, opening orders for the first Mini PCIe boards based on the chip and publishing full PCIe design files and bills of materials for system integrators.[^21] Mini PCIe boards initially shipped at a list price of USD 499, with later M.2 modules added in early 2025 at USD 249.[^21][^22] BrainChip CEO Sean Hehir, who succeeded DiNardo as chief executive, described the launch as the culmination of "a decade-long development effort" in autonomous incremental learning at ultra-low power.[^21]
A rapid sequence of partner announcements followed commercialization. In November 2021, the Japanese fabless ASIC house MegaChips Corporation signed an Akida IP license to integrate the technology into next-generation edge AI ASICs across automotive, IoT, gaming and industrial markets.[^23] In late 2022, Renesas Electronics America signed BrainChip's first IP license with a major MCU vendor, committing to incorporate Akida 1.0 IP into Renesas system-on-chip products under a single-use, royalty-bearing, worldwide design license.[^12] Around the same time, NASA's Ames Research Center purchased an Akida Early Access Evaluation Kit to assess the processor's suitability for spaceflight environments, citing the fact that the Akida chip can run feed-forward networks without an external host CPU, memory, or deep learning accelerator.[^12]
In November 2022, edge AI training platform Edge Impulse announced formal support for Akida, making BrainChip the platform's first strategic IP partner and adding learning blocks that target the AKD1000 directly.[^11][^24] In 2023, BrainChip's Akida processor family was integrated with Arm's Cortex-M85 processor for use in low-power MCU subsystems.[^14]
On 29 January 2023, BrainChip announced that it had completed the design of its second reference chip, AKD1500, taped out on GlobalFoundries' 22 nm fully-depleted silicon-on-insulator (FD-SOI) process.[^25][^26] The AKD1500 reference design integrates Akida event-based IP with quad/octal SPI for direct MCU attachment and a PCIe interface for x86, Arm and RISC-V host platforms; product literature reports 800 giga-operations per second (GOPS) at under 300 mW.[^25][^26] On 6 March 2023, the company unveiled the second-generation Akida 2.0 IP platform, adding 8-bit weights and activations, Vision Transformer acceleration, hardware skip connections for residual networks, multi-pass sequential processing of large models with configurable scratchpads, and a new family of Temporal Event-Based Neural Nets (TENNs) for spatiotemporal streaming data such as video, radar and biomedical signals.[^7][^27]
On 9 December 2024, BrainChip announced a US$1.8 million development award from the U.S. Air Force Research Laboratory under topic AF242-D015, "Mapping Complex Sensor Signal Processing Algorithms onto Neuromorphic Chips," focused on micro-Doppler radar processing for power- and thermally constrained platforms such as drones, drone defense systems and missiles.[^13] In April 2025, BrainChip and RTX's Raytheon division announced a teaming agreement to deliver the AFRL neuromorphic radar prototype.[^28] BrainChip has also publicly cited evaluations by Information Systems Laboratories, the Rochester Institute of Technology, and the European Space Agency-funded ship-detection experiments described in arXiv preprints by Lemaire, Tossou and collaborators.[^29][^30]
Akida is a fully digital implementation of an event-based (spiking) neural processing fabric.[^4][^17] Unlike clock-synchronous tensor accelerators that propagate dense activations through every layer on every cycle, Akida only performs computation when its inputs change state, that is, when a "spike" is produced or received. BrainChip and independent analysts characterise this as event-domain processing: "Computations are only performed when new sensory input is received, dramatically reducing the number of operations."[^4] Synapse weights remain stored on-chip in local SRAM tightly coupled with each processing element, eliminating most DRAM traffic during inference.
