Is This Analog AI's Best Hope?

analog-aichip-architecturein-memory-computingenergy-efficiencyneural-networks

Abstraction: Capacitor-based analog AI chip overcomes noise limitations of current-based designs

Key points:

Traditional analog AI stores neural network weights as conductances in RRAM/phase-change memory and computes matrix multiplies via Ohm's Law + Kirchhoff's Current Law, but device variability and temperature dependence cause noise accumulation across layers
EnCharge AI's EN100 stores weights in digital memory but computes using capacitors: charge = capacitance × voltage; capacitance is geometry-determined in CMOS, highly controllable and noise-resistant
EN100 delivers 200 TOPS at 8.25W for PC laptops; a 4-chip card targets 1,000 TOPS for workstations; claims up to 20x better performance-per-watt than digital competitors
Technology is based on switched-capacitor operations (used for decades in ADCs); key innovation is applying it to in-memory computing for AI
EnCharge raised $100M from Samsung Venture, Foxconn, and others; products with early-access developers as of 2025
Competitors D-Matrix and Axelera use digital compute-in-memory (custom SRAM); traditional analog competitor Sagence also in market; Nvidia GB10/GB300 targets same PC/workstation segment

Connections: Encharge AI · Naveen Verma · Nvidia · Analog Computing · In Memory Computing · Hardware Accelerator