A research team from the University of Illinois Urbana-Champaign and the University of Notre Dame has reported measuring a 2.75 GHz modulation bandwidth in a blue micro-LED with a 5µm diameter, in a paper accepted by IEEE Electron Device Letters. Modulation bandwidth indicates how quickly the intensity of light can be varied in response to an electrical signal. The device reached this value at a current density of 15.1 kA/cm² and an electrical input of 25 mW. The university describes the result as a world record for quantum-well LEDs, but if the scope is not limited that way, a different research group reported 3.73 GHz in 2024. What matters most in this study is that the team fabricated five device sizes, from 100µm to 2µm in diameter, from epiwafers with the same layer structure and examined systematically how performance varies with size. The fastest was the 5µm device, and bandwidth fell in the even smaller 2µm device. Through measurements and calculations, the team clarified how miniaturization speeds devices up until a point, after which other factors begin to limit performance.
2.75 GHz at 5µm diameter, measured as small-signal modulation bandwidth
The paper has seven co-authors, led by graduate student Abdelghany Abouelnagga, and an author version before final editing is available. It was submitted on July 10, 2026, and its DOI (10.1109/LED.2026.3745737) was registered on October 21. The research groups of Professor Can Bayram, Associate Professor Peter Dragic, and Professor Emeritus Milton Feng of the University of Illinois, and Professor Patrick Fay of the University of Notre Dame took part, and the university has also introduced the results in a press release.
The team used a blue-emitting epiwafer grown on a 6-inch Qromis Substrate Technology substrate. The light-emitting layer is a six-layer multiple quantum well structure of InGaN (indium gallium nitride) and GaN (gallium nitride), and circular emitters were fabricated in five diameters: 100, 20, 10, 5, and 2µm. The highest modulation bandwidth was recorded by the 5µm device, which reached 2.75 GHz at a current density of 15.1 kA/cm². The paper's abstract describes the result as "over 2.7 GHz."
What was measured is the frequency response of optical output to a small signal. The team drove the device by superimposing a weak −20 dBm high-frequency signal on a DC bias current, collected the emitted light with a multimode fiber, and measured it with a photodetector responsive up to 25 GHz. They examined the optical output response while raising the signal frequency, and defined the modulation bandwidth as the frequency at which the response drops 3 dB relative to the low-frequency response. The paper reports no bit rate (Gbps), bit error rate, or eye diagram. The 2.75 GHz figure therefore indicates the device's response speed and is not a number showing how many Gbps of data could actually be transmitted.
The 25 mW figure is likewise the electrical input supplied to the device at the operating point where 2.75 GHz was achieved, not the optical output. The body of the paper gives no specific optical output value. Calculating the circular junction area from the nominal 5µm diameter gives about 1.96×10⁻⁷ cm², so a current density of 15.1 kA/cm² corresponds to roughly 3.0 mA. However, the paper notes that sidewall damage from etching may make the effective junction area smaller than the nominal value, so this conversion is only a rough guide.
Behind the need for such high current density is the internal electric field in the light-emitting layer of InGaN/GaN LEDs. According to the paper, in InGaN/GaN LEDs grown on the crystal plane known as the c-plane, crystal polarization creates an electric field within the emitting layer that spatially separates electrons and holes. As a result, it takes longer for them to recombine and emit light, and modulation bandwidth has traditionally been limited to around 1 GHz or lower. Raising the current density injects large numbers of electrons and holes, and these carriers screen the internal field, speeding up recombination and widening the modulation bandwidth. In other words, high-speed operation requires devices that operate stably at high current density.
Tolerance for high current density varied with device size. In the 5µm device, the relationship between current and voltage stayed linear up to 25 kA/cm², whereas in the 20µm device self-heating changed the slope of the characteristic, and in the 2µm device degradation began at around 10 kA/cm². The team therefore measured modulation bandwidth at current densities before the effects of heating and degradation became significant. The current densities at measurement were 6.4 kA/cm² for 20µm, 9.5 kA/cm² for 2µm, and about 15 kA/cm² for 5µm. The bandwidths for each size are not compared at the same current density; they are values obtained within the range where each device operates stably.
