The hardware behind the future


NEO Semiconductor Stacks the Deck: 3D X-DRAM Clears Silicon POC on NAND Tooling

Cell-level test chips hit sub-10 ns latency and better than a second of retention at 85C, and an investor group led by Acer founder Stan Shih is funding the next phase.

NEO Semiconductor Stacks the Deck: 3D X-DRAM Clears Silicon POC on NAND Tooling

NEO Semiconductor says its 3D X-DRAM technology has passed silicon proof of concept on existing 3D NAND process equipment, and has taken a strategic investment led by Stan Shih, founder and former chairman and CEO of Acer and a TSMC board director for more than 20 years.

The performance figures are the easy part. NEO reports read/write latency below 10 ns, data retention beyond one second at 85C, bit-line and word-line disturbance beyond one second at the same temperature, and endurance above 10 to the 14th cycles. Blocks & Files, reporting the same results, puts conventional DRAM latency at roughly 10 to 20 ns and above, calls the disturbance numbers better than DRAM, and calls the endurance figure a match.

Retention is the number worth staring at. NEO measures its cells against the 64 ms JEDEC standard and claims 15 times better, which is not a benchmark score so much as a refresh budget. A DRAM array that holds charge for a second instead of 64 milliseconds spends less of its power and bus time refreshing itself, and refresh overhead is precisely what scales badly as density rises.

“These results validate a new scaling path for DRAM,” said Andy Hsu, founder and CEO of NEO Semiconductor. The scaling path, not the speed, is the pitch. NEO's argument is that stacking DRAM cells in a NAND-like structure lets the industry build density on tooling that is already installed, already amortised and already running past 300 layers, rather than on the capacitor scaling that conventional DRAM has been fighting for a decade.

Jeongdong Choe, senior technical fellow and SVP at TechInsights, is quoted in the announcement saying “NEO’s silicon POC represents a significant milestone, demonstrating real-world viability beyond theoretical concepts.” Worth noting where that appears: it is a named analyst at a named firm, but the quote arrives inside the vendor's own press release rather than in independent published analysis, and TechInsights has not put a report number against it.

Read NEO's description of what the money funds and the limits of the POC come into focus. The next phase covers array-level implementation and multi-layer test chip development, which is a plain admission that what has been demonstrated so far is not a multi-layer array. The manufacturability case rests on stacking hundreds of layers; the silicon so far does not.

Other gaps are conspicuous. NEO does not disclose the size of the investment, the valuation, or who else is in the Shih-led group. It names no memory maker as a co-development partner, which matters because a fabless company pursuing what it calls a scalable licensing and partnership model needs Samsung, SK hynix, Micron or Kioxia to sign something before any of this reaches a wafer in volume.

The POC was fabricated and tested in Taiwan, developed with National Yang Ming Chiao Tung University's Industry-Academia Innovation School and built at the Taiwan Semiconductor Research Institute. That is a credible academic route to first silicon and a long way from a qualified production process.

Bootnote: the 64 ms figure NEO benchmarks against is the JEDEC retention specification that forces every DRAM row to be rewritten many times a second, whether or not anything is reading it. It is the reason DRAM burns power sitting idle, and it is why a memory cell that simply holds its charge longer is an economic claim rather than a speeds-and-feeds one.

Leave a Reply