The AKD1000 reference chip is built around 80 Akida Neural Processing Units (NPUs) arranged on a 2-D mesh, organised into 20 nodes of four NPUs each.[^17][^31] Each NPU contains 8 Neural Processing Engines (NPEs), each NPE acting as a compute unit for synaptic and neuronal operations, plus 100 kB of local SRAM (60 kB for spike/data buffering and 40 kB for weights).[^17] A small Cortex-M4 microcontroller serves as the on-chip host processor for configuration, sequencing and external I/O.[^17] The AKD1000 die runs at clocks up to 300 MHz, with synaptic weights configurable at 1, 2, 4 or 8 bits.[^17][^18][^31]
Independent reviews note that Akida does not implement the leaky-integrate-and-fire neuron model common in academic neuromorphic chips. Instead, it "operates similarly to an event camera, converting pixels to events and using Rank Order Coding (ROC) to encode the input," and propagates ternary-valued events through configurable convolutional or fully connected node functions.[^4] Akida supports convolutional neural networks, deep neural networks, recurrent neural networks and (in Akida 2.0) Vision Transformers, with native event-domain execution achieved by converting the trained model to a sparse spiking equivalent.[^4][^7]
Akida supports incremental, one-shot or few-shot on-device learning on its final classification layer.[^4][^32] As BrainChip and independent reviewers describe it, the AKD1000 "supports on-chip continual learning for the last layer only. The layer must be FullyConnected, it must have binary weights, and receive binary inputs."[^4] This restriction makes Akida well suited to personalisation tasks such as enrolling a new keyword for Keras-trained voice models, learning a new face or gesture, or adapting an industrial classifier to a particular machine without retraining the full network in the cloud. A 2024 IEEE paper by Karaman and collaborators demonstrated braking-intent personalisation on the AKD1000 using few-shot transfer learning on the last layer.[^32]
Akida exploits two complementary forms of sparsity. First, activations are sparse because event coding only fires when an input pixel, audio sample or sensor reading crosses a threshold. Second, weights themselves can be reduced to 1-bit, 2-bit or 4-bit values via BrainChip's quantization toolkit, exploiting weight sparsity in the first generation and 8-bit precision in Akida 2.0.[^7][^33] At the network-on-chip level, processing nodes exchange address-event packets rather than dense activation tensors, so power consumption tracks event rate. BrainChip and reviewers report typical AKD1000 power consumption in the tens of milliwatts for common edge workloads.[^4][^17]
Announced from Laguna Hills, California on 6 March 2023, Akida 2.0 is offered as IP only and is intended for license partners who want to integrate it into custom SoCs.[^7][^27] Key additions over Akida 1.0 include:
The AKD1500, announced as a tape-out on 29 January 2023 in GlobalFoundries' 22 nm FD-SOI (FDX) platform, is a co-processor reference chip that ports the Akida IP to an ultra-low-leakage process favoured for always-on sensor applications.[^25][^26] It offers PCIe Gen3 host attachment, quad/octal SPI for MCU-side integration and reports 800 GOPS within a 300 mW envelope.[^26] Publicly available product briefs position the AKD1500 as an MCU-attached co-processor for AIoT, industrial, consumer and automotive markets, validating that the same RTL can be migrated between foundries.[^25]
BrainChip's software environment, MetaTF, is a TensorFlow- and ONNX-compatible toolchain that lets developers train, quantize and deploy models for Akida.[^33][^34] MetaTF comprises four Python packages installable via pip:
By building on TensorFlow's TensorFlow and Keras front-ends, MetaTF lets machine learning engineers reuse standard training pipelines and frameworks like PyTorch via ONNX export, rather than authoring spiking models from scratch.[^33][^34]
| Variant | Form | Process | Status | Notable specs |
|---|---|---|---|---|
| Akida architecture (2018) | IP announcement and Akida Development Environment | n/a | Historical | First public disclosure 10 Sep 2018; sampling target Q3 2019.[^16] |
