FeFET Charge-Based Computing Cuts Inference Loss Below 0.2 Points, Beating 5.7-Point Current-Based Drop
Updated
Updated · Nature.com · Jul 21
FeFET Charge-Based Computing Cuts Inference Loss Below 0.2 Points, Beating 5.7-Point Current-Based Drop
2 articles · Updated · Nature.com · Jul 21
Summary
Researchers reported that charge-based FeFET in-memory computing kept average HDC language-classification accuracy loss below 0.2 percentage points, versus as much as 5.7 points for current-based designs.
Device tests on 28 nm fabricated FeFETs and TCAD simulations showed much lower variability in non-volatile capacitor mode, including a 0.66% sigma-to-mean ratio in simulated on-state capacitance and 72.6% less variability than current readout.
At circuit level, the team built an 8-bit SAR-ADC array around a 255/256-cell charge-based CAM block, reaching about 90%-91.7% sensing accuracy with 40.64 pJ energy per compute cycle.
That lower variability translated into system resilience: the charge-based array's average error probability was 8.1%, yet HDC inference remained largely intact because the architecture scales better than current-based arrays, whose nonlinear accumulation limits block size to about 15 cells.
This new AI chip promises huge efficiency gains, but what are its hidden trade-offs in speed or durability for real-world applications?
Could this memory-computing breakthrough solve the energy crisis in data centers, not just on small edge devices?
As startups race to market, can legacy chipmakers pivot to charge-based computing before being left behind in the edge AI revolution?
FeFET In-Memory Computing Achieves 96.6% AI Accuracy with 900x Power Reduction: The 2026 Charge-Based Breakthrough
Overview
Artificial intelligence computing is being transformed by the urgent need to overcome the limitations of traditional architectures, where the von Neumann model separates memory and processing. This separation creates a 'memory wall,' causing most system energy to be spent on moving data and making data access much slower than computation. As AI models grow more complex, these issues severely limit efficiency. In-Memory Computing (IMC) has emerged as a promising solution, performing computations directly within memory to reduce energy use and latency. This breakthrough paves the way for more efficient and powerful AI systems.