The "world record" is limited to quantum-well LEDs; comparisons with earlier work need care
It is the university's press release, not the body of the paper, that positions the 2.75 GHz as a "world record." The university describes it as a "world record for modulation speed in quantum-well LEDs," limiting the claim to LEDs with the quantum-well structure that is standard in industry. It also explains that some non-standard LED structures operate at even higher frequencies but emit weak, unstable light. It would therefore not be appropriate to take this record as the top speed for micro-LEDs overall.
Table I of the paper compares the result with six earlier room-temperature results for standard multiple-quantum-well micro-LEDs. However, we have not been able to individually confirm the light-emitting-layer structure of all six cited works. The highest figure in the comparison table, 2.64 GHz by Li et al. in 2024, comes from a paper whose title features a "Nano-Engineered InGaN Active Region," which is not necessarily the same as a typical quantum-well structure. The table below lists the six works from the paper's Table I, plus the 2024 paper by Rogers et al., which is not in that table, and results announced by Taiwan's Ennostar in August 2026.
| Report | Diameter (µm) | Color | Bandwidth (GHz) | Definition | Current density | Type |
|---|---|---|---|---|---|---|
| This paper (2026) | 5 | Blue | 2.75 | E-E | 15.1 kA/cm² | Paper |
| Li 2024 | 20 | Blue | 2.64 | E-E | Not stated | Paper (value cited in Table I) |
| Rao 2024 | 20 | Blue | 2.30 | E-E | Not stated | Paper (value cited in Table I) |
| Xu 2025 | 20 | Blue | 1.60 | E-O | Not stated | Paper (value cited in Table I) |
| Jin 2025 | 10 | Green | 1.27 | E-E | Not stated | Paper (value cited in Table I) |
| Pezeshki 2025 | 7 | Blue | 2.00 | E-E | Not stated | Paper (value cited in Table I) |
| Lin 2026 | 20 | Yellow | 2.14 | E-E | Not stated | Paper (value cited in Table I) |
| Rogers 2024 | Not stated | Blue | 3.73 | E-O | 11 kA/cm² | Paper (IEEE Photonics Technology Letters) |
| Ennostar 2026 | Not disclosed | Blue | 3.3 | Not disclosed | 5 kA/cm² | Company announcement |
The 2.75 GHz exceeds 2.64 GHz, the highest of the six entries in the paper's comparison table. On the other hand, in 2024 a different research group reported 3.73 GHz in a blue GaN micro-LED, measured at room temperature and a current density of 11 kA/cm², with the bandwidth defined as E-O. Rogers et al. described this 3.73 GHz as the highest value reported for micro-LEDs at the time. However, because the definition of bandwidth differs between this paper and the work of Rogers et al., the numbers alone cannot be used to judge directly which performs better. Whether the devices of Rogers et al. qualify as quantum-well structures, and why they are not included in this paper's comparison table, are also unclear from the paper.
The "Definition" column in the table shows which relationship between input and output signals was used to obtain the bandwidth. E-O stands for "electro-optical," the response of optical output to an electrical signal. For E-E, however, no clear definition is given in this paper, so the table uses the original notation as is. The paper itself lists its own result as E-E in Table I while explaining in the text that it measured the response from electrical signal to optical output, so the descriptions are inconsistent. Bandwidths measured under different definitions cannot be compared simply, and because current densities are not given for the six comparison works, it also cannot be said that the paper exceeded previous records under identical conditions.
Companies, too, have announced figures that exceed this result. In August 2026, Taiwan's Ennostar announced that a single channel of its blue micro-LED reached a −3 dB modulation bandwidth of 2.1 GHz at a current density of 2 kA/cm² and up to 3.3 GHz at 5 kA/cm². However, this figure is based on a company announcement rather than a peer-reviewed paper, and neither the device size nor the definition of modulation bandwidth has been disclosed.
