The patent defines links and RF links as follows.
The patent states that a satellite can have 32 RF link pairs, and each RF link may include multiple frequency bands:
“In an embodiment, the SAT 104 can include 32 RF link pairs (e.g., 32 UL RF links and 32 DL RF links), and each of the DL and UL links can include a plurality of frequency bands (e.g., a DL can have 8 frequency bands, a UL can have 8 frequency bands).”
We already know that the Ku-band user link consists of:
A natural interpretation is that a single RF link can span the entire bandwidth by including all eight channels.
However, the wording is ambiguous.
Does the patent mean that each link can have eight frequency bands, or that each RF link can have eight frequency bands?
Does “can have” mean that eight bands are always active simultaneously, or simply that up to eight bands are supported?
Must these eight bands be eight non-overlapping frequency channels, or could they instead represent reuse of the same channel across eight geographically separated beams?
In my previous analysis, I assumed that each RF link simultaneously carried eight 8*250 MHz downlink channels. This resulted in a total downlink capacity of 320 Gbps for 32 RF links.
However, 320 Gbps is much higher than the publicly reported capacities of current Starlink satellites. Public sources typically report approximately 20 Gbps for first-generation satellites and around 80 Gbps for V2 Mini satellites. Even the not-yet-launched V3 satellites are reported to provide about 1 Tbps of downlink capacity —— not super higher than 320Gbps.
Therefore, either my previous blog interpretation was incorrect, or the 2021 patent describes a configuration that was considerably more ambitious than the satellites eventually deployed so far.
Additional information in the patent provides further insight.
Regarding the relationship between the modem and the RF link, FIG.5 suggests that a modem contains one PHY transmitter and one PHY receiver, implying a one-to-one correspondence between a modem and an RF link.
However, later figures suggest a more complex implementation. A modem may contain two transmitter/receiver pairs, supporting two independent 250 MHz Ku-band downlink channels. In that case, one modem would contain two RF links.
FIG.20 illustrates an example satellite configuration with four modems for Ku band downlink. Each modem supports two 250 MHz channels, and each channel corresponds to one beam, resulting in a total of eight 250 MHz downlink channels. According to the earlier RF link definition, this is an example of one modem (or RF link) supporting two frequency bands, rather than eight.
These eight 250 MHz channels provide a total downlink capacity of 10 Gbps, noticing 5 bit/s/Hz in my previous blog.
The patent also notes that the four-modem configuration in FIG.20 is only an example and could instead use 16 modems.
With 16 modems, the satellite would support 32 independent 250 MHz downlink channels, providing a total downlink capacity of 40 Gbps. This example also aligns naturally with the patent’s reference to 32 RF links.
Based on this interpretation, the Ku-band downlink capacity explicitly described in the patent falls in the range of 10–40 Gbps, which is much closer to the reported capacities of current Starlink satellites than the 320 Gbps estimated in my previous blog analysis.
]]>According to FIG.21, Starlink uses two types of modem chips/modules based on capability of demodulation bandwidth and data rate.
Modem for the Ku-band User Link
This modem is used for:
It provides a peak data rate of 1.25 Gbps over 250MHz bandwidth.
Modem for the Ka-band Feeder Link (satellite — gateway)
This modem is used for:
It provides a peak data rate of 2.5 Gbps over 500MHz bandwidth.
From these informations, the Starlink modem achieves a spectral efficiency of 5 bit/s/Hz.
The highest modulation supported is 64QAM, whose ideal uncoded spectral efficiency is 6 bit/s/Hz. After accounting for channel coding, packet headers, and other protocol overhead, the resulting 5 bit/s/Hz is both efficient and a reasonable engineering result.
In deployment:
Each Starlink satellite user link provides:
Each RF chain supports eight parallel frequency channels:
32 chains in total offer 80Gbps (2.5x32) UL and 320Gbps (10x32) DL capacity.
Each RF chain supports approximately 200 simultaneously active terminals. Each satellite supports approximately 6,400 (32 chains times 200 UT/chain) simultaneously active terminals.
Assuming equal resource allocation:
In practice, depending on service tiers and network oversubscription, a satellite can serve significantly more than 6,400 subscribed users.
