GaN on silicon Primed for RF power
Can GaN-on-silicon move on from just serving 600 V and below power management systems, wireless charging and LIDAR, to win deployment in next-generation RF applications, such as 5G and the Internet of Things? By Markus Behet and Joff Derluyn from EpiGaN
For several decades, innovations in electronics have tended to draw on silicon and GaAs-based technologies, devices and integrated circuits. Take cellular infrastructure: its increases in complexity from 2G to 4G-LTE have been underpinned by continuous improvements in GaAs HEMTs, HBTs and silicon LDMOS devices.
One downside of all these incumbent semiconductor technologies is that they are now encroaching their theoretical limits. Consequently, any improvement at the system level requires a huge effort by design engineers, and draws on ever increasing developments costs. Given this state of affairs, there is an urgent need for a new RF semiconductor technology to step in and fulfil the promises of 5G RF systems "“ and it is here that GaN technology comes to the stage.
Wideband GaN technology is definitely not a newbie to the semiconductor market. It is already being deployed in the likes of power supplies and motor drives, which are benefitting from highly efficient 600 V GaN-on-silicon HEMTs. In addition, there are GaN-on-SiC RF products that are being used in base stations, Satcom, military and CATV systems. Due to all this activity, on both of these formats of GaN, the total addressable market in 2017 was worth $400 million.
Another opportunity for revenue growth will come with the dawn of the next generation of cellular infrastructure "“ 5G. Initial deployments, which will take place by 2020, promise to lead to: peak data rates of 20 Gbit/s; a large number of users or sensing nodes for any given area; a high power efficiency, enabling less power per transmitted bit; a latency of less than 1 ms; and ubiquitous connectivity. This revolution will not only enhance existing telecom services drastically, but enable emerging applications, such as virtual/augmented reality, autonomous cars, massive IoT, and mission critical services.
These changes have led market analysts ABI to argue that 5G should be viewed as a general purpose technology that will act as a catalyst for transformative changes of work processes, and will establish a new set of rules of competitive economic advantages. The impact will be so great, says ABI, that by 2035 it will trigger a global economic output of $12.3 trillion "“ that's comparable to current levels of annual US consumer spending.
Figure 1. EpiGaN's optimized RF GaN-on-silicon HEMT structure produces very low RF losses, according to measurements on transmission line structures.
The pillars of 5G
To meet stringent user requirements, workgroups defining the 5G roadmap have devised a novel architecture for the 5G network. One of its three key features is the use of many small network cells, known as pico- and femto-cells. They will allow a high user density and low latency. Another key ingredient in 5G systems is the use of transceivers that operate at a higher baseband frequency and have a very large (analogue) bandwidth that can meet digital bandwidth requirements. The third important innovation associated with the introduction of 5G is the move from omni-directional broadcasting to the use of directional beams for each connection. This will enhance energy efficiency, and enable more efficient use of the RF spectrum.
Figure 2. A typical structure for a GaN-on-silicon HEMT includes a GaN channel, a barrier made from AlGaN, and either a GaN cap or a SiN passivation layer.
What are the requirements for the RF technology that can serve 5G architecture? Firstly, it will need to support RF networks operating at much higher frequencies. This domain is less congested than that used today, but spans 6 GHz to as high as 100 GHz. New frequency bands within this domain will have a much higher useable bandwidth, and will ideally be covered with a single amplifier module.
Another major change is the introduction of massive MIMO (multiple-in, multiple-out) beamforming. This technology features multiple transmit and receive chains for each antenna element in the phased-array configuration of the transceiver system. Compared with a single omni-directional antenna, the power per antenna element is reduced, so it is possible to trim the size of the amplifier.
One of the challenges with MIMO is that the antenna elements must all emit signals that have a well-defined phase relationship with one another. Due to this criterion, the technology adopted for the RF amplifier must allow close integration with the driving digital controllers. Note that it is much easier to realise this using silicon substrates than those made from SiC.
The requirements listed above define the mandatory attributes for power amplifier technology for 5G infrastructure and its associated handset architectures. These requirements imply the need for devices that combine compactness with low cost, high power density, linearity at millimetre-wave frequencies and integratability with CMOS technology.
Here GaN-on-silicon technology comes to the fore. It is a fundamentally superior RF semiconductor technology that is ideally placed for fulfilling demanding performance targets. Compared with GaAs, the key, proven attributes of GaN are: a power density, per millimetre of gate periphery, that is ten times higher; a higher efficiency; operation at a higher voltage, leading to a reduction in impedance transformation challenges; and a superior broadband operation at high frequencies. What's more, GaN can operate at higher device operating temperatures than GaAs. Thanks to this, cooling requirements are easier, maintenance costs lower, and there is an increase in device reliability.
