ULTRARAM: The Memory That Refuses to Forget

A single cell sits on a sliver of silicon no wider than a human hair. Inside it, electrons slip through quantum barriers thinner than a strand of DNA, locking data into place with a grip so firm that the information will outlast the person who stored it. One thousand years from now, if the chip survives, the bits will still be there. This is ULTRARAM. At a time when AI systems are pushing ever harder against the constraints of memory speed, energy, and durability, that kind of persistence is more than a curiosity. It may well be an inflection point.

Where Today’s Memories Fall Short

Every smartphone, laptop, and data center relies on a familiar trio of memory technologies. Flash stores photos and files without needing constant electrical current, but it demands voltages above 10 V to shuffle charges through a thick oxide barrier, and that brute force wears the device out after roughly 10,000 write cycles. DRAM keeps processors fed with data at lower voltages, yet it must refresh itself every 32 milliseconds or the information evaporates. SRAM responds the fastest of all, but a single bit requires six transistors, making it far too bulky and expensive for mass storage.

For decades, researchers have searched for something better. The goal has been a so-called “universal memory,” a single device that could replace today’s layered hierarchy by combining the speed of SRAM, the density of DRAM, and the persistence of flash, all while consuming minimal energy. Several contenders have entered the ring over the years, from ferroelectric RAM to spin transfer torque MRAM. Each brings strengths to the table, yet none has managed to check every box at once.

A Quantum Trick at the Nanoscale

ULTRARAM, developed by physicists and engineers at Lancaster University in the United Kingdom along with collaborators at UC Berkeley and IIT Roorkee, takes a fundamentally different approach. At its core, the device works by making electron movement highly selective rather than forceful. Rather than shoving electrons through a single thick oxide barrier the way flash does, ULTRARAM guides them through a triple barrier resonant tunneling (TBRT) structure built from indium arsenide (InAs) and aluminum antimonide (AlSb).

The physics behind this structure is elegant. Two quantum wells sit between three ultrathin AlSb barriers, each only a few atomic layers thick, forming the “triple barrier” stack that gives TBRT its name. When no voltage is applied, the energy levels inside those wells sit far above the electron states in the surrounding layers. The result is a 2.1 electron volt (eV) energy barrier, high enough to effectively eliminate leakage at room temperature and roughly twice what a conventional flash cell manages with its silicon dioxide layer. This is the locked vault that keeps data in place for over a millennium.

However, once a modest voltage of plus or minus 2.5 V is applied to the control gate, the conduction band tilts just enough for the quantum well energy levels to line up with the occupied electron states in the channel, creating a resonant path through the barrier. At that instant, resonant tunneling kicks in and electrons traverse the barrier rapidly. Think of it less as pushing a ball over a wall and more as briefly opening a perfectly aligned set of doors. When the alignment is right, electrons pass through effortlessly. When it is not, the barrier is effectively impenetrable. A negative voltage on the control gate produces the opposite effect, pulling electrons back out. The write and erase operations are therefore fast, low voltage, and consume switching energy per unit area roughly 1,000 times less than flash and 100 times less than DRAM.

From GaAs to Silicon

Early ULTRARAM cells were grown on gallium arsenide (GaAs) substrates, which are well suited to III-V compound semiconductors but expensive and fragile compared to silicon. Recognizing that any serious commercial memory technology must be compatible with existing silicon fabrication infrastructure, the researchers tackled the integration head on.

Using molecular beam epitaxy, the group deposited ULTRARAM layers on 3-inch silicon wafers with a 4 degree offcut in the [011] crystal direction. That slight tilt in the wafer surface helps suppress the formation of antiphase domains, a common defect when growing polar III-V materials on nonpolar silicon. A carefully engineered GaSb buffer layer bridged the 12.3% lattice mismatch between silicon and the memory stack. Transmission electron microscopy confirmed that the buffer was free of antiphase domains, with most threading dislocations confined to the first 500 nm of GaSb, well below the active memory layers. This matters because defect-free integration on silicon is a prerequisite for scaling beyond lab prototypes.