]]>The UW is used differently on different communication links.
Satellite–Terminal Link
Uplink: When multiple terminals access the same satellite, each terminal is assigned a different UW. The UW is generated by a 7-stage LFSR-based PN generator, driven by the Channel ID.
Downlink: When multiple satellites cover the same geographic area, for example to improve reliability or avoid signal blockage, each satellite is assigned a different UW.
Satellite–Gateway Link
Communication between a satellite and a gateway station is always point-to-point. There is no scenario in which multiple satellites communicate with the same gateway simultaneously, or multiple gateways communicate with the same satellite simultaneously.
However, this link uses two Ka-band polarizations, LHCP and RHCP, and a different UW is assigned to each polarization.
]]>Starlink supports the following modulation schemes:
Unlike Wi-Fi and LTE, Starlink includes 8-QAM and 32-QAM. A possible reason is that satellite links are highly power-limited due to large propagation losses. Finer modulation granularity allows Adaptive Modulation and Coding (AMC) to utilize the available link margin more precisely and avoid wasting transmit power.
The patent explicitly mentions the Adaptive Modulation and Coding (AMC) mechanism.
User data is encoded using LDPC. Data scrambling uses a 15-stage Linear Feedback Shift Register (LFSR).
The physical-layer packet header, analogous to the Wi-Fi SIGNAL field, is encoded using a 64-state (constraint length 7), rate-1/3 convolutional code with BPSK modulation. This is similar to the coding used for LTE control/broadcast channels such as the PDCCH and PBCH.
The packet header contains:
Within a radio frame, the physical-layer header of the first PDU contains 27 bits of information, while the headers of subsequent PDUs contain 23 bits.
PDUs are divided into data PDUs and signaling PDUs. Signaling PDUs are used for communication between MAC entities, such as the random access procedure (RACH) between the terminal MAC and the satellite MAC.
Starlink also employs a PN-based symbol scrambling method for modulated symbols. A Linear Feedback Shift Register generates a random 2-bit value that rotates each transmitted symbol by 0°, 90°, 180°, or 270°. This scrambling can transform a BPSK symbol stream into a QPSK constellation.
All frequency-domain subcarriers, including both data and pilot subcarriers, are scrambled using PN phase rotations generated by a 15-stage LFSR. As a result, BPSK pilot symbols become QPSK after scrambling.
The 128-sample base sequence used for the time-domain Unique Word (UW) is also originally BPSK and is converted to QPSK through PN phase rotation scrambling. UW scrambling uses a 7-stage LFSR.
]]>A small change in clock rate is also caused by special (motion) and general (gravity) relativistic effects but it is quite small in low Earth orbits (about 0.2 ppb or parts per billion).
This is the relativistic effect considered by Starlink when analyzing synchronization error sources in an OFDM system.
The conclusion is that the clock-rate difference caused by relativistic time dilation is much smaller than other synchronization error sources, such as Doppler shift and sampling-time drift due to satellite motion, crystal oscillator manufacturing tolerance, and temperature drift. Therefore, it can be safely neglected.
Perhaps I simply have not read enough papers, but this is the first time I have seen relativistic effects explicitly discussed in the context of OFDM synchronization algorithms. （of course I know that GNSS does consider it)
I do wonder whether, in deep-space communication scenarios—such as solar probes or Mars missions—the clock errors introduced by relativity become significant enough to require explicit compensation for communication synchronization.
Although I have not yet finished reading the patent, I now strongly feel that this is not merely a patent. It is a comprehensive treatise on modern OFDM communication engineering, covering fundamental principles, system design, and implementation details with remarkable completeness. Anyone who thoroughly understands every section of this patent would gain an excellent foundation in communication theory, system design, and engineering practice.
The first author of the patent, Martin McCormick, once remarked with characteristic humility:
“I’ve earned the somewhat dubious distinction of holding more patents than anyone else working for Elon Musk. I’ve invented about 10% of SpaceX’s entire patent portfolio.”
After reading this patent, I understand that statement very differently. He was not simply writing patent applications—he was writing what is effectively a technical textbook. There is nothing “dubious” about that achievement. The breadth and depth of his expertise are truly remarkable.