Improving the foundation
Within the technology toolbox, there are at least two variants of GaN to choose from. One is GaN-on-SiC, a pairing of materials that provides the ultimate power levels in the most demanding applications. The common alternative is GaN-on-silicon. This can serve cost-sensitive, high-volume markets that may require large-diameter wafers; and it can deliver an RF performance that is vastly superior to low-frequency silicon LDMOS and expensive, low-power GaAs HEMT technology.
At EpiGaN of Hasselt, Belgium, we are at the forefront of developing GaN-on-silicon epiwafer technologies for RF devices. Since our founding in 2010, we have achieved significant technological milestones that will help to drive the adoption of GaN-on-silicon for the next stage of cellular infrastructure "“ 5G.
Several hurdles must be overcome if GaN-on-silicon is to compete directly with GaN-on-SiC in high-end RF applications. Some are intrinsic, such as the lower thermal conductivity of silicon compared to SiC, which can be countered by slashing the substrate thickness to 50 μm during device processing; and the need to minimize the conductive interface between the silicon substrate and the III-nitride buffer layer by an optimized process during the initial stages of epitaxial growth. A too conductive interface causes a parasitic conduction path for the RF signals, which leads to an undesired dissipation of those signals that is exacerbated at higher frequencies. It is paramount to avoid this, as transistors manufactured on such lossy substrate/buffer combinations will never attain good performance at high efficiencies.
To mitigate this conductive path, we have developed a robust, optimized interface technology. It trims the RF signal loss on GaN-on-silicon material to below 0.2 dB/mm for today's basestation frequencies, which is very close to the value obtained on the much more expensive GaN-on-SiC material. Even in the E-band up to 100 GHz, where future 5G networks may operate, RF signal loss is well below 1 dB/mm (see Figure 1).
Today, the most common platform for manufacturing low-loss RF GaN-on-silicon technology is resistive, float-zone silicon substrates with a diameter of up to 150 mm. We are breaking new ground, having recently released this RF GaN process on high-resistive 200 mm silicon substrates produced by the Czochralski process. Devices formed with these epiwafers deliver comparable performance and showcase the potential for further reductions in the cost of RF power GaN-on-silicon technology. Another asset of these larger epiwafers is that they will ease entry into today's mainstream 200 mm lines at silicon IDMs and foundries. 
Figure 3. Researchers at IEMN-CNRS have used EpiGaN material to produce 0.15 μm gate length AlN/GaN-on-silicon prototype transistors that deliver encouraging levels of power, gain and power-added efficiency at 6 GHz; courtesy of F. Medjdoub, IEMN, Lille.
Beautiful binary barriers
At the heart of the most common structure for an RF GaN HEMT is a ternary AlGaN barrier, typically 20 nm thick and sporting an aluminium content of around 25 percent. This barrier is capped and protected by an ultra-thin GaN layer, having a thickness of between 2 nm and 3 nm (see Figure 2). 
We developed a markedly different structure. To enable the ultimate high-frequency RF performance, we combine pure AlN barrier layers, rather than AlGaN ternaries, with an in-situ SiN cap layer. Armed with this pairing, the thickness of the barrier can be slashed from typically 20 nm to just 4 nm to 6 nm, allowing the transistor's gate to be positioned very close to the densely populated channel, and maximising the electrostatic coupling between the two "“ in other words, improving gate control.
These modifications create superior transistors. Bringing the gate closer to the channel maximises the transistor's transconductance, a key characteristic for RF amplifiers. There is also a dramatic suppression in the so called "˜short channel parasitic effects', unwanted effects that include the reduction of the transconductance in transistors with gates that are below 0.15 μm and have a poor aspect ratio for the gate length to gate-to-channel distance "“ ideally, it should be 15 or more. 
Figure 4. The CW power performance of a 2 à— 25 μm AlN/GaN HEMT produced by a team at IEMN-CNRS. The device is operating at a frequency of 40 GHz, and drain-source voltages of 15 V and 20 V. Results are reproduced courtesy of F. Medjdoub, IEMN, Lille.
Another merit of the switch from an AlGaN barrier to one made of AlN is an increase in the inherent piezoelectric effect to the maximum possible value. Due to this, carrier densities exceed 2 x 1013 cm-2 in the two-dimensional electron gas in the channel, leading to an increase in power density. If a well-considered thermal chip layout is adopted, significant reductions in chip size can follow.
New device features
We have developed and optimized a process for sealing the top of GaN-on-silicon wafers directly after growth. To do this, we use an in-situ grown SiN passivation layer. This approach prevents exposure of (Al,Ga)N layers to the fab environment "“ that's a major concern for CMOS fabs, which are far more willing to process our epiwafers than standard GaN-on-silicon structures.
