{"slug": "z2ram-wants-to-grow-sram-speed-memory-on-top-of-ai-chips-but-its-best-numbers", "title": "Z2RAM Wants to Grow SRAM-Speed Memory on Top of AI Chips, but Its Best Numbers Are Still Projections", "summary": "Nagare Labs, a Georgia Tech spinout with former Intel CEO Pat Gelsinger on its board, has raised $5 million in pre-seed funding led by Playground Global for its capacitorless Z2RAM gain-cell memory, which it says can be stacked directly on top of an AI accelerator with SRAM-like speed and DRAM-like density. The published research shows a boosted three-transistor cell retaining data for more than 300 seconds at 85°C, versus about 300 milliseconds for the early two-transistor cell, but Nagare's first-generation targets — five times the density of high-density SRAM, access 14 times faster than DRAM and a third of SRAM's read energy — are all marked as projections, and co-founder Suman Datta told The Register: \"Where we are is literally lab prototypes.", "body_md": "**Nagare Labs, a Georgia Tech spinout with Pat Gelsinger on its board, says its capacitorless Z2RAM can sit directly on top of an AI accelerator with SRAM-like speed and DRAM-like density. The lab data behind it is real, but the headline numbers are still projections.**\n\nEvery AI chip designer is fighting the same war, and it is not about compute. Peak FLOPS have been climbing far faster than the bandwidth needed to feed them, which is why HBM stacks keep getting taller, pricier and harder to buy. Nagare Labs thinks the answer is to stop shipping data across an interposer altogether and grow the memory straight on top of the logic.\n\nThe Atlanta startup came out of stealth late last month with [$5 million in pre-seed funding led by Playground Global](https://convergedigest.com/nagare-labs-5m-z2ram-ai-memory/), the venture firm where former Intel CEO Pat Gelsinger is a general partner. Gelsinger has taken a board seat. This week, [Tom’s Hardware went through the published research](https://www.tomshardware.com/tech-industry/semiconductors/gelsinger-backed-nagare-labs-capacitorless-z2ram-promises-sram-speed-with-dram-density-latest-published-cell-retains-data-for-over-300-seconds-at-85-c-but-further-figures-are-projected) behind the company’s pitch, and the picture is more interesting, and more cautious, than the slogan.\n\n## Two transistors and a shield\n\nZ2RAM is a gain cell. Instead of the capacitor that every DRAM bit relies on, it stores charge on an insulated node between a write transistor and a read transistor, with a shield in between to keep neighbouring cells from disturbing each other. Because the read transistor senses the charge without draining it, reads are non-destructive. A conventional DRAM cell has to write its data back after every read; SRAM avoids that, but burns six transistors per bit and has all but stopped shrinking at the leading edge.\n\nThe trick is in the materials. The transistors are built from oxide semiconductors, which can be deposited at low temperatures in the back end of the line, after the logic is already finished. That means the memory layers can be patterned directly above an accelerator on a single die, using standard 300 mm DUV tools rather than EUV, with no hybrid bonding between layers. [Nagare’s roadmap](https://www.nagarelabs.ai/) runs past 100 layers and talks about hundreds of thousands of vertical connections, compared with the few thousand lanes an HBM stack exposes.\n\nThe company was founded by two Georgia Tech professors, Suman Datta and Shimeng Yu. Datta co-invented Intel’s high-k/metal gate and Tri-Gate transistors; Yu is known for the NeuroSim modelling framework. This is their own research lineage, stretching from a 2020 IEDM paper on an oxide two-transistor gain cell to work presented at VLSI this year.\n\n## What has actually been measured\n\nRetention is the number that decides whether a capacitorless cell is usable, and it has improved quickly. The early two-transistor cell held its data for about 300 milliseconds at 85°C. A “boosted” three-transistor variant shown at VLSI 2026 kept it for more than 300 seconds at the same temperature, enough that the researchers describe it as effectively refresh-free. For comparison, standard DRAM is refreshed every 64 ms or so.\n\nEverything beyond that is a projection. Nagare’s first-generation target, a six-layer stack, promises five times the density of high-density SRAM, access 14 times faster than DRAM and a third of SRAM’s read energy. All three are marked as projected on the company’s own site, and Nagare declined to share measured latency or energy figures. Nothing published yet shows retention above roughly 95°C, and it is not yet proven that the tested three-transistor cells behave like the two-transistor product the company is selling.\n\nDatta was candid about where things stand, telling The Register: “Where we are is literally lab prototypes.” He also flagged single-electron effects that can nudge a cell’s threshold far enough to flip a bit. Nagare has no fab and is looking for an R&D partner, with sampling targeted within about four years.\n\nSo, not a product, and not close to one. But it is a credible attempt at the problem every GPU vendor is currently paying HBM prices to paper over, and with a former Intel CEO in the boardroom, the foundries will at least take the meeting.", "url": "https://wpnews.pro/news/z2ram-wants-to-grow-sram-speed-memory-on-top-of-ai-chips-but-its-best-numbers", "canonical_source": "https://hwbusters.com/news/z2ram-wants-to-grow-sram-speed-memory-on-top-of-ai-chips-but-its-best-numbers-are-still-projections/", "published_at": "2026-11-08 10:09:32+00:00", "updated_at": "2026-11-09 01:35:32.133734+00:00", "lang": "en", "topics": ["ai-chips", "ai-infrastructure", "ai-startups"], "entities": ["Nagare Labs", "Z2RAM", "Pat Gelsinger", "Playground Global", "Suman Datta", "Shimeng Yu", "Georgia Tech", "Tom's Hardware"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/z2ram-wants-to-grow-sram-speed-memory-on-top-of-ai-chips-but-its-best-numbers", "markdown": "https://wpnews.pro/news/z2ram-wants-to-grow-sram-speed-memory-on-top-of-ai-chips-but-its-best-numbers.md", "text": "https://wpnews.pro/news/z2ram-wants-to-grow-sram-speed-memory-on-top-of-ai-chips-but-its-best-numbers.txt", "jsonld": "https://wpnews.pro/news/z2ram-wants-to-grow-sram-speed-memory-on-top-of-ai-chips-but-its-best-numbers.jsonld"}}