The University of Illinois research group itself also has earlier work reporting higher modulation speeds. In 2009, Milton Feng, Nick Holonyak Jr., and colleagues reported in the June 15 issue of Applied Physics Letters that they had achieved 4.3 GHz with a light-emitting transistor and 7 GHz with a tilted-charge LED based on it. At the time, the LED modulation speed record was said to be 1.7 GHz. Tilted-charge LEDs differ in structure from ordinary quantum-well LEDs and so are not included in this comparison, but Feng is also a co-author of the present paper. Taking all this into account, what the 2.75 GHz can be said with certainty to have exceeded are the six room-temperature measurements in the paper's comparison table; it is not the highest value for micro-LEDs overall.
Miniaturization speeds devices up to 5µm; why performance drops at 2µm
In this study, modulation bandwidth grew as device diameter was reduced from 100µm to 5µm, reaching a maximum of 2.75 GHz at 5µm. In the even smaller 2µm device, however, bandwidth fell. The maximum bandwidth for each size was measured at current densities before heating and degradation became significant: 32 A/cm² for the 100µm device and about 9.5 kA/cm² for the 2µm device. The only measured modulation bandwidth given as a specific number in the body of the paper is 2.75 GHz for 5µm. In the university's press release, lead author Abouelnagga explains that after systematically shrinking devices from 100µm to 2µm, the results showed that smaller is not necessarily better, and that 5µm showed the highest performance.
The speed-up from shrinking and the opposing cause of performance loss arise from different electrical properties of the device. The pn junction of an LED stores charge like a capacitor, and the larger the junction capacitance, the more charging and discharging in response to high-frequency signals lag, limiting response speed. This limit is set by the RC time constant, the product of resistance and capacitance, and making a device smaller reduces junction capacitance in proportion to area. Meanwhile, the contact resistance at the junction between metal electrode and semiconductor, and the current-spreading resistance as current spreads from the electrode to the emitting layer, increase as device area shrinks. When the effect of increased resistance outweighs the benefit of reduced capacitance, modulation bandwidth no longer widens even if the device is made smaller.
Another important factor is impedance mismatch, which causes reflection of high-frequency signals. The measurement system in this study is designed around 50Ω, and when the LED-side impedance departs greatly from this value, part of the supplied high-frequency signal does not reach the device and is reflected. Using Smith charts, which relate reflection characteristics to impedance, the team analyzed each device's characteristics over a frequency range of 100 MHz to 5 GHz. They found that in the 100µm device, large junction capacitance limited bandwidth, whereas below 20µm in diameter the influence of capacitance dropped substantially while increased resistance and impedance mismatch with 50Ω became prominent. In other words, as miniaturization proceeds, the main factor limiting bandwidth shifts from junction capacitance to resistance and impedance mismatch.
The team built an equivalent circuit for each device size and extracted resistance and capacitance from the measurements. Using those values, they calculated the "RC-pole calculated bandwidth (fRC)," corresponding to the frequency at which the response begins to fall because of RC characteristics. The following figures are from Table II of the paper and are not actually measured modulation bandwidths. The active-layer capacitance in the table represents the LED's junction capacitance in the equivalent circuit. The differential resistance of the device used in the calculation was evaluated at a current density of 10 kA/cm² for diameters of 20µm and below and 500 A/cm² for 100µm.
| Diameter (µm) | Active-layer capacitance (pF) | Contact resistance (Ω) | Current-spreading resistance (Ω) | RC-pole calculated bandwidth (GHz) |
|---|---|---|---|---|
| 100 | 45.10 | 6.10 | 2.50 | 0.29 |
| 20 | 2.862 | 30.70 | 50.71 | 3.56 |
| 10 | 0.639 | 61.50 | 202.83 | 4.61 |
| 5 | 0.133 | 123.15 | 811.32 | 5.57 |
| 2 | 0.031 | 307.20 | 5071.08 | 4.09 |
In the 100µm device, the active-layer capacitance was large at 45.10 pF, and the RC-pole calculated bandwidth was only 0.29 GHz. At 5µm, by contrast, capacitance fell to 0.133 pF and the calculated bandwidth expanded to 5.57 GHz. In the even smaller 2µm device, however, the calculated bandwidth fell to 4.09 GHz even though capacitance dropped further to 0.031 pF. This is because, in addition to the increase in resistance, the parasitic capacitance of the pad, which exists regardless of device size, becomes relatively more significant. Current-spreading resistance in particular increases about 2,000-fold, from 2.50Ω at 100µm to 5071.08Ω at 2µm. The team explains that the combination of increased contact and current-spreading resistance, the 0.18 pF parasitic capacitance of the p-side pad, and reduced tolerance for high current density meant that the 2µm device could not fully exploit the capacitance-reduction benefit of miniaturization.
