====== Analog AI Chips ======

**Analog AI chips** are specialized semiconductor processors that perform AI inference using continuous analog signals rather than binary digital logic. By executing computations directly within memory arrays, analog chips achieve ultra-low power consumption and high-speed matrix operations, making them particularly suited for edge devices and always-on AI applications. ((Source: [[https://www.precedenceresearch.com/analog-ai-chip-market|Precedence Research — Analog AI Chip Market]]))

===== How They Work =====

Analog AI chips rely on **in-memory computing (IMC)**, where data processing occurs within memory arrays rather than shuttling data between separate memory and processor units. This eliminates the von Neumann bottleneck that plagues digital architectures. ((Source: [[https://www.technavio.com/report/analog-ai-chip-market-industry-analysis|Technavio — Analog AI Chip Market]]))

==== Resistive Crossbar Arrays ====

The core technology uses **resistive crossbar arrays** built from non-volatile memories such as Resistive RAM (ReRAM/RRAM):

  * Each memory cell's conductance represents a synaptic weight in a neural network
  * Voltage inputs across rows produce analog currents that sum along columns via Kirchhoff's current law
  * This yields multiplication results without clocked digital steps — up to 1000x faster throughput and 100x energy efficiency compared to digital processors for matrix operations ((Source: [[https://www.trendforce.com/news/2025/10/21/news-chinese-scientists-developed-a-novel-chip-crossing-a-century-old-hurdle/|TrendForce — RRAM Breakthrough]]))

Recent breakthroughs in RRAM precision have achieved digital-rival accuracy with five orders of magnitude improvement in precision. ((Source: [[https://www.trendforce.com/news/2025/10/21/news-chinese-scientists-developed-a-novel-chip-crossing-a-century-old-hurdle/|TrendForce — Chinese Scientists RRAM Chip]]))

==== Other Approaches ====

  * **Neuromorphic chips** mimicking brain spike patterns for real-time learning ((Source: [[https://www.technavio.com/report/analog-ai-chip-market-industry-analysis|Technavio — Analog AI Chip Market]]))
  * **Analog reservoir computing** — TDK's cerebellum-like circuits for time-series edge tasks ((Source: [[https://www.tdk.com/en/news_center/press/20251002_01.html|TDK — Reservoir AI Chip]]))
  * **Phase-change memory** — IBM's approach using material state changes to encode weights

===== Companies =====

^ Company ^ Technology ^ Applications ^ Notable Developments ^
| Mythic | Analog in-memory computing | Automotive ADAS | Honda partnership for 100x energy-efficient IMC chips (2026) |
| IBM | Phase-change memory prototypes | Complex AI tasks | Low-energy inference prototypes (2024-2025) |
| TDK | Analog reservoir circuits | Robots, sensors | Reservoir AI chip prototype for real-time edge learning (2025) |
| EnCharge | Analog multiplier accelerators | Broad AI inference | $100M DARPA funding; 20x energy reduction (2025) |
| Peking University | High-precision RRAM crossbar | Training/inference, 6G | RRAM chip with digital-level precision (2025) |
| Syntiant | Neural decision processors | Voice, driver monitoring | Auto OEM deal for cabin AI (2025) |

===== Advantages Over Digital Chips =====

  * **Energy efficiency** — Up to 100x lower power for continuous inference versus microcontrollers and GPUs ((Source: [[https://www.precedenceresearch.com/analog-ai-chip-market|Precedence Research — Analog AI Chip Market]]))
  * **Speed** — Order-of-magnitude lower latency for matrix operations; 1000x throughput for specific workloads ((Source: [[https://www.fanaticalfuturist.com/2025/03/low-energy-analogue-ai-chips-gets-a-100-million-boost-from-darpa/|Fanatical Futurist — DARPA Analog AI]]))
  * **Edge suitability** — Standalone operation on battery-powered devices without cloud connectivity
  * **Reduced data movement** — Eliminates 80% of digital AI power waste from memory-to-processor transfers

==== Limitations ====

  * Noise and precision challenges requiring software compensation
  * More complex fabrication compared to established digital processes
  * Limited programmability versus general-purpose digital chips

===== Market Outlook =====

The analog AI chip market is projected to grow from $251 million in 2025 to $2.45 billion by 2035, driven primarily by inference workloads at the edge. In-memory computing holds 38% market share, with inference applications comprising 52% of total demand. ((Source: [[https://www.precedenceresearch.com/analog-ai-chip-market|Precedence Research — Analog AI Chip Market]]))

===== See Also =====

  * [[neural_processing_unit|Neural Processing Unit (NPU)]]
  * [[sram_centric_chips|SRAM-Centric Chips]]
  * [[meta_mtia_chip|Meta MTIA Chip]]

===== References =====