Misunderstanding Moore’s Law
Tom Evslin, the former head of AT&T’s dial-up Internet services unit in the 1990s, has written a blog post for CircleID that mistates Moore’s Law’s application to networking. There’s a lot of wrong stuff on the Internet, of course, but this piece deserves particular attention because it contains a common pair of errors that are very widespread in today’s debates around networking and network policy. When advocates get the technical facts wrong, it’s likely that their policy wishes will be wrong as well.
Tom’s first mistake is to overstate the role of Moore’s Law in determining the cost of networks:
How much data you can transmit for a given price in a given period of time depends on how much of it can be squeezed into a copper, fiber, radio wave or other kind of pipe. The cost of data squeezing depends on the cost of the computer power which does the squeezing. So it shouldn’t be any surprise that data communication costs follow Moore’s law down further and further
This statement is partially true, but not completely true. While it’s true that some portions of network costs are affected by Moore’s Law, others are not. Some aspects of networking decline in price faster than Moore’s Law, and some decline more slowly. There are two reasons for this:
- Moore’s Law pertains to mass-market semiconductors, not to all chips. Prices of mass-market semiconductors decline because chip manufacturers invest in new fabs that are capable of squeezing more transistors onto a chip, and the factories are followed by redesigned mass-market chips. Chips that are sold to specialized, niche markets don’t follow this curve because they remain in the lower-tech fabs. The Content-Addressable Memories (CAM) used in edge IP routers for address lookup are in the niche category, for example.
- Not all aspects of a network are digital, and not all the digital parts are mass-market. Trenches and labor, for example, are necessary components of networking that don’t benefit from process improvements in chip manufacture at all. While the capacity of a given cable plant, the copper loop that carries DSL, for example, increases over time due to improvements such as VDSL, VDSL+, and VDSL2, replacing the copper with fiber isn’t an exercise in Moore’s Law inevitability. Similarly, the cost of mobile broadband is related to advances in batteries that also fall outside of the digital realm.
So Tom should have said that the cost of networking is partially affected by Moore’s Law, but only partially; therefore, we should not expect a price reduction of 50% every 18 months for bandwidth. Martin Cooper, the inventor of the cell phone, has postulated “Cooper’s Law” to the effect that bits per Hz doubles every 30 months, a much slower curve, and another formulation says that fiber capacity doubles every eight months, much faster than Moore’s Law would predict.
Tom then goes to claim that Skype is free because it operates independently of the “toll booths” on the legacy public switched telephone network:
Skype video calls can be free because they don’t use any of the voice infrastructure which is still under the control of the descendants of the old monopolies.
This is a more serious mistake than the Moore’s Law error. In the first place, Skype video calls aren’t “free” to Skype, even if they are “free” to the consumer (once we take the consumer’s broadband costs out of the picture.) One-to-one video Skype is a “loss leader” than Skype has chosen to give away in order to encourage more people to use Skype’s paid services, which primarily consist of paid interconnection to the PSTN (Skype Out) and paid multi-party video calling (Skype for Business.)
Skype is a business with a staff to pay and an investment in equipment, and as such it needs a revenue stream to remain in business, enhancing its code, rolling out new releases, and supporting its network. Skype’s principle revenue source is PSTN interconnection, so the reality of video Skype’s economics is just the opposite of Tom’s claim: Video Skype is free (in the one-to-one version) not because it bypasses the PSTN’s toll booths but because Skype earns revenue from its many users who of those very toll booths.
It’s generally the case that many services can be offered over the Internet without end-user subscription fees or usage fees because of the fact that end users pay enough to cover the bulk of the generic application’s communication expenses and some other source of revenue – generally ads – covers the rest. The only effect of Moore’s Law in the overall scenario is to lower the costs of the application components that directly pertain to data processing on the server side, preference mapping for ads, and the user interface functions of the device that runs the browser or the app on the user side. There are still substantial costs to running a business that are analog (like people) or too specialized to benefit from process improvements in the manufacture of mass market chips. I suspect Tom knows this, and is omitting them from his discussion for dramatic effect.
Tom Evslin has some serious misconceptions about wireless networks as well. He claimed that part 15 rule white space networks will make LTE obsolete. He claimed that a 4W ERP white space radio/tower has an “effective radius of up to 30 miles”. http://www.circleid.com/posts/20081126_fcc_white_space_regulations/
He also argues that wireless networks don’t need schedulers.
I’d also point out that Skype actually *does* charge consumers by charging hardware makers for access to their APIs. If you buy a webcam or mic, there’s a good possibility that part of the price you paid goes to Skype.
Skype also required their hardware licensees to lock out other VoIP protocols, including the IETF standard SIP. At the VON conference of 2008, I interviewed one of their hardware partners why they couldn’t offer me a dual-mode SIP/Skype Wi-Fi or wired handset. They explained me Skype’s licensing terms prohibited any VoIP alternatives.
Wireless systems with miles of propagation need schedulers because CSMA doesn’t work when edge-to-edge propagation time exceeds a few microseconds. Evslin doesn’t realize that the “command and control” aspect of WiMax is a consequence of that fact rather than the “bellhead pathology” he imagines.
This discussion is in reference to channels, wired and wireless. Faster processing also benefits source coding for video and audio, which reduces the number of bits the channel needs to carry.
I’ll also defend scheduling. It’s a source of gain in cellular systems. If there’s one user in a sector, experiencing a typical fading channel, there are times when that user can’t be served. One can calculate an average sector throughput. Add a second user, and the sector can serve that user during some of the time the first user can’t be reached. Average sector throughput is thus increased. With several dozen users employing a mix of traffic types there is significant gain from this diversity.
There’s a limit to how far you can compress video. Most of the metrics for video compression are based on “perceived” quality and not on the actual loss of data. It’s just a question of how well the lost data is masked with approximations and how well the artifacts are hidden.
Steve makes a good point.