Georgia Tech spinout Nagare Labs says its Z2RAM technology could combine SRAM-like speed with DRAM-like density by building memory directly above AI processors. Research by its founders has demonstrated improved data retention in experimental cells, but sampling of Z2RAM tiles is targeted within four years and production shipments, if they arrive, within the next decade.
The company projects that its first generation, using six stacked layers, would offer five times the density of high-density SRAM, operate 14 times faster than DRAM and use one-third as much read energy as SRAM, according to Tom’s Hardware. Those figures are targets, not measured results from a finished memory stack.
AI processors need substantial memory bandwidth, while stacking memory above computing circuits creates thermal challenges. SRAM offers speed but takes up more space; DRAM provides greater capacity, but its temperature limits constrain the heat that processors underneath can generate.
Nagare’s proposed design uses two transistors and a shield, without the capacitor found in conventional DRAM cells. One transistor writes charge to a storage node, while another reads it without draining the stored charge. The shield is intended to reduce interference from neighbouring cells.
The company plans to deposit and pattern memory layers directly on an accelerator, avoiding an interposer and bonds between memory layers. It says standard semiconductor manufacturing tools could pattern several layers in one exposure. Tom’s Hardware reported that the single-lithography claim has not yet been demonstrated in published work.
A boosted three-transistor experimental cell in the founders’ research held data for more than 300 seconds at 85 degrees Celsius, compared with 100 milliseconds for a conventional three-transistor design under the reported test conditions. It is not the proposed two-transistor product.
A separate simulation found that a four-tier, 256MB cache of two-transistor gain cells at 7nm could reach 5.7 times the bit density of 7nm SRAM in the same area. The simulation does not establish the performance of a manufactured Z2RAM stack. Neither the published cells nor Z2RAM have demonstrated retention above the roughly 95-degree ceiling Nagare cites for DRAM.
“Where we are is literally lab prototypes,” co-founder Suman Datta told The Register, as reported by Tom’s Hardware. Other hurdles include reliability, the lack of a standard for integrating this kind of memory with an accelerator, and whether the experimental design can translate successfully into the proposed product. Nagare might have to come up with the interface itself.
Nagare is reportedly looking for a research fabrication partner to move beyond the university clean room, and its plan to scale depends on getting partners for fabrication and packaging. Playground Global led its $5 million funding round. Pat Gelsinger, the former Intel chief executive and a general partner at Playground, sits on Nagare’s board.
Analysis
Why This Matters
- Memory speed, capacity and heat constraints affect how effectively AI processors can operate. Nagare is targeting those constraints together, rather than simply increasing memory capacity.
- The distinction between experimental cells, simulations and commercial projections is central: promising retention measurements do not yet establish that a multilayer product can meet the company’s targets.
Background
SRAM and DRAM serve different roles in computing. SRAM is typically used where rapid access matters most, while DRAM provides larger working-memory capacity. Putting memory closer to processing circuits can shorten connections, but it also makes manufacturing and heat management more demanding.
What to Watch
- Whether Nagare secures fabrication and packaging partners capable of producing a multilayer stack beyond university prototypes.
- Measurements of a manufactured Z2RAM stack, particularly latency, energy use, reliability and retention at higher temperatures.
- Progress towards the founders’ target of sampling memory tiles within four years, including development of an accelerator interface.
Checked
12 claims in this story checked against 1 source: 11 supported.
Show each claim
- SupportedNagare Labs is targeting memory with SRAM-like speed and DRAM-like density, stacked above AI processors.Latest from Tom's Hardware: “Nagare Labs, says its Z2RAM can put SRAM-speed memory , at DRAM density, directly on top of AI chips.”
- SupportedThe performance figures are projections, and fabrication partners and further testing are still needed.Latest from Tom's Hardware: “until a fab-built, multi-layer Z2RAM stack is measured, retention above 95°C included, its headline numbers remain projections.”
- SupportedNagare is a Georgia Tech spinout proposing to build memory directly on AI processors.Latest from Tom's Hardware: “A Pat Gelsinger-backed Georgia Tech spinout, Nagare Labs, says its Z2RAM can put SRAM-speed memory , at DRAM density, directly on top of AI chips.”
- SupportedNagare targets Z2RAM tile sampling within four years and production shipments, if they happen, within the next decade.Latest from Tom's Hardware: “sampling of Z2RAM tiles is targeted within four years and production shipments, if they arrive, within the next decade.”
- Partly supportedThe six-layer first generation is projected to have five times the density of high-density SRAM.Latest from Tom's Hardware: “Nagare Gen 1 density 5× SRAM Projected Six layers, vs. 2nm HD SRAM”
- SupportedThe first generation is projected to operate 14 times faster than DRAM.Latest from Tom's Hardware: “Nagare Gen 1 speed 14× faster than DRAM Projected Gen 1”
- SupportedThe first generation is projected to use one-third as much read energy as SRAM.Latest from Tom's Hardware: “Nagare Gen 1 read energy 3× less than SRAM Projected Gen 1”
- SupportedA boosted three-transistor experimental cell retained data for more than 300 seconds at 85°C, versus 100 milliseconds for a conventional three-transistor design.Latest from Tom's Hardware: “Boosted three-transistor cell, retention More than 300 s Measured 85°C, to a 100 mV decay”
- SupportedA simulation found that a four-tier, 256MB cache of two-transistor gain cells at 7nm could reach 5.7 times the bit density of 7nm SRAM in the same area.Latest from Tom's Hardware: “Bit density 5.7× that of 7nm SRAM Simulated Four tiers, 256MB, same area”
- SupportedNeither the published cells nor Z2RAM has demonstrated retention above the roughly 95°C ceiling Nagare cites for DRAM.Latest from Tom's Hardware: “neither the founders’ published cells nor Z2RAM have shown retention above that line yet.”
- SupportedSuman Datta said, “Where we are is literally lab prototypes,” to The Register.Latest from Tom's Hardware: ““Where we are is literally lab prototypes,” Datta told The Register”
- SupportedThe single-lithography claim has not yet been demonstrated in published work.Latest from Tom's Hardware: “That has not yet been proven, and nothing published yet shows it.”