The devices fabricated on silicon showed retention exceeding 1,000 years at room temperature, endurance beyond 10 million program/erase cycles with zero degradation observed, and clear logic state separation during continuous cycling. Those numbers put ULTRARAM in a category of its own among charge based memories.

Building a Compact Model

Without an accurate compact model, ULTRARAM cannot be meaningfully evaluated in real chip designs. To bridge the gap between laboratory measurement and circuit simulation, the authors developed a physics based compact model for the TBRT current by solving the Schrödinger equation inside the quantum wells and expressing the tunneling transmission coefficient as a Lorentzian function.

The model accounts for both resonance levels in the double quantum well structure and adds a thermionic component to capture thermal excitation over the barriers. An RC network then integrates the tunneling current over the applied pulse duration, yielding the real time charge trapped in the floating gate. That charge determines the threshold voltage shift, which in turn sets the drain current in either the programmed or erased state.

Validation against technology computer aided design (TCAD) simulations and experimental measurements from silicon based ULTRARAM cells showed excellent agreement. The model also captures how the memory window varies with pulse width and rise/fall time, making it suitable for exploring design trade offs at the circuit level.

Stacking Up Against the Competition

A side by side comparison of ULTRARAM with other memory technologies reveals where this device excels.

In terms of cell area, SRAM is the bulkiest at over 100 F², while DRAM sits at 6 F² and both flash and ULTRARAM come in below 4 F². FeRAM occupies 15 to 35 F², and STT MRAM ranges from 6 to 50 F². On operating voltage, SRAM and DRAM both work below 1 V, but flash requires over 10 V. FeRAM stays below 5 V, STT MRAM below 1.5 V, and ULTRARAM below 2.5 V.

The switching energy gap is where ULTRARAM stands apart most dramatically. SRAM and DRAM both consume on the order of 10⁻¹⁵ joules per switching event. Flash is slightly higher at 10⁻¹⁴ J. FeRAM jumps to 10⁻¹¹ J and STT MRAM to 10⁻¹³ J. ULTRARAM, by contrast, operates at roughly 10⁻¹⁷ J per unit area, which is three orders of magnitude below flash and two below DRAM.

Retention tells an equally striking story. SRAM holds data for about 40 ms and DRAM for about 60 ms before requiring a refresh. Flash, FeRAM, and STT MRAM all retain data for more than 10 years. ULTRARAM exceeds 1,000 years. On endurance, SRAM and DRAM lead with 10¹⁶ cycles, while flash trails at 10⁴. FeRAM exceeds 10¹⁰, STT MRAM reaches 10¹⁴, and ULTRARAM has shown zero degradation after 10⁷ cycles with the true ceiling still unknown. Switching time ranges from 1 ns for SRAM to 10⁵ ns for flash, with ULTRARAM at approximately 100 ns in TCAD simulations for 20 nm channel length devices.

Among the newer contenders, ferroelectric RAM (FeRAM) and spin transfer torque MRAM (STT MRAM) both offer nonvolatility with respectable endurance. Still, neither yet matches ULTRARAM’s combination of sub 4 F squared cell area, millennial retention, and ultra low switching energy.

Why ULTRARAM Could Reshape AI Hardware

Beyond conventional data storage, ULTRARAM shows strong potential as a synaptic element in neuromorphic computing architectures. What makes it especially interesting is not just that it stores data well, but that it updates data linearly and predictably, which is exactly what analog learning systems require.

In biological brains, synapses adjust their connection strength in response to activity, a process that artificial neural networks approximate by updating weight values stored in memory. When those weights live in off chip DRAM, the constant shuttling of data between processor and memory burns energy and slows training. An on chip analog memory that can store and update weights locally would cut that overhead dramatically.

The authors explored ULTRARAM as an analog synapse by applying sequences of programming pulses with varying amplitude but fixed width. This amplitude modulation scheme produced linear weight updates during both long term potentiation (LTP) and long term depression (LTD), with 32 accessible conductance states. Linearity in weight updates matters greatly because asymmetric or abrupt conductance changes can degrade training accuracy in deep neural networks. That linearity, combined with ULTRARAM’s low energy per switching event, also makes the device relevant for edge AI applications where every milliwatt counts.

