Precise synchronization and contention resolution in next generation optical networks, beam steering, LIDAR and waveforming represent only a few of the possible applications of optical buffers [1]. In nearly all practical applications, the wavelength of the input information should be preserved, whereas the buffering element should be able to address any delay throughout its tuning range. Consequently, a true in-line optical delay is: (i) continuously tunable and (ii) wavelength transparent. Even if the second condition is rescinded, the delay value and the output wavelength should be independently controlled variables to be suitable for the applications listed in [1]. New parametric wavelength converters have enabled delay elements [2–5], culminating in a recent 1.83μs measurement. In contrast to the dual conversion architecture [4], which relies on conjugated dispersion self-compensation, a new scheme based on triple wavelength conversion was recently proposed and demonstrated [2,3]. By introducing the additional (third) wavelength conversion, the requirement on higher-order engineering is greatly reduced [2,3,5]. Indeed, the additional degree of freedom, granted by the third conversion step, allows for the effective cancellation of higher-order dispersion, and the combination of large delay range and bandwidth of the processed signal. The latter benefit, however, does not come free, as one must engineer three, rather than two wavelength conversions.

In recently published work [1], the authors claim the introduction of a new delay architecture and experimental measurement of a 1-μs tunable delay. The purpose of this comment is to point out that these claims are not supported by the material reported in [1]. Indeed, the delay scheme in [1] appears to be identical to the scheme proposed and demonstrated by Namiki [2,3]. Specifically, the concept described in [1] is a reduced version of the Namiki scheme that does not include higher-order dispersion compensation. This simplification was enabled by the use of a moderate (10Gbps) data rate and, contrary to the authors’ claim, cannot be extended to considerably higher data rates. The true difference introduced in [1] is the use of a silicon waveguide instead of a high confinement fiber. Its performance appears to be a main reason for the system results reported in [1]. It appears that the claim of a measured all-optical delay is not substantiated as [1] reports only a partial experimental result. Specifically, only two of the three required wavelength conversions are reported, reducing the measured architecture to a new, yet unclassified functionality: each delay setting corresponds to a different wavelength output. We believe that it was reasonable to use only two conversions to demonstrate the principle of the scheme [2,3]. However, as it is now well established, subsequent demonstrations [5], and particularly those aiming for record performance, should perform all three conversions to be relevant.

The experiment reported in [1] is, to the authors’ credit, an attempt to perform a complex system measurement. Unfortunately, the reported measurements fall short of the rigorous requirements outlined above and should not be trivialized. They are of a more fundamental nature inherent in physical properties of silicon and should not be rationalized by the statement that the addition of the third conversion stage “will have no impact on the dispersion management and the delay range.” This rather important assessment is in direct contradiction to (a) the measured penalty and (b) the means used to measure the performance of the partial (two-conversion) architecture. Indeed, the measured penalty of 4.5dB represents a considerable impairment in any system operating at 10Gbps, even if the reader does allow that “wavelength conversion based on the Si waveguide will result in a small power penalty (~ 1 dB) to the system”. If only two conversions contribute a 4.5dB penalty, a question remains about the total penalty accumulated after a third conversion with an efficiency of only −20dB [1]. Would one actually measure BER below 10⁻⁹ by completing the experiment? The last issue becomes even more relevant if one considers the fact that the authors chose to measure performance using a short (2²³-1), rather than a standard (2³¹-1) pattern length. This is in spite of well-known pattern-dependent effects expected in non-ideal converters such as silicon. Two-photon absorption and pump depletion effects are only two of many impairment mechanisms that a researcher in field would test before deciding to use a non-standard test-word length. Indeed, the risks of understating the system penalty by using a short, rather than standard (2³¹-1) test pattern in wavelength converters were quantified more than a decade ago [6]. The reported back-to-back measurement (−19dB SNR at a BER of 10⁻⁹) is significantly below the standard reference level of about −34dBm [7], suggesting a large penalty which might mask other effects in this experiment.

While [1] simply does not provide enough experimental details, we believe that most of the discrepancy between the measured and claimed results originates in the performance of the conversion element and challenges stemming from the −20dB conversion efficiency and associated impairments. Indeed, there is a reasonable hope that high-performance silicon mixers could be created in 1550nm band, in spite of the basic limitations related to carrier generation in this band. While promising, the present silicon conversion technology has not yet demonstrated the performance to support the authors’ claims, and is not yet competitive with silica- [8] or chalcogenide-based [9] wavelength converters. That being said, it is clear that the authors undertook rather difficult system experiments with silicon converter elements and that such waveguides might have a potential as nonlinear platform for signal processing in the 1550nm band. Consequently, we eagerly await a report of rigorous measurements supporting the microsecond-delay claim with silicon mixers.