1. Introduction

Driven by the demand from mobile and Internet-of-things (IoT), global wireless traffic is growing explosively, with more than 70% occurring indoor [1]. Compared with 4G, dedicated indoor radio access network (RAN) become crucial for 5G and beyond. This is because outdoor (B)5G base stations will face more technical difficulties to serve indoor service, due to the fact that 5G signals with higher carrier frequency (even sub-6GHz) suffers from much more propagation loss and penetration loss through walls and windows [2,3]. In this context, an indoor cloud RAN (C-RAN) architecture with centralized processing and fronthaul (FH) transport would be promising, which can reduce cost/energy consumption and provide centralized control and management when a plurality of indoor radio units (RU) are deployed to improve coverage and user experience [4].

Multimode fibers (MMF) such as OM2 and OM3 [6] fibers deployed years ago still constitutes a large portion of cable infrastructures indoor, e.g., premises, shopping malls, factories, etc., with millions of kilometers installed worldwide [5–7]. Reusing these existing fibers for indoor (B)5G radio access system could significantly reduce capital expenditure (CAPEX) and deploying time. However, MMFs typically have limited bandwidth-distance product, e.g., 2000MHz*km effective modal bandwidth (EMB) for OM3. If traditional FH interfaces like common public radio interface (CPRI) [8] continues to be used, for instance, an RU with 16 antennas serving 50-MHz 5G NR traffic [9] would require a CPRI-equivalent rate of about 32.4-Gb/s, assuming an I/Q sampling frequency of 61.44-MHz. Considering on-off keying (OOK) signaling, this would severely limit the achievable distance to about 70m for OM3 MMF [10], which is far from enough to support the distance of in-building backbone cables plus horizontal cables [5]. For example, about 70% of in-building MMF cables are longer than 100m, and 40% over 200m [6], while 10GbE standard stipulates an achievable distance of 300m for OM3 [7]. Analog radio-over-fiber (A-RoF) approach that effectively alleviates this issue has been discussed [11–14]. However, the nonlinearity induced by optical and electrical components such as 850nm-band vertical cavity surface emitting laser (VCSEL) could easily degrade the fidelity of radio signals. The support of larger-scale multiple-input multiple-output (MIMO) is another challenge, as state-of-the-art technology only demonstrated 3×3 MIMO capability [14]. On the other hand, data compression approaches that aim at overcoming bandwidth bottleneck of digital radio over single-mode fiber (SMF) links have also been actively investigated, such as scalar quantization [15–19], delta-sigma modulation (DSM) [20] and vector quantization [21]. Compared with scalar quantization and DSM, vector quantization can achieve more bandwidth reduction at the expense of higher complexity. Despite significant progress, most conventional schemes only compress in time domain with moderate compression ratios around 40%∼50%, which may limit their application in MMF scenarios that have more severe bandwidth issue.

Here we introduce an alternative scheme, analog-to-digital-compression radio-over-multimode-fiber (ADX-RoMMF). In this scheme, radio signals received by multiple antennas of RU are jointly compressed and converted to digital stream, which can significantly improve (e.g., ∼10-fold) FH bandwidth efficiency by reducing redundant information in both space and time domain. Moreover, such compression DSP and coding/modulation of the MMF link are jointly designed and optimized, forming a converged wireless-wired link. As a result, ADX-RoMMF can greatly increase the capacity and/or reach of RoMMF system, enabling more operation margin and larger in-building coverage. Meanwhile, its digital nature brings advantages over A-RoF, such as robustness to nonlinear link distortion and low-cost transceiver module. Note that, in such digital-based scheme, DSP-induced latency must meet the stringent (B)5G requirement.

In this paper, we experimentally demonstrate the potentials of ADX-RoMMF for future indoor radio access. Uplink 16 MIMO channels of 1024QAM 5G NR-class signals, corresponding to a CPRI rate of 32.4Gb/s, are compressed by real-time field-programmable gate array (FPGA)-based ADX prototype and successfully transported over 850m OM3 MMF. The support of 3 different candidate (B)5G formats is also shown. Moreover, we validate latency overhead added by ADX DSP to be less than 500ns, which is negligible considering the overall FH latency requirement usually on the order 100us.

2. Network concept and the real-time ADX prototype

Figure 1(a) shows the schematic of a building with legacy MMF cable infrastructure, including backbone cables and horizontal cables [6]. The topology may be point-to-point or tree-like with switching. Due to the loss issue of 5G frequency band, the uplink 5G radio signals from user equipment (UE) cannot be directly served by outdoor base stations. Instead, they are firstly received by a multi-antenna RU, digitalized (e.g., with ADX), and transmitted over indoor MMF to a distribution unit (DU) in the building, such as in the equipment room. After possible processing (e.g., decompression and functional split), the traffic is then transmitted outside of the building, over another segment of RAN (e.g., based on single mode fiber), to a central unit (CU) for centralized processing.

