UCLA

As data centers are expected to manage the increasing demands in bandwidth, processing power and storage requirements, connectivity issues between blades/racks present a whole new set of challenges in maintaining a stable infrastructure. While data centers may grow to occupy thousands of square feet, current passive copper interconnects pose a real limitation with a run length of 10 meters at 10Gbps per wire pair. Optical fiber can extend the interconnection length from 10 meters to 100 meters, but the large power requirements and expensive opto-electric modules prove to be too uneconomical for practical application. As a compromise, through the use of the Infiniband standard, a 12Gbps cable link can be achieved that would extend the range of copper interconnects beyond the 100 meter threshold.

The proposed link leverages synchronous clock forwarding on one available data channel that improves jitter tracking, while greatly simplifying the design of the receiver and timing recovery circuits. Only a phase de-skewing is required at the receive side to retrieve the clock-data relationship. In the cable link architecture, the 12 Gbps data is repeated in 8 meter sections with clocking forwarding on a dedicated channel. Then the forwarded clock is dropped off every data repeating stage in order to be multiplied to half the data rate and be used to strobe the incoming data. The longer the quality of clock forwarding is maintained, the cleaner the data strobed at each repeater and the longer the cable can be extended. At each repeater, the clock resets the jitter accumulated from the previous repeater, allowing for data transmission with as much jitter as in the strobing clock.

Determining a fine balance in forward clock frequency is crucial in defining jitter performance of the cable link. Frequency beyond the cable bandwidth results in large attenuation of clock amplitude creating more noise and jitter accumulation along clock repeater. On the other hand, frequency well below the cable bandwidth will increase jitter accumulation time and will degrade jitter performance inside the clock multiplier. The trade-off between low frequency clock jitter accumulation in the Clock Multiplication Unit (CMU) and the high frequency jitter accumulation along the clock repeaters is one of the defining aspects of optimizing the active copper link.

To further reduce the clock jitter accumulation across repeaters, phase interpolation between the input clock and the divided output of the CMU is used to generate the forward clock for the next repeater stage. The addition of the phase interpolator has negligible power/area cost, dramatically reduces jitter accumulation, and adds another degree of flexibility in choosing the forwarded clock to reduce the total accumulated jitter. With the proper choice of forward clock frequency, application of the FIR filtering technique and a high performance CMU, a total run length of 115 meters is achieved at 12Gbps data rate.