Neuromorphic edge AI chip debuts on Raspberry Pi and Comet Lake dev kits

Neuromorphic edge AI chip debuts on Raspberry Pi and Comet Lake dev kits

Oct 25, 2021 — by Eric Brown

142 views

BrainChip has launched two “Akida Development Kits” for its self-learning, low-power “Akida NSoC” neural networking chip for edge AI. One uses the Raspberry Pi CM4 and the other a Comet Lake-S based Shuttle PC system.

BrainChip Holdings has opened pre-orders for two development kits that showcase its Akida neural networking processor (Akida NSoC). The Linux-driven, $4,995 Akida Development Kit — Raspberry Pi and Linux and Win 10 compatible $9,995 Akida Development Kit — Shuttle PC implement the Akida NSoC via a mini-PCIe module equipped with BrainChip’s AKD1000 silicon.

Akida Development Kit — Raspberry Pi (left) and Akida Development Kit — Shuttle PC
(click images to enlarge)

The Akida NSoC is a neuromorphic, event-based AI processor that mimics brain processing, espeiclaly the capability for “spiking” processing. The spiking neural networks (SNNs) enabled by the chip express information via both spatial and temporal sequences. As explained in this EETimes report, spikes typically result from changes in sensor data, including color changes from an event-based camera.

The Akida NSoC can process standard convolutional neural networks (CNNs) in addition to SNNs. The technology does so by converting CNNs into SNNs and running inference in the event domain. This capability allows for on-chip self-learning, which in turn enables the technology to handle changes in the perceived environment more flexibility than most AI chips. This self-learning and re-learning capability can “eliminate the need for data roundtrips to centralized CPUs for retraining,” says BrainChip.

Akida IP architecture (left) and BrainChips benchmarks showing MAC operations required for object classification inference (dark blue is CNN in non-event domain; light blue is Akida with event domain; green is event domain with further activity regularization)
(click images to enlarge)
Source: BrainChip via EETimes

Built for edge computing, Akida NSoC reduces processing cycles and latency by focusing on key events while discarding “data with no value,” says BrainChip. This translates into lower power consumption, with power budgets limited to microwatts or milliwatts, says the company. The technology is also touted for its performance, although we saw no TOPS rating. EETimes posted the chart shown above benchmarking for Akida MAC operations.

The Akida NSoC architecture is built on an array of up to 20 Nodes, each of which contains 4x Neural Processing Units for up to 80 NPUs in total. Like the brain, nodes are interconnected via a mesh network. Scalability is further improved by node parallelization and the ability to utilize a smaller group of nodes “in a recirculating fashion.” Akida NSoC is powered by a “M-Class CPU with FPU and DSP,” which likely means a Cortex-M core.

Akida workflow diagram (left) and MetaTF Development Environment architecture
(click images to enlarge)

While many AI processors focus on vision and audio, Akida is also suitable for olfactory, gustatory, and vibration/tactile sensing applications. Applications include smart home, city, transportation, and health, with specific examples including home automation and remote controls, industrial IoT, robotics, security cameras, sensors, unmanned aircraft, autonomous vehicles, medical instruments, object detection, sound detection, odor and taste detection, gesture control, and cybersecurity.

The Akida Development Kits ship with BrainChip’s MetaTF Development Environment, a machine learning framework that leverages Python alongside tools and libraries such as NumPy. MetaTF includes an Akida Execution Engine chip simulator, as well as data-to-event converters and a “model zoo” of pre-trained models.

Akida Development Kit — Raspberry Pi

The Raspberry Pi dev kit is enclosed in a 200.6 x 105 x 37mm chassis. Inside, a carrier board is powered by the up to 1.5GHz, quad-core, Cortex-A72 Raspberry Pi Compute Module 4. You can choose from CM4 models with 1GB to 8GB LPDDR4 and 8GB to 32GB eMMC. A CM4 with WiFi/Bluetooth is optional, and there appears to be an additional option on the carrier board for BT 5.0 with BLE.

Akida Development Kit — Raspberry Pi top view (left) and internal view with Akida-equipped mini-PCIe card
(click images to enlarge)

The Akida Development Kit — Raspberry Pi is equipped with a microSD slot along with GbE, 2x HDMI, 2x USB, and a micro-USB client port. A U.FL antenna connector and fan are also available.

There are 2x 100-pin high-density connectors to connect the CM4, as well as 40-pin GPIO, 2x MIPI-DSI, and 2x MIPI-CSI interfaces. The specs also list PCIe, but this is likely used up by the mini-PCIe slot holding the Akida module.

Akida Development Kit — Shuttle PC

The Intel-based version of the kit is a modified version of a Shuttle PC computer equipped with a 10th Gen Comet Lake-S processor. The 250 x 200 x 78.5mm desktop PC dwarfs the Shuttle PC mini-PCs we have covered, such as the Apollo Lake based EN01.

Akida Development Kit — Shuttle PC rear view (left) and internal view
(click images to enlarge)

The system supports up to 64GB DDR4 and offers a 2.5-inch SATA bay and M.2 M-key 2280 slot with SATA and NVMe support. An M.2 E-key 2230 slot supports an optional WiFi/BT module. There is also a full-size PCIe 3.0 x16 slot for up to 75W cards.

The Akida Development Kit — Shuttle PC provides GbE, HDMI 2.0a, and VGA ports. You also get 4x USB 3.2 Gen1 and 5x USB 2.0 ports, one of which is an internal USB 2.0 port for a USB stick. Other features include 2x audio jacks, an optional COM port, VESA mounting, and a 180W, 19.5V adapter. Dual fans with heatpipe support a 0~50°C range.

Further information

The Akida Development Kit — Raspberry Pi is available for pre-order at $4,995 and the Akida Development Kit — Shuttle PC can be ordered for $9,995. No ship date was listed. More information may be found in BrainChip’s announcement and Akida Development Kits product page. More on Akida NSoC may be found on the Akida NSoC page and this detailed user guide.