5G wireless systems are envisioned to provide ultra-high data rates, increased robustness and low latency for new applications such as machine-to-machine (M2M) and car-to-car (C2C) communications. This paper considers high-speed visible light communication (VLC) as a candidate serving technology to fulfill this set of requirements. Synchronization as the first step towards reliable data detection is studied here thoroughly aiming at high robustness and low complexity. Three schemes are reviewed, namely Schmidl-Cox and Park autocorrelation as well as a cross-correlation based approach, all using a predefined preamble. All schemes are examined concerning their robustness against multipath propagation and their precision in time estimation versus the preamble length. It is found that the simple Schmidl-Cox scheme, originally designed for complex-valued radio waveforms, can be applied also for real-valued optical wireless waveforms. Based on these investigations, an optimized preamble structure is finally proposed. Preamble boosting and fine synchronization can also be applied to reach a similar performance like optimal cross-correlation at reduced complexity. We study the fundamental trade-off between the preamble duration and the minimum signal-to-noise ratio (SNR) required for reliable synchronization and observe that ultra-low latency needs sufficient SNR.