AI Chip Rides Novel Networks

　　An emerging company with a novel machine-learning technology and equally unique financial structure debuted its first hardware product today. BrainChip rolled out an FPGA-based accelerator for its spiking neural network (SNN) software and hopes to deliver an ASIC within two years to expand its existing markets.

　　SNNs are related but different from the convolutional neural nets (CNNs) now widely used and, some say, hyped by web giants for jobs like voice and image recognition. SNNs use a simpler, one-shot training method and are well-suited to tasks such as face recognition in low-resolution and noisy environments such as surveillance video. To date, BrainChip’s products are used mainly by law enforcement and security.

　　The BrainChip accelerator packs six SNN cores in a Xilinx Kintex chip on a PCI Express board processing video at up to 600 frames/s at about 15 W max. It provides as much as a six-fold performance boost for the BrainChip Studio software for x86 computers that the company rolled out in July at a cost starting at $4,000 per video channel. The company first described its architecture in late 2015.

　　Today’s emerging CNN chips essentially accelerate sparse linear algebra to shorten training loops. By contrast, BrainChip’s accelerator speeds processes in digital pathways said to mimic neural synapses, reinforcing or inhibiting traffic and setting thresholds as appropriate.

　　The card will be available at the end of September at a $10,000 list price for single units. The company will sell the card to system integrators. It also may sell integrated bundles of cards, software, and servers directly to end users.

　　The company claims that the card is the first commercial hardware to accelerate SNNs. IBM’s True North is more widely known but has been more of a general-purpose research vehicle, although the U.S. Air Force said in June that it would use it in a supercomputer. Stanford and the European Union also support research efforts in SNN accelerators.

　　BrainChip licensed technology for an ASIC to accelerate unsupervised learning in SNNs from a research group in Toulouse, France. The company is studying the potential in automotive, cybersecurity, financial, and medical markets to determine how to tailor the silicon that could be available in 12 to 24 months.

　　The company got its start 10 years ago as a spinout from the Toulouse University research effort that was creating custom software for users in France. BrainChip now consists of a software team in Toulouse, a hardware group in southern California, and last year, it brought on new management mainly in Silicon Valley.

　　A co-founder from Australia balked at financial terms of traditional venture capitalists. As an alternative, the team engineered a reverse takeover of an underactive mining company in Australia and raised $15 million from public investors on the Australian stock market, where it is now listed as a small-cap stock.

　　The company would not comment on plans for further fundraising, which it will clearly need to fund ASIC development while its still-meager software and card revenues slowly expand.