Abstract
Phase Change Memory (PCM) is a leading candidate for next-generation non-volatile memory, offering speeds that bridge the gap between Flash and DRAM. However, read latency remains a critical bottleneck due to parasitic capacitance in long bit lines. This paper presents a 64 Mbit PCM chip fabricated in a 40 nm CMOS process that achieves a read latency of approximately 50 ns. The performance gains are driven by a novel sense amplifier (SA) architecture featuring shared bias voltages, optimized feedback loops, and a hierarchical bit line structure.
Introduction
The read operation in PCM involves sensing the resistance of a chalcogenide material, which varies between high-resistance amorphous (RESET) and low-resistance crystalline (SET) states. In large-scale memory arrays, the "electrical sludge" of parasitic capacitance in long bit lines often slows down this sensing process. Furthermore, traditional sense amplifiers require dedicated reference branches for every unit, consuming significant chip area. To address these challenges, researchers proposed a multi-faceted architectural overhaul to minimize capacitance and maximize sensing efficiency.
Architectural Innovations
The researchers implemented four key strategies to enhance read performance:
1. Shared Bias Voltage for Sense Amplifiers
Traditional designs often assign a dedicated reference resistor branch to every sense amplifier. In this work, the team introduced a shared bias voltage ($V_b$) architecture. By allowing 384 SAs to share a single group of voltage generators, the total area of the sensing circuitry was significantly reduced without compromising accuracy.
2. Feedback Loop Optimization
In conventional SA designs, the bit line—and its associated parasitic capacitance—is part of the amplifier's feedback network. This inclusion makes the system sluggish. The authors decoupled the bit line from the feedback loop, allowing the internal amplifier circuitry to stabilize much faster and providing a crisper transient response.
3. Separated Transmission Gates
PCM programming requires high currents (mA range), while reading requires very low currents (μA range). While a single large gate typically handles both, this design separates them into Programming Transmission Gates (PTG) and Reading Transmission Gates (RTG). The RTG, optimized solely for low-current read operations, is much smaller, which further strips away unnecessary parasitic capacitance.
4. Hierarchical Bit Line Architecture
The 64 Mbit array is partitioned into 32 "tiles" of 2 Mbit each. A multi-level hierarchy (Block Bit Line → Local Bit Line → Read Bit Line) ensures that during a read operation, the system only interacts with a short segment of the wiring. This minimizes the physical load the SA must overcome.
Experimental Results
The chip was fabricated using a 40 nm CMOS process, featuring 64 Mbit of capacity plus 32 Mbit of redundancy for Error Correction Code (ECC).
- Read Latency: While simulations predicted a sense time of 42 ns, hardware measurements on the fabricated chip demonstrated a sense time of approximately 50 ns.
- Comparative Performance: This represents a 2x speedup over a reference 130 nm PCM design, which exhibited a 98 ns sense time.
- Area Efficiency: The shared bias generator approach reduced the block area for 384 SAs to just 46,080 μm², demonstrating high efficiency through shared circuitry.
- Reliability: Experimental data confirmed a clear resistance margin of approximately two orders of magnitude between SET and RESET states, ensuring robust data integrity.
Conclusion
By isolating the sense amplifier from bit line parasitic capacitance and utilizing a shared-bias hierarchical architecture, this 40 nm PCM chip demonstrates that non-volatile memory can achieve read speeds competitive with high-performance systems. Future work will likely focus on managing parasitic resistance as scaling continues toward 28 nm and investigating the long-term endurance of these high-speed sensing circuits.
Summary updated