Edge is Cool ($) but Cloud is Hot ($$$)

This is becoming more real. Here Coolness of Edge & Hotness of Cloud refers to the tax in terms of Cost we need to incur to get the capabilities of Smart Edge & Smart Cloud.

Let me expand Tax part here with an example. I’ve taken variable cost into account here given the scale we are talking about & so let’s focus on effective contribution margin we get.

Sales = From SaaS service provided by using Smart Cloud & Edge

Variable Cost = Cost of (Network bandwidth + Cloud Storage + Cloud Kubernetes Cluster + Serverless compute + Smart Cloud Service OR Smart Edge Service)

Contribution Margin = Sales – Variable Cost

Cloud Tax = Extra $ amount to be paid for using Smart Cloud service(s)
Edge Tax = Extra $ amount to be paid for using Smart Edge service(s)

As you can see here, every business would like to optimize (of course minimize) on Cost part to maximize on Contribution Margin J However, it turns out that using the Smart Cloud Services is costly value proposition than using Smart Edge services. I’m referring to the increased cost due to Smart Cloud Services as Tax here.

Smart Services referred here are the ones that provide the Machine Learning (ML), Deep Learning (DL), Reinforcement Learning (RL) capabilities to us. We use them to solve variety of different use-cases for Edge. Understood that these services would need extra compute power (read GPU) to train our models (if we are building or tweaking cloud provided ones) OR run pre-trained models (if we are using cloud based services as is).

However, if you think about it, running such models on Cloud versus on Edge makes a lot of difference. On the Edge side, we have more control in terms of running the specific (pre-trained) ML/DL/RL frameworks (including Lite versions) as well as algorithms on a relatively commodity hardware – Yes Hardware is becoming COOL again! – thereby limiting the Tax for running the Smart Edge Service in control (How to run the model on Edge efficiently is another big topic; so let’s park it for now). This is not the case with Smart Cloud service usage. We do not have much control since we DO NOT own it end-to-end. After all it is “Managed”.

Let me take one example here wrt AWS ML/DL services. Let’s say we want to process Video stream coming at 30 fps (frames per seconds) for face detection use-case. So we can make use of Amazon Rekognition service for same. Video data generated can be moderate to huge depending on Edge use-case in picture. Here are some stats for same:

Sr # Assumptions & Facts Relative #
1 Video Stream Rate 30 fps
2 Max # of Frames/Images generated per Day ~2.5 millions
3 Max # of Frames/Images per Month ~75 millions
4 Amazon Rekognition Price per 1,000 Images Processed:

First 1 million images processed* per month

Next 9 million images processed* per month

Next 90 million images processed* per month

Over 100 million images processed* per month

*Each API that accepts 1 or more input images, counts as 1 image processed.





5 Approx cost for processing Frames/Images for 1 day ~$1900/day

If you think about it, for the Tax part, we have just discussed the cost angle here wrt Smart service usage. There are also additional considerations (some of which we have discussed in previous blogs) regarding real-time processing at the Edge to avoid network latency thereby improving response (QoS), geo-location of data processing & storage from compliance point of view – which also have significance.

There could be ways to make impact of such a tax less of a burden in the Cloud by discounting. For e.g. using ML/DL Inference Acceleration to run pre-trained models; AWS just announced new service – Amazon Elastic Inference https://aws.amazon.com/machine-learning/elastic-inference/ couple of days ago which needs to be validated & also made available across more regions. But now, it turns out that Edge is really Cool ($), giving complete end-to-end control while Cloud is still Hot ($$$).

Intelligent IOs,
IOPhysics Systems