 

Fig. 1. (a) An indoor RAN reusing the legacy MMF. (b) Illustration of ADX with space-time-type architecture and joint optimization.

Download Full Size | PDF

An example architecture of ADX is shown in Fig. 1(b) with space-time-type. It is a cascade of spatial compression which reduces MIMO channel count from M (number of antenna elements) to K (K < M), and time-domain compression which reduces the number of bit per sample from, e.g., 15 in CPRI. An optional optical spectrum efficiency (SE) enhancement stage (e.g., by multi-level pulse amplitude modulation, PAM) can be involved to further reduce bandwidth of MMF-based indoor FH. Joint optimization of spatial and temporal compressor, as well as wireless part and wired (optical) part is a key to improve the performance of such wireless-wired converged RAN [22].

In this work, the real-time ADX implementation is based on the abovementioned space-time architecture. Spatial compression, or spatial dimensionality reduction, can be modelled as a problem of Karhunen-Loeve transform of RU-received signal matrix having M spatial channels and N time samples. We suggest that only adaptive filter-like Karhunen-Loeve transform, such as the family of fast subspace tracking [23], can be potentially employed in FH context due to its tight constraints on latency and RU-side complexity. Here the spatial compressor (or “spatial filter”, SF) is designed based on projection approximation subspace tracking (PAST) [24] because of its low complexity, robustness to finite precision (i.e., suitable for fixed-point implementation) and compatibility to complex-valued baseband radio signal. On the other hand, for temporal compression circuit, we leverage the principle of low-complexity adaptive quantizer, considering the differentiated signal power due to MIMO channel fading and possibly varied signal bandwidth due to radio resource allocation. Here adaptive differential pulse code modulation (ADPCM) [25] is employed as the basic algorithm.

In the real-time circuit design, a challenging issue is that the feedback loop latencies in the concepts of compression algorithms severely limits the circuit throughput to only several MHz [26]. This is not sufficient to process 5G NR-level bandwidth signals, which would induce large buffering latency (e.g., at the order to radio signal packet duration) before compression and add heavy burden on the total FH latency. To eliminate such buffering latency, it is crucial to design the ADX circuit to break the algorithms’ limitations and to have NR-level throughput for all MIMO channels.

Figure 2(a) shows the overall architecture of our proposed ADX circuit. The spatial compression circuit has been designed to be a feedforward-feedback decoupled structure. The SF (i.e., multiplication of filter matrix W^H and input signal vector x) works at a throughput of hardware clock rate of 122.88MHz, while the filter update part shown in Fig. 2(b) works at 6.144MHz throughput. SF is updated every L input sample period. L is positively correlated with throughput of the spatial compressor, and negatively correlated with converging speed of the filter update part. We selected L=10 and the throughput of 61.44MHz is achieved. With PAST algorithm, the filter update part converges within several tens of µs according to the experiments, which is still much faster than the variation speed of MIMO channel (typically, ms-level [27]). Thus, the SF circuit with such decoupling can still deal with varied MIMO environment. On the other hand, in temporal part of ADX, we propose to use sub-banding filters (SBF) before adaptive quantizer as an efficient way of circuit parallelization. Moreover, the number of multipliers required for each SBF can be halved with polyphase implementation shown in Fig. 2(c). Considering a balance among throughput, SBF latency and out-of-band crosstalk, 2-band 16-tap Johnston type-B quadrature mirror filter (QMF) is employed, leading to throughput doubling. Also, the ADPCM algorithm has been modified to boost throughput, including modified gain adaptor [19] and halved prediction rate for the predictor, as shown in Fig. 2(d). In total, 61.44MHz throughput is also achieved for temporal part of ADX.

 

Fig. 2. (a) Overall architecture of the designed real-time ADX circuit. (b) Detail of the update circuit of spatial filter. Operator Tri{*} indicates that only the lower triangular part of the matrix is calculated, and its Hermitian transposed version is copied to the upper triangular part. Β=0.9999 in the implementation. (c) Detail of the SBF circuit with polyphase implementation. (d) Detail of the modified ADPCM circuit. “Pred.”: predictor. “Reg.”: register. I Photo of the designed hardware prototype.

