Optical data links could be in your (very near) future
BlogMarch 03, 2015
Ray Alderman
VITA Standards Organization
This article is not about the technical aspects of optical data links like light frequencies, single-mode versus multi-mode fibers, optical waveguides, PIN diodes, laser diodes, light beam losses over distance, connectors for optical links, optical engines, etc. There are plenty of website articles covering those topics, so let’s focus on why optical could be in your future in radar, sonar, signal intelligence, and medical applications. We moved from buses to differential serial fabrics because we encountered the bandwidth limit of single-ended signals on copper. Very soon, we will hit the bandwidth limit of serial links on copper and be forced to move to optical. How close are we? Alderman’s Laws of data transmission define that for backplane-based embedded computers:
Law 1: When (f+P) is increasing, then (s/n) is decreasing
On copper traces, as the frequency (f) of a link increases and the electrical signaling protocol (P) becomes more complex, the signal-to-noise ratio (s/n) declines dramatically. The higher the frequency and the more complex the signaling protocol on copper, the more the signal looks like noise.
Law 2: 2Bf=d/2
Every time the frequency of the signals going through a copper connector and along the copper backplane traces doubles (2Bf), the distance those signals can run declines by 50 percent (d/2). Otherwise, they look like noise again.
Law 3: 2Nf=B/2
Every time the network link frequency doubles (2Nf), the demand for copper-based backplanes and their associated connectors declines by 50 percent (B/2). When the network outperforms the backplane, there’s no need for a backplane.
Law 4: if Nf > Bf, then 2Nf = B/2
If the network link frequency (Nf) is greater than the backplane link frequency between the boards (Bf), then backplanes lose their primary performance advantage. Today, we are running multiple links on copper traces to each board to keep up with the network link bandwidth. When we run out of connector pins on the backplane for those additional copper links, Law 3 is invoked.
Understand that backplane-based computers are centralized systems. Each board in the rack is connected to the others through copper traces on a common backplane, and all the boards share a central power supply and chassis. Backplanes have two basic benefits: high-performance data links and modularity. Modularity, in turn, has two sub-benefits: maintainability and upgradeability. When the bandwidth of the network cable or optical fiber is greater than the bandwidth of the copper traces between the boards on the backplane, the boards in the rack will be broken out into separate boxes, into a distributed system, and connected together with the network cable (Law 3). That’s higher performance at a lower cost, since the rack, the big power supply, and the backplane are more expensive than the power supply and packaging for a single board (Law 4).
Also understand that up to 1994, all computers were CPU-bound: the data links could deliver more data than the CPU could process. After 1994, and the advent of GHz multicore processors, we have been I/O-bound: the CPU can process more data than the links can deliver. And, that situation is getting worse. Just look at the latest NVIDIA Tegra X-1 GPU. It has 256 GPU cores, eight 64-bit CPU cores (ARM), can handle 4K video at 60 frames per second (fps), and process data at the rate of one Teraflop (10^12 floating point operations per second (FLOPS)). You can feed data to this monster on copper connections if you only run them a few inches (Law 2). But, if you are bringing in data from remotely-mounted cameras on a Global Hawk, Predator, or a Reaper, you’ll have to go optical on those links or the cores will be data starved (Law 1).
If you are making products for the industrial markets, those users can run most of their applications on an Apple Watch CPU, so they don’t need the bandwidth. Industrial apps are register-oriented sequencers (“bingo” machines). If you are making products for telecom, they will stay with copper because they cannot afford optical. Telecom machines route data, they don’t process it. Industry leaders in telecom or industrial board sales are just the lepers with the most fingers. If you are anticipating making products for the IoT (Internet of Trash) market, they neither need the bandwidth nor care about optical.
But, if you are making radar, sonar, signal intelligence, or medical machines (streaming data processing), you have some studying to do.
VITA Standards Organization (VSO)
