NEWS: 公告在東京證券交易所JASDAQ標準市場新上市


神經形態處理器(2021-2022):Kisaco 領導力圖

Kisaco Leadership Chart on Neuromorphic Processors 2021-22

出版商 Kisaco Research Limited 商品編碼 1007523
出版日期 內容資訊 英文 39 Pages
商品交期: 最快1-2個工作天內
神經形態處理器(2021-2022):Kisaco 領導力圖 Kisaco Leadership Chart on Neuromorphic Processors 2021-22
出版日期: 2021年05月13日內容資訊: 英文 39 Pages

神經擬態計算誕生於基於與人腦具有直接生物學聯繫的技術的人工智能 (AI) 研究。大腦是一個模擬系統,它使用電尖峰在神經元之間發送信號。同樣,許多為其處理器選擇神經形態標籤的供應商在其模擬系統(通常是電路)中使用尖峰神經網絡 (SNN)。

在本報告中,我們將解釋區分神經擬態與傳統 AI 的具體點,評估神經擬態供應商,並介紹每個供應商的簡介。


  • Kisaco Research 的觀點
  • 動機
  • 主要調查結果
  • 解決方案分析:神經形態處理器
  • 技術情況
  • 市場情況
  • 解決方案分析:供應商比較
  • Kisaco 神經形態處理器領導力圖 (KLC)
  • 神經形態處理器供應商的比較
  • 神經形態處理器的 KLC
  • 供應商分析
  • AIStorm,Kisaco 評級:領導者
  • Kisaco 的評估
  • Aspinity、Kisaco 評級:選擇不參加 KLC
  • BrainChip、Kisaco 評級:領導者
  • Kisaco 的評估
  • iniVation、Kisaco 評級:選擇不參與 KLC
  • Innatera Nanosystems, Kisaco 評級:初創企業
  • Kisaco 的評估
  • 英特爾、Kisaco 評級:選擇不參與 KLC
  • Rain Neuromorphics、Kisaco 評級:創新者
  • 簡介
  • Rain 的模擬處理單元 (APU)
  • APU 芯片流片
  • Rain的神經學習能量平衡算法
  • Rain 的 APU 3D 突觸架構
  • Kisaco 的評估
  • SynSense,Kisaco 評級:競爭者
  • Kisaco 的評估
  • 附錄
  • 供應商解決方案選擇
  • 選擇標準
  • 調查方法
  • KLC 的定義
  • Kisaco 研究評估
  • 參考資料
  • 謝謝
  • 作者
  • Kisaco Research 的分析網絡
  • 版權聲明/免責聲明


Neuromorphic computing arises out of artificial intelligence (AI) research based on technology that has direct biological links with the human brain. The brain is an analog system that uses electrical spikes to transmit signals between neurons, similarly many vendors that choose the neuromorphic label for their processors use spiking neural networks (SNNs) in an analog system, typically electric circuits. However, other such vendors choose to use digital devices with a SNN, and yet again others use an analog device with non-spiking, continuous value signal neural networks.

Neuromorphic computing emerged in the 1990s but has had a slow evolution due to the challenges in training neural networks without use of a global learning rule, such as backpropagation. Backpropagation is critical in (non-spiking) deep learning neural networks, and it uses information at the output of the network to update neurons (more exactly the synapse weights) upstream in the network. To our best understanding at time of writing the human brain does not use a global learning rule and it has taken time for local learning rules to emerge for neuromorphic architectures, with success in the last two years, and this has given birth to a surge in startups in this space.

To find the common ground that can be pinned to the neuromorphic label there are two key characteristics: low power consumption and high efficiency, typically in the form of highly sparse connectivity - both characteristics of the human brain. We delve deeper into what exactly distinguishes neuromorphic from the traditional AI in this report. We also assess neuromorphic vendors with processors that span the range of possible architectures and learning rules. The Kisaco Leadership Chart (KLC) compares five of the pioneering vendors side by side: AIStorm, BrainChip, Innatera Nanosystems, Rain Neuromorphics, and SynSense. In addition to our in-depth profiles on these vendors, we have three more vendors profiled in-depth: Aspinity, Intel, and Inivation.