Download Full Size | PDF

Overall, an entire 61.44MHz ADX throughput was achieved at 122.88MHz clock frequency of FPGA [28], which fully supports 5G NR bandwidth of 50MHz. With ASIC implementation, the achievable clock frequency can be significantly increased, and the proposed circuit has the potential to support wider radio signal bandwidth such as 100MHz and beyond in the future. Figure 2(I) shows the photo of the implemented hardware prototype. We developed the real-time ADX based on a single-chip field-programmable radio platform, i.e., Xilinx RFSoC (XCZU29DR), which integrates 16-channel, 4GHz-bandwidth DAC&ADC array, digital RF chain and FPGA on one chip [29]. We also utilize this platform to experimentally emulate multi-antenna/channel radio reception in the RU.

3. Experimental setup, results and discussion

We conducted experiments to investigate the performance of ADX-RoMMF in the application of indoor (B)5G FH. The setup is depicted in Fig. 3. At UE side, P independent 5G NR-like signals were transmitted over 4×16 MIMO channel with independent and identically distributed (i.i.d.) Rayleigh fading that emulates rich-scattering indoor environment [33,34], resulting in 16-channel signals received by RU. 1024QAM was used as the modulation format to show the high fidelity of this RoF link. Detailed parameters of the 5G NR-like signals are listed in Table 1. 768-sample preamble was added to assist the convergence of the spatial compressor. The MIMO channel outputs were then noise-added to emulate a finite wireless signal-to-noise ratio (SNR). “Wireless SNR” here means the ratio between the total power of 16-channel signals and the power of the added white Gaussian noise. Then, three cases were experimentally investigated and compared as shown in Fig. 3:

 

Fig. 3. Experimental setup.

Download Full Size | PDF

Table 1. Parameters of tested 5G NR-like signals

View Table

Case (I): the signals were compressed by the real-time ADX, including 16-by-4 spatial compressor and 8-bit time-domain compressor;

Case (II): the signals were compressed by a conventional time-domain-only compressor, such as a differential quantizer in [19];

Case (III): the signals were not compressed as in the CPRI case.

In Case (I), the 16-channel signals were subsequently up-converted to an intermediate frequency (IF) of 100MHz and emitted by the on-chip 16-channel, 14-bit DAC array of RFSoC operating at 983.04Msa/s. In the RU, the on-chip 16-channel, 12-bit ADC array digitized these 16 IF channels at 983.04Msa/s, followed by down-conversion to baseband. The ADX as described in the previous section performed MIMO signal compression. The CPRI equivalent rate is 61.44(MHz)×(15×2)×16×overhead=32.4Gb/s, where the control word (CW) overhead of 16/15 and line coding overhead of 66/64 [8] were assumed. The compression ratio of ADX compared with CPRI is 4/16×8/15≈13.3%, which means that the required FH bandwidth can be reduced by nearly 90%. The compressed data stream with was then 64b/66b line coded and output by an on-chip serial transceiver (GTY) with peak-to-peak amplitude of 400mV. The bit rate was approximately 4-Gb/s (4.05504-Gb/s).

In Case (II), the 16 channels were individually compressed by differential quantizers with 8-bit resolution (same as Case (I)), emulated offline. Then the data streams were serialized and line coded, and output by an SHF bit pattern generator (BPG) with peak-to-peak amplitude of about 400mV. CW was not emulated, and the bit rate was 16-Gb/s. In Case (III), the 16 channels were linearly quantized to 15-bit per I/Q branch without compression as in the CPRI case, serialized and line coded, and output by the SHF BPG. The bit rate was 30-Gb/s.

The electrical OOK stream encapsulating MIMO radio signals was converted to optical signal employing a directly modulated vertical cavity surface emitting laser (VCSEL) with bandwidth of about 21-GHz at 850nm band. 850nm VCSELs are off-the-shelf product and commonly used together with MMF links [37], which have certain cost and reliability advantages over long-wavelength VCSELs [38]. In Case (I), the bandwidth of the bias-tee of VCSEL was 6-GHz, while in Case (II) and (III), the bandwidth of the bias-tee used was 26-GHz as the bit rate was much larger. The optical OOK signal was then transmitted over various distances (up to 1000m) of OM3 MMF to the DU. In the DU, the OOK signal was received by a multimode photodetector (PD, bandwidth about 25-GHz) and captured by an analog-to-digital converter (ADC) of a Tektronix oscilloscope. In Case (I), the sampling rate and bandwidth of the ADC were 12.5-Gsa/s and 6-GHz respectively, while in Case (II) and (III), the sampling rate and bandwidth were 50-Gsa/s and 33-GHz respectively. Offline processing includes down-sampling, OOK bit decision, line decoding, ADX decompression, MIMO demodulation, and EVM evaluation. In this work, we didn’t use forward error correction (FEC) or digital equalizer, which avoids extra latency, power consumption and transmission overhead of FH link [39–41].

