| 摘要 |
In a preferred embodiment, the invention uses an 8-to-1 data serialization circuit in the transmitter to convert 80-bit parallel 200 MHz data to 10-bit parallel 1.6 Mb/s date. On the receiver side, data are captured using a forwarded clock and de-serialized. A single global DLL generates 16 master phases without reference to the word boundaries of data being transmitted. These 16 unreferenced phases are input to a phase rotator that, via a series of calibration steps, maps the unreferenced phases into named phases, and in doing so references the phases to the word boundary of the data being transmitted over the slowest data line of the parallel channel. The named phases are then input to a data interpolator in each receiver, which generates 16 local phases. The 16 local phases correspond to the data-bit centers and data-bit edges for each of the 8 bits transferred per major channel clock period. In a bit-centering calibration step, a training pattern is evaluated by each receiver and each data interpolator dynamically adjusts a delay applied to the 16 local phases to establish the local center-data phases in the center of the bits received by the corresponding receiver. In an additional calibration step, on a per-wire basis, 8 contiguous bits are selected as the data outputs from a 10-bit window. The local center-data phases are used to serialize and de-serialize the channel data for the receiver. The present invention optimizes clock timing for each channel bit, thus providing the benefits of a dedicated DLL per channel bit, without the associated cost.
|