| AKD1000 | Standalone NSoC | TSMC 28 nm digital CMOS | Shipping | 80 NPUs, 1.2 M neurons, 10 B synapses, 300 MHz, 1/2/4-bit weights, Cortex-M4 host.[^17][^18][^31] |
| Akida 1.0 IP | Synthesizable RTL | Foundry-agnostic | Licensed | Same architecture as AKD1000; licensed by MegaChips, Renesas Electronics and others.[^12][^23] |
| AKD1500 | Co-processor reference chip | GlobalFoundries 22 nm FD-SOI | Sampling / silicon received | 800 GOPS at <300 mW, PCIe Gen3, quad/octal SPI for MCUs.[^25][^26] |
| Akida 2.0 IP | Synthesizable RTL | Foundry-agnostic | Available to early-access partners since Q3 2023 | Adds 8-bit weights, Vision Transformer, TENN cores, hardware skip connections, multi-pass execution.[^7][^4] |
| AKD1000 Mini PCIe / M.2 boards | Development boards | n/a | Shipping | Mini PCIe from USD 499 (Jan 2022); M.2 from USD 249 (Jan 2025).[^21][^22] |
BrainChip's highest-profile deployment to date is in the Mercedes-Benz Vision EQXX concept electric vehicle, unveiled at CES 2022.[^8][^35] Mercedes-Benz integrates Akida silicon for in-cabin "Hey Mercedes" wake-word detection, exploiting the chip's event-based design to keep the always-on listener at very low standby power. Mercedes-Benz stated at the EQXX reveal that BrainChip's neuromorphic solution was "five to ten times more efficient than conventional voice control" at spotting the wake word.[^8][^35] The Vision EQXX was promoted as the most efficient Mercedes-Benz vehicle ever built, claiming a single-charge range of more than 1,000 km, with neuromorphic processing contributing to the overall power budget.[^8] BrainChip has positioned the EQXX deployment as evidence that event-based inference can meaningfully improve automotive cabin AI energy efficiency relative to GPU-class or DSP-class accelerators.[^35]
Under the IP license signed in 2022, Renesas Electronics America committed to integrate Akida 1.0 IP into Renesas system-on-chip products, particularly the RZ vision MPU line, with on-going royalties tied to per-unit volume.[^12][^36] Renesas executives have publicly described licensed Akida processors as enabling "hyper-efficient acceleration for today's mainstream AI models at the edge, with advanced temporal convolution and vision transformers" in MCU-class power envelopes targeting industrial and consumer IoT and personalised healthcare.[^36] Renesas has also integrated Akida-related capabilities into its RZ/V high-end vision AI MPUs (which combine quad Cortex-A55 cores with the DRP-AI3 accelerator), with the RZ/V2H supported by Edge Impulse alongside the AKD1000.[^36][^37]
In November 2021, MegaChips Corporation, a Japanese fabless ASIC house, licensed the Akida IP for next-generation edge AI ASICs across automotive, IoT, cameras, gaming and industrial robotics markets.[^23] The license gives MegaChips access to the full Akida RTL, simulation testbenches, scripts and timing constraints needed for custom integration. BrainChip has since announced additional license and evaluation agreements with companies including Information Systems Laboratories (an early access partner working on defense applications) and Neuromorphyx (announced in 2026 as a strategic customer and go-to-market partner for the AKD1500).[^29][^38]
In late 2022, NASA's Ames Research Center at Moffett Field, California, acquired an Akida Early Access Evaluation Kit to assess the processor's suitability for spaceflight, citing its ability to run feed-forward networks without an external CPU, memory or deep learning accelerator.[^12] Independent academic groups have since published peer-reviewed and arXiv preprints exploring Akida for spacecraft pose estimation and onboard satellite ship detection, validating event-based inference under embedded power envelopes.[^29][^30]
On 9 December 2024, the U.S. Air Force Research Laboratory awarded BrainChip a US$1.8 million development contract for neuromorphic radar signal processing under topic AF242-D015, with a focus on micro-Doppler signature analysis for power- and thermally constrained weapon and platform systems.[^13] In April 2025, BrainChip and RTX's Raytheon announced a teaming agreement to deliver the AFRL prototype.[^28]
