From the 1950s through about 1993, all computers were processor-bound. The I/O interconnects could feed more data than the processors could handle. After 1993, with the availability of multi-GigaHertz and multicore processors, designers rapidly became I/O-bound. The processors could handle more data than the interconnects could provide. As processor speeds and capabilities increased exponentially, I/O bandwidth increased incrementally with PCI, PCI Express, InfiniBand, RapidIO, and Ethernet. Today, designers are still horribly I/O-bound because the semiconductor bandwidths, both at the processor and interconnect chip level, are increasingly faster than copper connector bandwidths.
Copper-based backplane connectors are the single most nefarious impediment to dramatic increases in computing power. As the frequency of the copper interconnects goes up, designers run into all types of transmission line problems including discontinuities, crosstalk, noise, reflections, and impedance mismatches. Once reaching approximately 10 G on copper, the gods of physics display their tremendous sense of humor.
Entering RF-absorbing materials
A number of benign applications needing increased computing power are ruled by commodity buyers, such as telecom. They want higher I/O bandwidths, but are not willing to pay much for it. So the connector industry is rallying around this low-cost commodity demand by mixing RF-absorbing lossy metallic particles into the plastic resins of connector shells and wafers. These materials are similar to the particles used in radar-absorbing paints applied to stealth military aircraft, missiles, and Unmanned Aerial Vehicles (UAVs).
The energy levels of the crosstalk and noise inside the connector are very low-power signals, so the shells and wafers do not need to be grounded like the all-metal Faraday shields presently used on medium-speed (<10 G) connectors. The connector shells just dissipate the RF energy by heating up a degree or so. This concept does abate the noise and crosstalk levels between signal pairs inside the connector, but it does not eliminate it. It is just another incremental improvement, and it is still seriously I/O-bound.
It is all about the pin field
The bigger problem holding designers back from increasing bandwidth is the connector pin field, the footprint of all the copper connector pins on the circuit board. In the VITA Standards Organization (VSO), in the backplane standards for critical embedded systems, the rule of thumb is that every single-ended signal must have a ground pin for every signal pin to guarantee reliable operation (a 1:1 ratio). For differential signals, every pair of signal lines must have a ground pin (a 2:1 ratio) to guarantee reliable operation. As one can see, there is a large number of pins (signal and ground) on the backplane and on the daughtercard to which to attach a pin-dense connector. This high pin density creates more noise, crosstalk, and routing problems for designers.
However, the real culprits, those creating the discontinuities and signal reflections, are the stubs: the via-holes on the PCB where the connector pins are inserted. The only way to solve these problems is to use micro-vias (very small holes in the PCB) and very thin connector pins (micro-pins). That will reduce the size of the stub and diminish the reflections and discontinuities. The tiny drill bits used for the holes in the PCB tend to walk around in the hole (they are also very thin and very flexible under pressure), making the hole (the stub) larger than needed for the pin (defeating the purpose of making the via-hole as small as possible). The drill bits wear out quickly since they have a very small cutting surface. There are thousands of these via-holes on a backplane, so the bits must be changed regularly. Yet again, this is another incremental improvement, with a lot of process problems, and designers remain I/O-bound.
Perfect alternative for telecom
These new “stealth” connectors, with tiny brittle pins, will be implemented in telecom equipment built to the New Equipment Building Systems/Network Equipment Building Standards (NEBS) specifications. So all these incremental improvements will fit right in with telecom’s goal of not paying much for them. Those small brittle pins would never make it through the shock and vibration testing needed for military weapons platforms, avionics, ground vehicles, and space applications.
The real alternative
The only cogent alternative to completely break the I/O-bound annoyance is optical interconnects. Optical links eliminate all the physics problems associated with copper; the same optical connections running at 10 G can also run at 40 G without a backplane replacement, and optical connections do not radiate any RF energy. As long as copper is used for interconnects, designers will remain I/O-bound and the gods of physics will be laughing at the industry.