Nagare Labs, a Georgia Tech spinout with former Intel chief executive Pat Gelsinger on its board, says it can build memory that runs at SRAM speed at DRAM density and sits directly on top of an AI accelerator. The company calls it Z2RAM, and its founders' published cells now hold data for more than 300 seconds at 85 degrees Celsius, a figure that matters because a GPU's last-level cache only needs to keep a line alive until it is overwritten.
The pitch lands in the middle of the industry's most expensive bottleneck. Compute has grown about 70 percent a year for three decades while memory bandwidth has grown about 20 percent, according to figures Nagare cites. High-bandwidth memory has been the workaround, and it is hard to yield at high layer counts. Stack DRAM on the logic and heat becomes the binding constraint: a 2025 simulation from co-founder Shimeng Yu's group put the ceiling on logic power at 220 watts per square centimetre with HBM stacked above it, or 300 with better thermal vias.
Two transistors and a shield
Nagare describes the cell as "two transistors and a shield. No capacitor." A write transistor puts charge on a storage node, a shield blocks interference from neighbours, and a read transistor senses the charge without draining it. Because the read gate's capacitance holds the value, reads are non-destructive and the cell skips the write-back that DRAM needs after every read.
The manufacturing idea is the part that would be hard to copy. Nagare deposits its layers on top of the accelerator and patterns them together on one die, the way 3D NAND adds layers, with no interposer and no bonds between memory layers. One lithography exposure covers several layers at once, and standard 300 mm DUV tools do the work, so the process avoids EUV entirely at a moment when ASML's EUV business is the scarcest link in the chain. The roadmap aims past 100 layers with hundreds of thousands of vertical connections.
Projections for the first generation: five times SRAM density across six stacked layers, 14 times DRAM speed, and a third less read energy than SRAM. The boosted cell measures 0.0168 square micrometres at 7nm, against 0.0199 for a high-density SRAM cell on TSMC's 3nm process.
Retention is the whole argument
The founders have been publishing toward this for six years, and the numbers trace the progress. A 2020 paper on a tungsten-doped indium oxide cell from co-founder Suman Datta's group held data for one second at 25 degrees and 300 milliseconds at 85. A 2025 update with three transistors held over 1,000 seconds at room temperature and simulated a GPU L2 cache at twice SRAM's capacity in the same area, with a 75 to 80 percent lower energy-delay-area score.
The folded asymmetric boosted cell, shown six months later, reached over 300 seconds at 85 degrees where a conventional three-transistor design managed 100 milliseconds. It is also refresh-free, which is what makes it usable as cache: data does not decay before the workload overwrites it.
That still sits below the roughly 95 degrees Nagare says DRAM must stay under. Nothing published shows retention above that line, and the company declined to share measured retention, latency or energy figures with Tom's Hardware. A separate simulation from Yu's group found a four-tier, 256 MB cache of two-transistor cells at 7nm could reach 5.7 times the bit density of SRAM in the same area, with returns falling off past seven or eight tiers because masks scale with layers.
No fab, no standard, no samples
Datta told The Register that the work is "literally lab prototypes," with sampling of Z2RAM tiles targeted within four years and production, if it arrives, within the decade. There is no standard yet for integrating this kind of memory with an accelerator, which means Nagare may have to define the interface itself. Single-electron effects can shift a cell's threshold enough to flip a bit, and nothing has shown that the tested three-transistor cells behave the same as the two-transistor product.
The company also has no fab. Scaling depends on partners for fabrication and packaging, and Nagare is looking for an R&D fab partner to move beyond a university clean room. Datta cited AMD's 3D V-Cache as the model for working with chipmakers, which is the closest thing shipping today to memory stacked on compute.
It is worth remembering that this pitch has been made before. Innovative Silicon's Z-RAM promised a similar multiple over SRAM 20 years ago, AMD licensed it and did not pursue it. Gelsinger, meanwhile, once phased out Intel's Optane in favour of CXL, and now sits on the board of a startup attacking the same gap from the other direction. The same pressure is pushing memory makers to reallocate capacity toward data centres, which is why a credible answer to the bandwidth wall attracts capital long before it attracts a customer.