On the streaming market, moving from the “Moore’s Law” to “Bezos Law”

Moore’s Law (from Gordon Moore, co-founder of Intel) is a 50+ year old observation that “the number of transistors in a dense Integrated Circuit doubles every two years” (according to Wikipedia). This observation, made in 1965, is still pretty much valid today, as depicted by the figure below:

Moore's law

Although purists will stir up scandal, one could rephrase that law as “The processing power of a computer will double every two years while staying at the same price”. Any doubts? Evidence is that the price of a laptop has been around $1000 for the last decade, while the CPU capacity of these laptops followed the Moore’s law consistently.

The Moore’s Law in the cloud

As we all know, the Cloud is becoming more and more prevalent in all the industries. Streaming is late in embracing the cloud (although modern streaming technlogies are changing that), but the transition is unavoidable. Based on Moore’s Law, a quick assumption could be that the CPU compute cost will drop by 50% every two years (equivalent to 30% every year). Let’s see if that works!

Benchmarking the Moore’s Law for Live Streaming

For this exercise, we’ll consider AWS cloud as a reference, for two reasons:

  • It’s quite easy to find old articles about the compute instance pricing for AWS, as they were the first on the market.
  • It’s still possible to launch 8 years old instances in AWS today, making it easy to compare performances. You can still launch c3 (released in 2013, based on SandyBridge architecture) and c4 (released in 2015, based on Haswell architecture) instances in most of the AWS regions.

While prices were simply gathered on the Internet, we used an open-source project based on ffmpeg (a widely used open-source transcoder software) to benchmark the performances. For the records:

  • The performance is expressed in “number of HD transcodings per hour”,
  • The cost is the instance on-demand price per hour, in USD, in the Oregon (us-west-1) datacenter.

And the results are …

Price and performance Evolution in AWS since 2013

To make this graph more readable, let’s “normalize” it, i.e. get the cost for transcodings 10 HD over the years:

normalized price for transcoding 10X HD in AWS

As we can see, the cost reduction has been systematic since 2013 (and was probably before that, but we couldn’t get any data). When doing the maths, the price drop is approximately 15% per year.

Here comes the “Bezos Law”!

Let’s look at these 15% from another angle, and let’s try to adapt the Moore’s Law to Live streaming in the cloud:

“ The price of processing your live streams in the cloud will be divided by two every four years” (because (1-0.15)⁴ ~ 0.5).

One could wonder why moving to the cloud apparently makes the Moore’s Law less “aggressive”. Well, it doesn’t! But keep in mind that when buying compute capacity at AWS, you don’t only buy a CPU: you buy the rack, the electricity, the real estate, and the human beings running these instances for you. There is no such thing as a Moore’s Law for these elements, and they represent a large portion of the overall compute cost (probably more than the CPU itself), “slowing down” the Moore’s Law.

If you remember the pizza analogy we used in a previous article: it’s not because the price of the flour drops by 50% that your favorite pizza price will drop by the same amount.

Wrap-up: how to fix your 5 year-streaming TCO

We often see on-premise vs cloud TCO comparisons that consider that the cloud price will remain constant over the TCO duration. But after five years, the price of the compute cost in the cloud will have dropped by more than 55%! To leverage the Bezos Law, you just need to make sure that you use a technology that is not tied to a given instance type. This is of course a native benefit of Just-In-Time functions, which are the basis of all the Quortex workflows. With Just-In-Time functions, you will be able to use newer instances as soon as they are released by the cloud provider, hence reducing your infrastructure cost by 15% a year.

Hence, it’s time to fix the graph that we computed in a previous article with the “Bezos Law”. A very interesting finding is that the cloud cost and the on-premise cost follow the same trend after year 3 … Hence, they will never cross and the cloud will remain more efficient than on-premise, whatever the amortization duration is!

on premise cost versus just in time cloud functions

Quortex

Société par actions simplifiée au capital de 101.242€
1137A Avenue des Champs Blancs
35510 Cesson Sévigné
FRANCE
RCS Rennes 842 176 141
email: info@quortex.io

Publication Director: Marc Baillavoine