Introduction

Artificial intelligence models are growing faster than the hardware that runs them, creating a bottleneck often called the memory wall. A startup spun out of Georgia Tech says it has a potential answer: a new memory layer that stacks vertically to deliver speed close to traditional SRAM while using far less space on a chip.

Nagare Labs recently emerged from stealth with $5 million in seed funding led by Playground, the venture firm where former Intel CEO Pat Gelsinger is a general partner. The company’s technology, called Z2RAM, is designed to sit alongside existing memory rather than replace it, offering AI chips and robotics systems a faster way to access data.

What Happened

Professors Suman Datta and Shimeng Yu founded Nagare Labs to commercialize Z2RAM. Datta, a former Intel transistor researcher, describes the name as a reference to two concepts: stacking in the vertical z-axis and eliminating the capacitor found in conventional memory cells.

Instead of storing data as charge on a capacitor, each Z2RAM cell uses two transistors. One transistor writes the bit, while a second reads it without disturbing the stored charge. Because the charge sits on the read transistor’s gate, the signal is amplified, allowing the system to sense small amounts of charge quickly. According to Datta, the design could achieve read speeds as fast as one to two nanoseconds under ideal conditions. That is far quicker than DRAM, which typically takes 40 to 50 nanoseconds, though still slightly slower than standard SRAM.

The real advantage is density. Nagare’s first generation targets six stacked layers and roughly 185 megabits per square millimeter—about five times the density of SRAM. Datta says the architecture could eventually scale to 106 layers, potentially reaching 90 times SRAM’s density. The transistors are fabricated using plasma-enhanced atomic layer deposition at temperatures below 450 degrees Celsius, cool enough to build memory layers on top of a finished CMOS wafer. Nagare plans to produce control circuits on a mature logic process, stack the memory array above, and bond the finished chip to a processor as a chiplet.

Why This Matters

Memory innovation has stalled for decades. SRAM, DRAM, and flash still dominate the market, and high-profile alternatives such as Intel’s Optane have failed. Gelsinger argues that AI is forcing the industry to rethink the entire computing stack, with memory playing a central role.

Z2RAM is positioned as an addition to the memory hierarchy, not a direct DRAM replacement. That distinction matters because data-center operators and AI accelerator designers are increasingly open to custom memory solutions rather than rigid industry standards. The technology could be especially useful for edge AI and robotics, where vision models may need little capacity but require extremely low latency to interpret surroundings in real time.

Key Takeaways

  • Nagare Labs, a Georgia Tech spinout, launched Z2RAM with $5 million led by Playground.
  • Z2RAM uses two transistors per cell and no capacitor, enabling non-destructive reads.
  • Estimated read speeds are one to two nanoseconds, faster than DRAM but slightly slower than SRAM.
  • First-generation chips target six layers and 185 megabits per square millimeter, about five times SRAM density.
  • The design stacks vertically using low-temperature deposition and can be bonded to a processor as a chiplet.
  • Potential applications include AI accelerators, data centers, and low-latency edge robotics.

Conclusion

Significant hurdles remain. Write operations can disturb stored charge through capacitive coupling, so Nagare uses a patented metal shield inside each cell—a trade-off that widens the cell and reduces density, though stacking may offset that loss. Oxide semiconductor transistors also face threshold-voltage stability issues under heat and electrical stress, and the company is testing various material recipes involving indium, gallium, zinc, and tungsten.

There is no established standard for connecting Z2RAM to an AI processor, though the industry’s shift toward custom high-bandwidth memory designs may create an opening. Nagare still needs a fabrication partner and must demonstrate the technology on standard 300-millimeter wafers. Commercial chips are likely years away, and Datta notes the company will require substantial additional funding and time to move from university research to the market.