As with other analog synaptic devices, however, variability, noise, and device-to-device mismatch will ultimately determine how well these idealized updates translate into large scale hardware.

On Chip Training Results

To quantify on chip performance, the researchers used the DNN+NeuroSim framework to simulate a VGG 8 neural network trained on the CIFAR 10 image classification dataset. Each synapse in the 128 by 128 crossbar array was modeled using the ULTRARAM compact model at a 32 nm node.

The ULTRARAM based accelerator achieved 91.28% training accuracy, virtually identical to the 91% reached by an SRAM based design, effectively reaching parity with conventional digital training. Meanwhile, it consumed 1.36 times less chip area, finished each training epoch 1.87 times faster, and used 1.1 times less dynamic energy. Throughput reached 1.46 tera operations per second (TOPS), and the energy performance ratio came in at 2.12 TOPS per watt. Those gains stem not from raw compute improvements but from reduced data movement, which reinforces the broader thesis that memory, not logic, is the bottleneck in modern AI training.

Benchmarking against other analog synaptic devices, including memristors, resistive RAM (RRAM), epitaxial RAM (EpiRAM), and ferroelectric FETs (FeFET), further reinforced ULTRARAM’s competitive standing. Memristors and RRAM achieved lower training accuracy of 49% and 37%, respectively, while requiring much longer write pulses. EpiRAM performed better at 85% accuracy but still fell short. FeFET came closest, reaching 91.12% accuracy with slightly better latency and energy numbers, though ULTRARAM matched it within a narrow margin across all metrics.

What Could Go Wrong

No emerging memory technology arrives without risk, and ULTRARAM is no exception. Several open questions remain before the device can move from isolated cells to commercial products.

First, III-V compound semiconductor fabrication is inherently more expensive than standard silicon CMOS processing. Even with successful integration on silicon substrates, the epitaxial growth steps add cost and complexity that foundries would need to absorb. Second, yield at scale is unproven. The researchers have shown impressive results on individual cells and small test structures, but large crossbar arrays will expose any variability in the TBRT stack, where even a 0.6 nm monolayer variation can shift device characteristics. Third, ULTRARAM enters a crowded field. FeFET and STT MRAM continue to advance rapidly, and both benefit from more mature manufacturing ecosystems. Whether ULTRARAM’s physics based advantages translate into a compelling cost per bit at production volumes remains to be seen.

Taken together, these results suggest that ULTRARAM is not just another memory candidate but a serious attempt to collapse the boundary between storage and computation.

What Comes Next

The work presented so far is based on single cell measurements and model driven simulations. Array level fabrication is currently underway, and the researchers have pointed to a scalable array architecture with separate read and write paths as the template for future integration. Hardware validation of neuromorphic training on physical ULTRARAM arrays remains a key milestone ahead.

Still, the combination of properties assembled in one device is noteworthy. Retention measured in millennia, endurance that shows no wear after millions of cycles, switching voltages below 2.5 V, energy consumption several orders of magnitude below flash, and compatibility with silicon substrates make ULTRARAM a compelling candidate for both storage and compute in memory applications.

Whether ULTRARAM ultimately becomes the universal memory that researchers have sought for decades will depend on yields, costs, and manufacturing scalability at advanced nodes. The remaining questions are no longer about whether the device works. They are about whether it can be manufactured at scale.


Citation: A. Kumar, P. D. Hodgson, J. Ashforth-Pook, M. Hayne, and A. Dasgupta, “ULTRARAM: An Emerging Memory Technology for NVM and Low-Power Neuromorphic Systems,” 2026 27th International Symposium on Quality Electronic Design (ISQED), 2026, doi: 10.1109/ISQED69900.2026.11534662.

Related Articles

PHP Code Snippets Powered By : XYZScripts.com
★

Track This Company

Create a free account to add companies to your watchlist and get notified when patents, funding, or key developments change.