Plate Nº 74 · recorded October 10, 2026
PhysicsReported finding
Stanford-built quantum RAM picks any of 7 cells at 1.2% error
Stanford team builds a seven-cell quantum memory prototype that lets one processor pull data from any cell on demand, with each access adding roughly 1.2% error on the active cell.
By Marcus Bennett4 min read756 words
In brief
- The prototype carries 7 memory cells; each access introduces about 1.2% average error on the active cell
- Individual swaps run at roughly 99.3% fidelity and take about half a microsecond
- Published in Nature Physics on September 25, 2026 (DOI: 10.1038/s41567-026-03418-w); senior author David I. Schuster of Stanford, co-first authors Ziqian Li and Eesh Gupta
- Device combines four parts: a transmon, a 7-mode aluminum cavity, a buffer cavity, and a tunable coupler
- In classical chips more than 99% of transistors serve memory; current superconducting processors have no equivalent
A Stanford-led team has built a prototype quantum memory that lets one processor pull data from any of seven storage cells on demand, with each access introducing about 1.2% error on average. The device, described in a paper published in Nature Physics on September 25, 2026, points toward giving superconducting quantum computers a separate memory layer rather than storing and processing data in the same hardware.
Why does quantum computing need RAM at all?
Classical computers split tasks between a central processing unit (CPU) and random-access memory, the short-term storage that lets the CPU grab any piece of data on demand. Most superconducting quantum computers today do not make that split. They use the same hardware for storing and processing information.
Each new qubit needs its own microwave control electronics and wiring that descends into a dilution refrigerator, the ultra-cold chamber that keeps superconducting circuits near absolute zero.
"If every additional qubit costs another set of control hardware, scaling stops being a physics problem and becomes an engineering one," David I. Schuster, senior author of the paper, said.
What does the new device look like?
The team's prototype has four parts: a transmon (a widely used superconducting qubit that does the computing), a multimode aluminum cavity that hosts seven distinct microwave modes (each acting as an individually addressable memory cell), a buffer cavity between them, and a tunable coupler that manages the swap. Modes are distinct oscillation patterns in which microwave energy can sit inside the cavity, and each one can hold one qubit's worth of information.
"The transmon does the computing, the multimode aluminum cavity has seven modes that act as individually addressable memory cells, and the buffer cavity between them works rather like a cache, or a workbench," Schuster said.
Selecting a cell works by sending a control tone at the right microwave frequency. The tone triggers an exchange that swaps the chosen cell's quantum state into the buffer, where the transmon operates on it. The process runs in reverse to write the state back. Each swap takes about half a microsecond. The other six cells stay exactly where they were.
How well did it perform?
The team repeatedly accessed the seven cells thousands of times under randomized operations and tracked how errors accumulated. Each access operation introduced about 1.2% error per memory mode. Single swaps ran at roughly 99.3% fidelity.
The researchers identified the main culprit: weak unwanted interactions between the seven modes. Operating on one cell slightly nudges the others out of step, even though they sit idle.
"We tracked not only the cell being used but every cell sitting idle in storage, which let us see exactly where the fidelity was going," said Ziqian Li, co-first author of the paper. "The dominant loss is not decay, and it is not imprecise control — it is that the modes interact weakly with one another."
That diagnosis gives the team a concrete target: better couplers that keep the modes more isolated from one another, especially as more cells are added.
Why put a buffer in the middle?
The buffer cavity is the defining architectural choice. In earlier designs, a transmon connected directly to memory modes created unwanted couplings. The transmon's strong interactions are what make it useful for computing, but those same interactions disturb the fragile quantum states meant for storage.
"Fast processing introduces parasitic interactions which can corrupt the coupled memories," said Eesh Gupta, co-first author of the paper. "In our cascaded design, the buffer effectively shields them and thus brings the best out of the processor and the memories."
What comes next?
The team plans to combine the device with quantum error correction, the family of techniques that detect and fix mistakes on logical qubits built from many physical qubits.
"The reason to build memory at all is error correction," Schuster said. "We are now exploring how this type of memory can be co-designed with quantum error correction algorithms to improve both memory and quantum processing."
The team says a working memory layer would let future machines reach the same logical computing power with many fewer control wires than today's designs need. In classical hardware, more than 99% of transistors serve memory rather than logic. Current superconducting processors have no analogue.
The paper appeared in Nature Physics. It credits Ziqian Li, Eesh Gupta, and colleagues at Stanford University, the University of Chicago, and the SLAC National Accelerator Laboratory. The DOI is 10.1038/s41567-026-03418-w.
via Phys.org Physics (Source)