That's not the only benefit of our SiN passivation layer, however. It can also serve far broader purposes, including the creation of a unique gate dielectric with a smooth, contamination-free surface. By controlling the filling of surface states during device operation, our SiN layer can provide enough charge to neutralize the surface charge of the AlGaN barrier layer in a GaN-on-silicon device, so that its surface potential no longer contributes to depletion of the two-dimensional electron gas "“ a phenomenon known as current collapse. Another benefit of the SiN layer is that it improves device stability at elevated temperatures.
Using in-situ SiN deposition, the AlGaN barrier may be replaced with pure AlN without any material degradation. For such a SiN/AlN/GaN heterostructure the sheet resistance falls well below 350Ω/sq. Such a low value enables the fabrication of transistors with higher current densities "“ thus smaller, cheaper devices for the same current rating.
We have collaborated with researchers at IEMN-CNRS, who have produced prototype transistors with our GaN-on-silicon RF wafers. The resultant devices, on un-thinned silicon substrates, have a power-added efficiency approaching 60 percent at 6 GHz (see Figure 3).Great performance continues at higher frequencies.
At 40 GHz, in-situ SiN capped AlN/GaN transistors with a 120 nm gate length can deliver a peak output power density of 4.5 W/mm, and an associated power-added efficiency of 46.3 percent (see Figure 4). And that's not the limit: optimising the buffer to refine heat dissipation has enabled the power-added efficiency at 40 GHz to climb to more than 50 percent, with an associated gain exceeding 10 dB. We have also partnered with Ommic, a leading European GaN foundry. It has used our unique technology to develop a 100 nm gate-length, open foundry MMIC process with a complete design kit.
Figure 5. A fully integrated Ka-band transmit/receiver built at Ommic, utilizing EpiGaN's SiN/AlN/GaN-on-silicon RF HEMT epiwafer technology. The PA is located at the top side of the chip, while the LNA is at the bottom side and the switch at the left side (antenna port). Image produced courtesy of Ommic, Limeil Brévannes, France.
Compared to standard GaAs pHEMT processes, the RF GaN-on-silicon devices produced by this process have a far higher breakdown voltage "“ it is 40 V. Thanks to this, the output power density in the Ka-band is much higher, with a typical value of 3.3 W/mm and a peak of up to 5.7 W/mm. What's more, the device features far greater robustness to input mismatch conditions.
Designers at Ommic have demonstrated the capability of our technology with a fully integrated transceiver operating in the Ka-band. It has a chip size of just 11 mm2. Operating at 30 GHz, the output power of the power-amplifier-and-switch transceiver is above 35.5 dBm, while the power amplifier alone offers 37 dBm. Meanwhile, gain of the low-noise-amplifier-and-switch transceiver part, evaluated in receive mode, exceeds 18 dB (see Figure 6).
The results obtained by engineers at IEMN-CNRS and Ommic add further weight to the view that RF GaN technology will come to the masses "“ in fact, there is no doubt about that. Both GaN-on-SiC and GaN-on-silicon are set to enjoy success, with levels of sales of the two formats depending on the cost-to-performance trade-off for the specific RF application.
Today, we are already seeing GaN-on-SiC HEMTs displacing incumbent GaAs pHEMT and silicon LDMOS technology in 4G-LTE infrastructure systems, while GaN-on-silicon is rapidly closing the performance gap and offering an increasingly tempting bang-per-puck. The benefits that RF GaN technology brings to the table are simply too attractive to ignore, and they are already spurring innovations in many existing applications.
To reach true mass-market adoption of RF GaN technology for 5G, there will need to be a mature supply chain and ecosystem. While this not in place today, big industry players "“ that is, IDMs and pure-play foundries "“ are positioning themselves to build up a strong supply chain for RF GaN device manufacturing. It is becoming ever more evident with GaN that the dawn of a new RF technology has begun, and that the incumbent silicon and GaAs devices will lose considerable market share in next-generation cellular RF infrastructure applications.
Last but not least: will RF GaN-on-silicon technology win deployment in future smartphones? It works with relatively high voltages of 10 V or more today, so it is not the most suitable technology for handsets, which currently use between 3 V and 5 V. But it could quickly become a viable candidate, as standards evolve, carrier aggregation is introduced to increase bandwidths, and there are ever increasing performance requirements for multi-mode, multi-band PAs.
Figure 6: Transmit and receive small-signal gain (left) and transmit PA output power (right) of a Ka-band transceiver MMIC (with and without switch), for the chip in Figure 5; courtesy of Ommic, Limeil Brévannes, France.