It is also said that he was among the last students of Alan V. Oppenheim. If so, his work certainly lives up to that reputation.
]]>The most distinctive feature of the Starlink pilot design is the pilot structure embedded within data OFDM symbols. It uses block-based pilot clusters. Two pilot clusters are placed at the two edges of the allocated frequency band, and each cluster contains eight consecutive pilot subcarriers. In contrast, Wi-Fi and cellular systems use distributed pilot subcarriers scattered across the frequency band.
A Starlink packet (burst) begins with an OFDM symbol consisting entirely of pilot subcarriers, similar to the Long Training Field (LTF) in Wi-Fi. This is followed by pilots embedded within data OFDM symbols for tracking. Wi-Fi follows a similar approach.
At the burst level, Starlink is architecturally closer to Wi-Fi. Cellular systems are closer to a continuous streaming structure, where pilots are periodically distributed across the entire time-frequency grid. However, Starlink also defines a periodic radio frame structure above the burst level. In this sense, Starlink combines concepts from both Wi-Fi and cellular systems, taking advantage of each.
The use of two pilot clusters instead of distributed pilots reflects the different situations faced by different systems.
The phased-array antennas on both the satellite and the user terminal provide strong directivity, suppressing multipath propagation. As a result, the channel exhibits relatively weak frequency selectivity. However, because of the high velocity of LEO satellites and the higher carrier frequencies, the channel exhibits much stronger time variations.
Block-based pilot clusters make it convenient to estimate inter-carrier interference (ICI). The ICI pattern resembles an FIR filter response. Once the ICI coefficients are estimated, equalization across multiple adjacent subcarriers can suppress ICI. The main sources of ICI include:
Block-based pilots also allow averaging across adjacent pilot subcarriers to improve the estimation accuracy of Common Phase Error (CPE), resulting in better tracking of the low-frequency components of phase noise. This averaging is effective because adjacent subcarriers experience similar channel responses, especially when frequency selectivity is weak.
Finally, placing the two pilot clusters at the edges of the allocated frequency band makes them more sensitive to the subcarrier-dependent phase rotation caused by Sampling Frequency Offset (SFO), which accumulates from one OFDM symbol to the next. This improves SFO tracking and compensation. SFO includes both the intrinsic frequency error of the sampling clock and sampling-time drift caused by the motion of the LEO satellite.
]]>Each burst begins with a Unique Word (UW), followed by multi-user OFDM symbols.
For an individual user:
The pilot clusters are placed at the two edges of the user’s allocated frequency band. Their offsets from the band edges are configurable, and the placement is symmetric.
A radio frame may contain multiple bursts. Starting from the second burst, the UW is optional.
Each user may occupy multiple Resource Blocks (RBs), with multiple RBs mapped to contiguous subcarriers. Different users occupy non-overlapping frequency bands, beginning with the non-overlapping channel estimation subcarriers in frequency domain.
A channel supports up to four users.
The UW is used for:
The Unique Word (UW) at the beginning of each packet is indeed likely different for different users. However, it is associated with the channel ID, rather than the terminal’s MAC address. FIG.12 confirms this.
The meaning of CFR is also confirmed. It indeed stands for Crest Factor Reduction, not Channel Frequency Response. The patent states:
“Transmitter components including at least a crest factor reduction module and a digital pre-distortion module.”
The patent further confirms the following:
Why is the gateway-to-satellite channel the widest?
Most likely because it carries the aggregated Internet DL traffic for a large number of user terminals, including web browsing and video streaming.
Channel bandwidth is changed by adjusting the sampling rate, while keeping the FFT size unchanged. This matches my earlier guess based on measured uplink signals.
The modem’s configurable parameters include:
Facts
Each uplink packet begins with a Unique Word (UW) consisting of:
However, it is still unclear whether this 128-sample sequence is identical for all user terminals. If different terminals use different sequences, how are they assigned? For example, are they derived from the terminal’s unique MAC address?
What the patent says about the Unique Word
According to the patent, the UW is used to:
A hypothesis
Based on these descriptions, it seems likely that each terminal uses a different UW (i.e., a terminal-specific signature sequence). Otherwise, it would be difficult to estimate the CFO of individual terminals when their uplink signals overlap.
