BrainChip launches neuromorphic hardware accelerator

BrainChip launches neuromorphic hardware accelerator

Technology News |
By Peter Clarke

The PCIe Gen.3 card is for use as a hardware accelerator for applications developed within the previously released BrainChip Studio, a software product that was pitched at vision in surveillance, security and defense applications.

BrainChip Accelerator is an 8-lane, PCI-Express add-in card that increases the speed and accuracy of the object recognition function of BrainChip Studio software by up to a factor of six, while increasing the simultaneous video channels monitored to up to 16 per card.

BrainChip’s approach to neural networks differs from other companies offering artificial neural networks, in that BrainChip models the brain and its neuron and synapse architecture more closely and includes models of spiking signals and connection reinforcement and connection inhibition. Bob Beachler, senior vice president of marketing and business development, said that this makes BrainChip the first company with a commercial offering of a true neuromorphic hardware, albeit implemented in a Xilinx, Kintex Ultrascale 115 chip.

The spiking neural network approach does not require the protracted, usually cloud-based training programs, required for convolution neural networks and even allows “one-shot” training and self-learning procedures. One of the things the spiking neural network (SNN) excels at is finding patterns in noisy environments with low computational and energy requirements, Beachler said.

For applications such as facial recognition this does require the application of some conventional computer graphics pre-processing to perform scaling and define a region of interest within a scene.

For now the hardware is being offered to accelerate BrainChip Studio which has been deployed with law enforcement and intelligence organizations to rapidly identify objects in large amounts of archived or in live streaming video. By processing multiple video streams simultaneously, the BrainChip Accelerator add-in card enables these organizations to search increasing amounts of video faster, with a higher probability of object recognition and lower total cost of ownership. The system learns from a single low-resolution image, which can be as small as 20 x 20 pixels, and excels in recognition in low-light, low-resolution, noisy environments.

Next: Six cores

The processing is done by six BrainChip Accelerator cores within the Xilinx FPGA. Each core performs fast, user-defined image scaling, spike generation, and spiking neural network comparison to recognize objects. Scaling images up and down increases the probability of finding objects, and due to the low-power characteristics of spiking neural networks, each core consumes approximately one watt while processing up to 100 frames per second.

When compared with GPU-accelerated deep learning classification neural networks such as GoogleNet and AlexNet, this is a 7x improvement of frames/second/watt.

“BrainChip’s spiking neural network technology is unique in its ability to provide outstanding performance while avoiding the math intensive, power hungry, and high-cost downsides of deep learning in convolutional neural networks,” said Christoph Fritsch, senior director for industrial, scientific and medical business at Xilinx, in a statement issued by BrainChip.

BrainChip Accelerator is compatible with Windows or Linux computing platforms and is currently available to select law enforcement and intelligence agencies as an integrated server appliance. It may also be purchased as an add-in card for security integrators and OEMs.

BrainChip Studio is licensed on a per-video-channel basis at about $4,000 with a 20 percent annual maintenance fee. The accelerator is being offered at about $10,000 per card. Beachler added: “You can use the card for other applications or research but within the context of one-shot training.”

A dedicated processor based on the SNN approach could be the next development from BrianChip. The company has previously discussed a SNAP 64 processor, but Beachler said that for now the focus is on meeting the detailed needs of applications via the FPGA-based hardware.

Related links and articles:

News articles:

Neural network teaches itself to count cars

BrainChip appoints former Exar CEO to lead company

BrainChip buys French computer vision firm

If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :    eeNews on Google News


Linked Articles