In late 2022, the cloud-based embedded ML platform Edge Impulse added official support for Akida, becoming BrainChip's reference tooling partner for developers without deep neuromorphic background.[^11][^24] Edge Impulse's classification and transfer learning blocks can target the AKD1000 directly, with a quantized model zoo optimised for Akida deployment.[^11] In March 2023, BrainChip announced that the Akida processor family integrates with the Arm Cortex-M85 processor, allowing Cortex-M based MCU designs to offload spiking inference to Akida co-processors.[^14]
Documented Akida deployments and prototypes span six broad domains:
BrainChip routinely positions Akida as competing not against NVIDIA H100-class data-centre GPUs but against the small AI accelerators bundled into low-power MCUs and at-sensor ASICs.[^7][^33]
Akida is the first neuromorphic processor IP that has been commercially licensed at scale into a mainstream MCU and ASIC supply chain. It is the only major neuromorphic family available both as a standalone reference chip and as foundry-portable IP, validated on at least two distinct processes (TSMC 28 nm and GlobalFoundries 22 nm FD-SOI).[^17][^25][^26] Whereas other widely cited neuromorphic chips have remained research-only platforms, Akida silicon ships on standard development boards and is supported by mainstream toolchains via MetaTF, Edge Impulse and Keras-based workflows.[^4][^33][^11] Its event-based design, with on-chip storage of weights in SRAM and address-event communication, illustrates how a strictly digital ("neuromorphic-inspired" rather than analog) implementation can deliver many of the sparsity advantages of more exotic mixed-signal neuromorphic chips while remaining manufacturable on commodity CMOS.[^4][^17]
Several limitations of Akida have been noted in independent reviews and academic benchmarks.
Akida sits in a small but expanding field of dedicated neuromorphic and event-based AI accelerators. The most widely cited contemporaries are described below; none currently match Akida's combination of commercial shipping silicon and foundry-portable IP.
| Project | Sponsor | Type | Status (2025–2026) |
|---|---|---|---|
| Akida (AKD1000/1500, Akida 2.0 IP) | BrainChip | Digital event-based, on-chip learning, IP + chips | Shipping; licensed; deployed in Mercedes EQXX.[^7][^17][^8] |
| Loihi 2 | Intel Labs | Asynchronous digital neuromorphic, research | Research-only; used in the Hala Point system at Sandia National Laboratories (1.15 billion neurons), with continued work on subsequent generations.[^40][^41] |
| NorthPole | IBM Research | Inference-only neuromorphic-inspired digital accelerator | Research prototype; described as TrueNorth's successor with roughly 4,000x higher performance on inference workloads.[^42] |
| TrueNorth | IBM Research | Digital neuromorphic, low-power | Earlier-generation IBM platform; not in active development; effectively superseded by NorthPole.[^42] |
| DynapCNN / Speck | SynSense AG (founded 2017, Zürich / Chengdu, formerly aiCTX) | Mixed-signal SNN processor and integrated vision SoC | Commercially available; targets event camera vision and audio.[^43] |
| GrAI VIP / GrAI One | GrAI Matter Labs | Event-driven NeuronFlow architecture | Commercial; targets industrial robotics and anomaly detection.[^44] |
| Pulsar | Innatera Nanosystems | Mixed-signal SNN for always-on sensor SoCs | Commercial; sub-milliwatt audio and sensor inference.[^45] |
Unlike Intel's Loihi 2 and IBM's NorthPole, which remain primarily research vehicles distributed via collaborator programs, Akida is sold as commercial development boards and licensed RTL.[^40][^42] Unlike SynSense's mixed-signal Speck and DynapCNN, Akida is fully digital and synthesisable on any modern logic node, easing integration into existing ASIC flows.[^43][^4] These complementary positions make BrainChip the most prominent example of "neuromorphic computing as a commercial IP play" rather than as a long-horizon research program.