In this sense, the UW resembles a signature sequence in a CDMA system.
An interesting observation
From the captured uplink waveform, the full-bandwidth UW is immediately followed by partial-bandwidth (RB: resource-block) OFDM symbols.
This suggests that a UW lasting only one OFDM-symbol duration is sufficient for uplink access. The satellite appears able to identify the transmitting terminal, estimate its CFO, and perform compensation using only the UW at the beginning of each packet, even multi terminals’ signal are overlapping.
Compared with the conventional random access procedures used in terrestrial cellular systems (e.g., the four-step contention-based or three-step contention-free procedures in LTE), the Starlink approach appears much simpler and more efficient.
Because satellite links have much longer propagation delays, multi-step stateful random access procedures may introduce unacceptable overhead. This appears to be an optimization specifically designed for satellite communications.
Why might this work?
My hypothesis is that Starlink terminals perform relatively accurate propagation-delay and Doppler pre-compensation before transmission.
As a result, even if the UWs from multiple terminals overlap, they arrive at the satellite in approximate time alignment (perhaps with timing errors smaller than the CP duration). This would significantly reduce the complexity of detection or enable more advanced multi-user detection algorithms.
Another possible reason is that the number of active terminals within a single satellite beam is relatively small—much lower than in a terrestrial cellular cell—making multi-user detection more manageable.
In contrast, terrestrial cellular systems such as LTE generally do not assume timing or Doppler pre-compensation during PRACH random access. Consequently, the base station must handle a fundamentally asynchronous, CDMA-like multi-user detection problem. This is more challenging for the network, but it keeps the terminal implementation simpler. Reliable identification of different terminals is then achieved through the subsequent multi-step random access signaling procedure.
]]>Today I only read the first small section, but I already found two interesting points.
1 .
A radical departure from the traditional satellite modem design philosophy
In conventional satellite communication systems, the modem design differs significantly between the user terminal, the satellite payload, and the feeder link (the satellite-to-gateway link). This is because these links operate under substantially different conditions, including frequency band, bandwidth, data rate, operating environment (mobile vs. fixed, multipath characteristics, antenna capability, etc.). As a result, each modem is typically optimized specifically for its own role.
In the Starlink system, however, the user terminal, the satellite, and the gateway station all use a common OFDM modem architecture. The same modem is adapted to different links and hardware characteristics primarily through what the patent describes as a “configurable” design.
This is yet another major break from traditional satellite communications. In terms of how disruptive it is, I would say it rivals Starlink’s earlier decision to deploy OFDM on satellites at large scale.
What surprised me even more is that the feeder link (satellite to gateway) also uses essentially the same OFDM modem as the user link (satellite to terminal).
Traditionally, feeder links operate in higher frequency bands. Gateway stations are equipped with high-gain antennas and high G/T, and there are many mature high-speed single-carrier modems available for such links. Compared with wideband OFDM, high-speed single-carrier waveforms are much more friendly to power amplifiers.
In the Starlink system, however, essentially the same or a very similar OFDM modem architecture is used throughout the entire network. This undoubtedly simplifies the overall system design, reduces hardware cost, and decreases the number of different modem vendors that must be managed. It is an aggressive and bold engineering decision.
This is absolutely not the kind of design that would come from a traditional aerospace organization. Had such a proposal gone through an internal design review at a conventional aerospace company, it would almost certainly have been heavily criticized.
Only an outsider who is unconstrained by conventional thinking would be willing to attempt something like this.
2 .
There appears to be an error in the patent’s Summary section
The patent states:
“The modem can further include an orthogonal frequency division multiplexing baseband processing component and a digital front-end processing component for at least one of digital pre-distortion and a channel frequency response (CFR).”
Here, I believe CFR is incorrect.
It should most likely refer to Crest Factor Reduction (CFR) rather than Channel Frequency Response, because Crest Factor Reduction is commonly paired with DPD (Digital Pre-Distortion) in digital front-end processing for OFDM power amplifier linearization. The pairing of DPD and CFR is standard practice, whereas “Channel Frequency Response” does not fit naturally in this context.
]]>