We first investigate the impact of indoor wireless SNR on the performance of ADX-RoMMF, i.e., Case (I). To single out its impact, we assume the MMF link is error-free. Figure 4(a) shows the measured EVM versus wireless SNR. The ADX supported P=1, 2, 3 and 4, and here we show the scenarios of P=4 and P=2. In P=4 case, EVM was less than 2.5% (threshold for 1024QAM [20]) when wireless SNR was larger than 30dB. The EVM floor, mainly caused by the time-domain compressor, was around 1.4%. In P=2 case, since the number of antennas (i.e., 16) and the number of ADX spatial compressor outputs (i.e., 4) are both unchanged, more spatial diversity gain [22] and lower EVM was observed. Note that throughout this work, the training symbol (TS) for MIMO channel estimation [9,36] were also compressed by ADX without being separately processed from payload, which affected the accuracy of MIMO channel estimation to some extent. Nevertheless, such high EVM performance was still achieved, highlighting the fidelity of the ADX. The experimental results agree well with floating-point simulation results.

 

Fig. 4. Experimental results of Case (I). (a) EVM vs. wireless SNR, with P=4 and P=2. Matlab floating-point simulation results are also plotted. (b) EVM of OFDM, SCFDM and F-OFDM vs. wireless SNR, with P=4. We also show by dashed line the EVM performance (wireless SNR at 50dB) without DAC, ADC or IF conversion.

Download Full Size | PDF

Toward (B)5G radio access, there are multiple candidate waveforms or radio formats besides OFDM, such as single-carrier FDM (SCFDM, or DFT-s-OFDM) [30], filtered OFDM (F-OFDM) [31] and filter-bank multi-carrier (FBMC) [32]. Hence, it would be interesting to investigate the compatibility of ADX to these different kinds of wireless formats. Here the EVM performance of OFDM, SCFDM, F-OFDM with the same ADX circuit were investigated. The signal parameters are given in Table 1. The experimental results (P=4) are shown in Fig. 4(b). Notably, negligible EVM performance difference was observed for all the tested signals. This indicates that the ADX is highly compatible with (B)5G radio formats. We also show by dashed line the EVM performance (approximately 1.2%) without DAC, ADC or IF interface, at wireless SNR=50dB. The small penalty could be attributed to the finite DAC/ADC resolution and the slight distortion during IF conversion. Nevertheless, the experimentally achieved EVM performance well meets the 1024QAM requirement. This highlights the fidelity of the ADX. It indicates that the designed ADX has the potential to handle multiple indoor scenarios/areas in (B)5G which may have different radio formats, without hardware modifications. In the following experiments, OFDM format was tested.

Next, the ADX-RoMMF link performance was investigated. The L-I-V curve of VCSEL is shown in Fig. 5(a). Figure 5(b) shows electrical SNR of the received OOK signal in Case (I) versus bias current in optical back-to-back (BtB) scenario. The signal quality is acceptable when the bias current is larger than 3mA. For the subsequent RoMMF transmission measurements, bias current of 6mA was used for Case (I). For Cases (II) and (III), 5mA was adopted. Figure 5(c) shows the electrical SNR of received OOK signal, with insets showing the respective eye diagrams. It can be seen that Case (I) which applied the most aggressive signal compression achieved the lowest FH rate, and the electrical SNR was the highest. In Case (II), the FH rate of 16Gb/s did not exceed the transceiver bandwidth. The signal quality was still acceptable, and no bit error was detected. However, the achievable reach would be limited by the modal dispersion of MMF. In Case (III) without signal compression, the FH rate was 30Gb/s, which was difficult to support even with the high-speed transceiver components in the experiment, and the eye opening was severely affected in optical BtB scenario.

 

Fig. 5. (a) L-I-V curve of VCSEL. (b) Electrical SNR of received OOK signal in Case (I) versus VCSEL bias current (optical BtB). (c) Electrical SNR of received OOK signal in Case (I), (II) and (III), after bias current optimization. Insets show respective eye diagrams.