What you will learn:

  • How neuromorphic processors differ from other AI processors on the market.
  • Which is the strongest market segment for neuromorphic processors.
  • Our report has assessed five neuromorphic processor vendors and we provide a high-level heatmap on the key features available
  • We compare the processors from the five participating vendors side by side and assess these in our Kisaco Leadership Chart.
  • We provide an in-depth profile on each of the participating vendors together with three strengths and three weaknesses.

Table of Contents

  • Kisaco Research View
  • Motivation
  • Key findings
  • Solution Analysis: Neuromorphic processors
  • Technology landscape
  • Market landscape
  • Solution analysis: vendor comparisons
  • Kisaco Leadership Chart on Neuromorphic Processors 2020-21
  • Neuromorphic processor vendor comparisons
  • The KLC chart for neuromorphic processors
  • Vendor analysis
  • AIStorm, Kisaco evaluation: Leader
  • Kisaco Assessment
  • Aspinity, Kisaco evaluation: chose not to participate in KLC
  • BrainChip, Kisaco evaluation: Leader
  • Kisaco Assessment
  • iniVation, Kisaco evaluation: chose not to participate in KLC
  • Innatera Nanosystems, Kisaco evaluation: Emerging Player
  • Kisaco Assessment
  • Intel, Kisaco evaluation: chose not to participate in KLC
  • Rain Neuromorphics, Kisaco evaluation: Innovator
  • Introduction
  • The Rain Analog Processing Unit (APU)
  • Taping out APU chips
  • The Rain energy equilibrium algorithm for neural learning
  • The Rain APU 3D synaptic architecture
  • Kisaco Assessment
  • SynSense, Kisaco evaluation: Contender
  • Kisaco Assessment
  • Appendix
  • Vendor solution selection
  • Inclusion criteria
  • Methodology
  • Definition of the KLC
  • Kisaco Research ratings
  • Further reading
  • Acknowledgements
  • Author
  • Kisaco Research Analysis Network
  • Copyright notice and disclaimer


  • Figure 1: Comparing the brain, neuromorphic chip, and GPU in AI inference mode.
  • Figure 2: Comparing the KLC vendors on key technology features.
  • Figure 3: Heat map analysis of participating vendor technical features.
  • Figure 4: Kisaco Leadership Chart on Neuromorphic Processors 2020-21.
  • Figure 5: Kisaco Leadership Chart on Neuromorphic Processors 2020-21: ranking of vendors.
  • Figure 6: Comparing digitization of input with AIStorm's AI-in-Sensor.
  • Figure 7: AIStorm imager with "always on" cascaded wake-on approach.
  • Figure 8: Aspinity AnalogML typical use case.
  • Figure 9: Aspinity AnalogML core.
  • Figure 10: BrainChip Akida NPU architecture and IP solution.
  • Figure 11: BrainChip Akida software development environment and training workflow.
  • Figure 12: Inivation sensors only capture image changes.
  • Figure 13: Innatera spiking neural processor architecture.
  • Figure 14: Innatera spiking neural processor: segment zoom view.
  • Figure 15: Audio processing with a temporal feedforward SNN on the Innatera SNP.
  • Figure 16: Loihi benchmarks: Recurrent networks with bio-inspired properties give the best results.
  • Figure 17: Loihi Research Systems currently available.
  • Figure 18: Loihi projects pursued by INRC members.
  • Figure 19: Efficient sensing and pattern learning.
  • Figure 20: Rain's Analog Processing Units (APUs).
  • Figure 21: Rain APU 3D architecture vs traditional 2D crossbar.
  • Figure 22: SynSense hardware families.
  • Figure 23: CNN based processing stack. Backpropagation-based training of visual features.