There are several caveats in interpreting these calculations. First, from 20µm to 5µm the active-layer capacitance falls by more than would be expected from the area ratio, which the paper attributes to sidewall damage from dry etching reducing the effective junction area. Also, whereas the RC-pole calculated bandwidth of the 5µm device was 5.57 GHz, the measured modulation bandwidth was only 2.75 GHz. According to the paper, in devices of 5µm and below, the bandwidth set by the electron-hole recombination rate and the bandwidth set by RC characteristics come close to each other, and the combined effect makes the actual bandwidth lower than either upper limit. The conclusion further points out that reflection of high-frequency signals also limits the measured bandwidth.
Note also that only five sizes were compared, with diameters of 100, 20, 10, 5, and 2µm, and no sizes between 5µm and 2µm were measured. The result that 5µm was fastest therefore holds for this device structure and these measurement conditions, and does not mean that a 5µm diameter is optimal for all micro-LEDs.
Aimed at data-center optical interconnects; demonstrating data transmission remains for the future
The main application the paper envisions is optical interconnects inside data centers. As AI and high-performance computing spread, the bandwidth density required of interconnects keeps rising, but in electrical wiring the power consumed by circuits such as SerDes (serializer/deserializer) grows as the per-lane transmission rate is raised. The team therefore focuses on arranging micro-LEDs at high density and transmitting data in parallel over many optical channels. The idea is to secure high overall bandwidth density by sharing communication capacity among many channels rather than making each channel extremely fast.
An earlier study that presented this concept as a concrete system is Microsoft Research's "MOSAIC." Presented at SIGCOMM 2025, MOSAIC uses many relatively slow channels in parallel in place of a few ultra-fast ones, and its prototype ran 100 channels at 2 Gbps each. Over a 20 m transmission distance, it measured the bit error rate of 25 channels and reported that all of them fell below the threshold correctable by error correction. Microsoft explains that whereas a conventional 800 Gbps link is built from 8 channels × 100 Gbps, MOSAIC can build it from 400 channels × 2 Gbps, and it assumes micro-LEDs with diameters of roughly 5–10µm. However, extension to 800 Gbps and beyond and transmission distances of up to 50 m were presented as configuration examples, not performance demonstrated by the prototype.
The 5µm device that was fastest in this study falls within the 5–10µm range MOSAIC assumes. However, modulation bandwidth and bit rate are different metrics, and this paper did not conduct actual data-transmission tests. As a result, how many Gbps of data a device with a 2.75 GHz modulation bandwidth can send is not known at present. In the university's press release, Abouelnagga says that 5µm devices could allow many emitters to be integrated into a small two-dimensional array, each carrying different data. The paper also points out that devices miniaturized to a few micrometers emit more directional light, which may reduce reliance on external microlens arrays.
The team also presents ways to widen modulation bandwidth further. The paper's conclusion explains that the devices measured had a large impedance mismatch with the 50Ω measurement system, and that proper impedance matching could improve performance substantially. As device-structure improvements, it lists optimizing the p-side contact and deeply etching the conductive n-GaN beneath the pad to reduce the 0.18 pF pad parasitic capacitance.
This study shows that beyond a certain point of miniaturization, resistance and impedance mismatch, rather than junction capacitance, begin to hinder micro-LED speed. Next, data will need to be transmitted with devices that improve on these limits, measuring the achievable bit rate and bit error rate. Those results will reveal how far the 2.75 GHz modulation bandwidth translates into practical performance as a transmitter for data-center optical interconnects.