Download Full Size | PDF

Subsequently, various MMF distances ranging from 0 to 1000m were tested. The wireless SNR was set at 50dB. Figure 6(a) shows the EVM performance versus MMF distance when P=4, i.e., 4 independent 5G NR-like signal streams were transmitted. In Case (III) (without compression), the FH data rate of 30Gb/s exceeded transceiver bandwidth, and thus even optical BtB transmission could not be supported. On the other hand, in Case (II), conventional time-domain-only compressor reduced the data rate to about 16-Gb/s, and thus the MMF reach was extended to about 200m regarding 1024QAM EVM requirement. Finally, in Case (I), with the real-time ADX compression, the link data rate was reduced to about 4Gb/s. Although the compression-induced signal quality degradation became larger, the benefit of bandwidth reduction dominates, and 850m MMF reach was supported for the CPRI-equivalent rate of 32.4-Gb/s. Moreover, considering the low bandwidth requirement, the cost of optoelectronic components can be lowered down. This highlights the attractiveness of the ADX-RoMMF scheme, which enables significantly more operation margin and larger in-building coverage. We also measured the case P=2, and the experimental results are shown in Fig. 6(b). In P=2 case, since the number of antennas in the RU was fixed to be 16, there was more spatial diversity gain. Therefore, lower EVM was observed for both Case (I) and (II) when MMF length=0. Still, the ADX provided substantial reach extension compared with conventional compressor.

 

Fig. 6. Measured EVM versus MMF length, in (a) P=4 case and (b) P=2 case.

Download Full Size | PDF

We also investigated the link bit error tolerance of the ADX-RoMMF scheme (i.e., Case(I)). The RoMMF link bandwidth reduced with increased MMF length, which caused inter-symbol interference (ISI) and bit errors. Figure 7(a) shows the ADX-RoMMF link bit error rate (BER) versus tested MMF distance. For the maximum reach 0f 850m satisfying 1024QAM EVM requirement of 2.5% shown in Fig. 6(a), the BER of 1.25e-6 can be tolerated. Figure 7(b) shows demodulated 1024QAM constellation diagrams or 4 signal streams at 750m MMF distance.

 

Fig. 7. (a) Link BER vs. MMF length in Case (I). (b) Demodulated 1024QAM constellation diagrams or 4 signal streams (P=4) at MMF distance of 750m.

Download Full Size | PDF

Finally, we used Vivado Integrated Logic Analyzer (ILA) to measure the latency induced by the real-time ADX processing. Figure 8(a) shows the latency of OFDM signal compression by ADX, which is 16 clock cycles or 131ns. Here “compression latency” means the delay between the first input signal sample and the first output compressed data. In addition, according to Vivado FPGA simulation, the ADX decompression latency at DU was 359ns or 44 clock cycles. In total, the one-way ADX latency overhead, including compression and decompression, was slightly less than 500ns. As reference, the indoor FH propagation delay is about 5ns/m. In addition, we measured the latency induced by ADX compression for SCFDM and F-OFDM formats as well. The results are shown in Figs. 8(b)–8(c). It was confirmed that from latency aspect, the ADX is also agnostic to OFDM, SCFDM and F-OFDM radio formats.

 

Fig. 8. ADX compression latencies for (a) OFDM, (b) SCFDM, (c) F-OFDM radio formats, measured by Vivado ILA.

Download Full Size | PDF

In this work, the solution path for the indoor (B)5G MMF FH is ADX, which focuses on 3GPP Option 8 (or “PHY-RF split”) and aims at transporting MIMO waveforms to the DU precisely without knowing/touching the information inside. On the other hand, low-PHY split such as Option 7-2 is another promising solution, which can also achieve ∼10-fold FH data rate reduction [42,43] by performing partial wireless (de)modulation based on the prior knowledge of signals sent from CU/DU. RU with ADX has the potential to support multiple radio formats with the same hardware, as demonstrated above. Detailed comparison of ADX and low-PHY split would be an interesting future work.

In the proof-of-concept experiments shown above, single frequency allocation (FA) case was demonstrated. On the other hand, the use of multiple FAs is a key to enhance the capacity of indoor 5G/B5G communication [44,45]. If multiple FAs are involved, after the RF front-end, we can utilize multiple parallel ADX each handling one FA, and multiplexing the ADX output data before fronthauling. Another interesting research direction would be designing intermediate-frequency ADX with high bandwidth which processes multiple FAs jointly.

4. Conclusion

We have presented an indoor (B)5G FH scheme for future C-RAN based on legacy MMF and real-time ADX, termed ADX-RoMMF. Experimental results have shown substantial MMF reach extension of the scheme up to 850m, while still maintaining signal quality suited for even 1024QAM. In addition, we have also confirmed multi-format compatibility and 500ns processing latency of the ADX. The presented scheme could promote to reuse installed MMF for future indoor RAN, and could also be useful in other applications with suitable network scale and data rates, such as campus RAN deployment.