オーストラリアBrainChip Holdings(ブレインチップ ホールディングス)CEOのSean Hehir氏が来日した。同氏によれば、同社はニューロモロフィック(脳型)演算を行うIP(Intellectual Property)コアを商用化した唯一の企業である。現在、ルネサス エレクトロニクスとメガチップスの2社が同IPコアのライセンスを取得している。
xtech.nikkei.com
Licensed to Renesas and MegaChips, brain-inspired computing IP core BrainChip
Ikutaro Kojima
Nikkei Cross Tech / Nikkei Electronics
2023.06.07
Paid members only
2450 characters total
Australian BrainChip Holdings CEO Sean Hehir (
Fig. 1 ) came to Japan and gave a presentation to the Japanese media on May 30, 2023. According to him, the company is the only company that has commercialized an IP (Intellectual Property) core that performs neuromorphic (brain-type) computation. The product name of the IP core is "Akida", and currently two companies, Renesas Electronics and MegaChips
News Release,have obtained licenses for Akida (
Fig. 2 ). In addition, Germany's Mercedes-Benz and others are said to be evaluating Akida.
Figure 1 Mr. Sean Hehir
(Photo: Nikkei Cross Tech)
[Tap image to enlarge]
Figure 2 Outline of BrainChip
(Source: BrainChip)
[Tap image to enlarge]
Machine learning, generally called AI (artificial intelligence), is based on mathematical models such as CNN (Convolutional Neural Network) and DNN (Deep Neural Network). Both CNN and DNN are processed by circuits that imitate the operation of the brain, but brain-type arithmetic circuits operate more like the brain than such circuits. However, CNN/DNN processing circuits and brain-type arithmetic circuits are quite similar when implemented on semiconductors (ICs). A neural network is processed by arranging a large number of small-scale operation nodes in a matrix. Akida also has a matrix of operation nodes (
Fig. 3 ).
Fig. 3 Overview and chip structure of Akida (1st generation product)
(Source: BrainChip)
[Tap image to enlarge]
When I asked Mr. Hehir about the difference between Akida and other CNN/DNN processing circuits, he said, "Akida's processing is event-based, and it is characterized by eliminating the need to process zero model parameters. More than half of the parameters in neural network models are zero, and by taking advantage of that property, Akida has higher performance and power efficiency than other CNN/DNN processing circuits," he said. Akida seems to perform the same processing as pruning in CNN/DNN.
According to him, by using the company's SDK (Software Development Kit), CNN/DNN models developed (learned) in frameworks such as TensorFlow and PyTorch can be deployed on Akida for inference processing. He described four advantages over other institutions' Neural network processing units (NPUs) for processing CNN/DNN models. First, as mentioned above, Akida is event-based. The second is the application of CIM (Computation In Memory) technology. "Each of Akida's computing nodes has a sufficient amount of built-in SRAM, so data communication with external SRAM is minimal and power consumption is kept low," he said.
Third, all processing can be done within Akida. “NPUs of other institutions often process only some layers of CNN/DNN, and other layers are processed by CPU cores.The effect of NPUs is limited. If so, the load on the CPU core can be considerably reduced." (Mr. Fourth, it is possible to perform optimization learning processing that matches the execution environment. Currently, when using machine learning, it is common to use a model learned in the cloud and perform inference at the edge. For this reason, the execution environment such as lighting and cameras differs between learning and inference. Expected processing accuracy is often not obtained due to the difference in the environment. Akida is capable of learning to absorb this difference, and the results can be added to models trained in the cloud (
Fig. 4 ). This makes it possible to optimize for the execution environment.
Fig. 4 Learning for optimization is possible at the edge
(Source: BrainChip)
[Tap image to enlarge]
Hehir also explained the difference between Akida and brain-like computing circuits developed by other institutions. First, as introduced at the beginning of the article, Akida is the only brain-type arithmetic circuit IP that has been commercialized. The second is that Akida is a completely digital circuit. “There are cases where brain-type calculations are processed by analog circuits, but analog circuits are difficult to transfer to other manufacturing processes. a Chip) can also be integrated” (Mr.