Memory & StorageLOT 261122 · 3 min read
Nagare Labs Exits Stealth With $5 Million for Z2RAM Stacked Memory
Georgia Tech spinout Nagare Labs has $5 million from Playground to build Z2RAM, a capacitor-free stacked memory targeting 1-2 ns reads and five times SRAM density on 300 mm wafers.
Key takeaways
- Nagare Labs emerged from stealth on 26 October with US $5 million led by Playground.
- First-generation Z2RAM targets six layers and 185 Mbit/mm², about five times SRAM density.
- Estimated best-case read speed is 1-2 nanoseconds, versus 40-50 ns for DRAM.
- Founders project up to 106 layers, which would give 90 times SRAM density.
- Next milestones: sign a fab partner and demonstrate the memory on 300 mm wafers.

Nagare Labs emerged from stealth on 26 October with US $5 million to develop Z2RAM, a stacked, capacitor-free memory that its founders claim will reach 185 megabits per square millimeter in its first generation. That figure, if confirmed on silicon, would be roughly five times the density of SRAM at read latencies of one to two nanoseconds.
The funding round was led by Playground, the venture firm where former Intel CEO Pat Gelsinger now serves as general partner. Gelsinger has personal reason for caution. "We haven't had a new memory for three decades," he said, noting that SRAM, DRAM and flash still dominate while alternatives such as Intel's Optane have failed.
The company is a spinout from Georgia Tech, cofounded by professor Suman Datta, a former Intel transistor researcher, and colleague Shimeng Yu.
What does Z2RAM actually change?
The name encodes the design. The first Z stands for stacking in the vertical axis; the second stands for zero capacitor. A conventional DRAM cell stores a bit as charge on a capacitor, and every read drains it, forcing a write-back and stretching access to 40 to 50 nanoseconds.
Z2RAM uses two transistors per cell instead. One writes the bit; the other reads it without disturbing it, so the two operations no longer share a port. Placing the charge on the read transistor's gate also amplifies the signal, which means a smaller stored charge can still be sensed quickly.
Datta frames the positioning narrowly. "This is an SRAM addition, rather than a DRAM replacement," he said. The technology is not quite as fast as SRAM, but SRAM carries a structural cost: it is built in the same front-end process as the processor, so every bit competes for area on an expensive logic die. SRAM cells have also proven increasingly difficult to shrink at advanced nodes.
How dense can the stack go?
Nagare's answer is vertical. The first generation targets six layers, for the 185 Mbit/mm² figure. Eventually Datta projects up to 106 layers, which would yield a density 90 times that of SRAM — a projection, not a demonstrated result.
Stacking is feasible, Datta says, because Z2RAM's transistors can be built with plasma-enhanced atomic layer deposition below 450 degrees Celsius, cool enough to add memory layers on top of a finished CMOS wafer without degrading it. The plan is to tape out control circuits on a mature logic process, build the array above with multiple layers per lithography step, then bond the finished chip onto a processor as a chiplet.
What stands in the way?
The hurdles are real. When the write transistor switches off, capacitive coupling from its control line disturbs the stored charge, narrowing the read margin and risking errors. Nagare's patented answer is a metal shield inside each cell, but the shield widens the cell and cuts density; vertical stacking is expected to offset the loss.
Stability is the second obstacle. "Right now the biggest challenge for oxide semiconductor transistors is the threshold voltage stability under bias and temperature stress," Datta said. If the turn-on voltage drifts, a stored bit can be misread. The company is still testing channel recipes using oxides of indium and gallium, sometimes with zinc, and tungsten-doped indium oxide.
The third hurdle is attachment: no standard exists for bonding this class of memory to an AI chip. Datta sees an opening in high-bandwidth memory, where AI accelerator builders are moving away from JEDEC-standard designs toward custom solutions.
What comes next?
Gelsinger's milestones are concrete: leave the university lab, sign a fabrication partner, and demonstrate the device on standard 300-millimeter wafers. The $5 million is a first step. "This will be multiple tranches of real funding," he said. Datta agreed, saying the company "requires a lot of runway, both in terms of capital and time."
Beyond data-center accelerators, Datta sees an edge opportunity in robotics, where vision models may need little capacity but demand extremely low latency for real-time environmental understanding. Commercial chips remain years away, and Datta keeps expectations modest: "The application will emerge just based on the capability of your technology."
linkedin.com (Original)
More from Nathan Brooks
Show full bio
Senior reporter covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,201 articles