Options emerge for 10Gbps chip-to-chip interfaces

Engineers have been rapidly increasing chip-to-chip I/O speeds in an effort to keep pace with the bandwidth needs of increasingly integrated silicon. Consequently, a variety of parallel and serial options at speeds up to 10Gbps are becoming available that a designer would do well to evaluate carefully.

When selecting a chip-to-chip interface, factors that should be considered include size, power, number of required package signal balls and latency. However, when comparing 10Gbps Serdes, the 10Gb attachment unit interface (XAUI), and SPI-4 in dynamic mode, a 10Gbps Serdes will have the advantage on size, power and signal pins.

With respect to power consumption, a 10Gbps Serdes can be about two-thirds that of a XAUI and about one-third that of an SPI-4 interface. The physical size of a 10Gbps Serdes core can be about half and one-eighth that of XAUI and SPI-4 interfaces respectively.

The number of signal pins can vary depending on the package, but in general, the interface with fewer high-speed signal I/Os will have the advantage. For example, a 10Gbps Serdes uses four high-speed signal lines while XAUI uses 16 and SPI-4 uses 72.

Interfaces with more data lines generally provide links with lower latency at the PHY. This is because, at a high level, as you transmit and receive at a higher speed over a single line, more serialization and deserialization is being performed on the data before it is transmitted and after it is received.

A 10Gbps Serdes, for example, will perform 64-to-1 or 32-to-1 serialization and deserialization. Each lane of a XAUI, on the other hand, will only need to perform 10-to-1 serialization and deserialization. So, for applications where link latency is critical, SPI-4 type PHY or a XAUI will generally provide better performance.

Depending on the application, other items to consider when selecting an interface include the capability to be backward compatible with legacy interfaces, packaging options and channel characteristics.

To dig deeper into the high-speed interconnect issue, let us take a look at an example of a 10Gbps interface and how it could be implemented in an ASIC.

In this example, the ASIC logic transmits 32bit or 64bit parallel data. Depending on the application, this data can be 64bytes/66bytes encoded or SONET scrambled. The data is serialized and sent using an non-return to zero signal over a single differential pair at 9.95Gbps to 11.1Gbps.

On the receive side, the differential serial data stream is received, the clock is recovered and the data is deserialized. The data is then provided to the ASIC logic at 32bits or 64bits wide with respect to the recovered clock. The core employs on-chip termination (100 ohms, differential) and for most applications transmits with a nominal signal amplitude of 500mV differential, peak-to-peak.

Such a 10Gbps Serdes core may be optimized for low power and small size by targeting applications which require up to approximately 20cm of FR4-type PCB and a single connector. These applications would include chip-to-chip applications where both chips are on the same board, or when one chip is on a daughter card.

By designing the Serdes core so that it supports the 10Gb Small Form Factor Pluggable Multi Source Agreement (XFP MSA), it could then also be used to interface with XFP optical modules. These modules do not perform serialization and deserialization, and have significantly smaller form factors than modules that employ parallel interfaces.

Rather than pushing 10Gbps over a single differential pair, an alternative is to use XAUI. The XAUI employs four Serdes, or lanes, each transmitting and receiving differential 8bytes/10bytes-encoded data at 3.125Gbps. The effective bandwidth after the overhead for the 8bytes/10bytes coding is removed is 10Gbps.

XAUIs may be used to transfer data between chips, over a backplane or to an optical module such as a Xenpak or XPAK (two 10Gb MSAs). When the XAUI is integrated into an ASIC, it provides 10-to-1 serialization and deserialization and interfaces to an XGXS (10GbE extended sublayer) block that performs 8bytes/10bytes encoding/decoding, lane alignment, character substitution and resynchronization of the received data stream to the local clock.

Another alternative for 10Gbps chip-to-chip links is to use LVDS I/Os consisting of multiple data lines. SPI-4 interfaces are examples of such links. They feature 16 LVDS transmit data lines, one transmit control line, 16 receive data lines, one receive control line, and separate transmit and receive LVDS clocks which are forwarded with the data in each direction. In a typical application, each data line operates at a minimum of 622Mbps, with 800Mbps and above being common. The clocks run at half the baud rate—for example, 311MHz for data-line operation of 622Mbps.

There are two primary operating modes for the electrical interface in SPI-4 links. In static mode, which is for lower data rates, such as 800Mbps and below, the skew between data lines and between clock and data must be carefully managed so that the clock may latch in the data on the receive side. Dynamic mode requires a more sophisticated receiver that determines the optimum timing for sampling each data line. In this mode there are no restrictions on clock-to-data skew and on-chip de-skew logic can resolve multiple bit times of skew between data lines.

Dynamic mode uses an initialization sequence to allow the receiver sampling circuitry to determine the optimal timing for latching each data line. With no restriction on clock-to-data skew, relaxed requirements for data-to-data skew and higher speed capability, dynamic mode has proved to be a popular interface. For both modes, the interface generally uses on-chip 8-to-1 Serdes for each data line.

- Rich Hovey

Manager, CoreWare Product Management Group

LSI Logic Corp.