Simple can be harder than complex: you have to work hard to get your thinking clean to make it simple. But it's worth it in the end because once you get there, you can move mountains. - Steve Jobs

We offer Sparrosense AI Supervisor which helps manufacturing firms better monitor their shop-floors by leveraging the power of AI Video Analytics. For context check out this blog and this video clip.

In our experience, while taking Sparrosense to manufacturers, a frequent misconception comes to the fore that hardware requirements are complex, expensive and this often makes them sceptical of exploring its full potential.

Here is a rough guide based on years of trials & errors on how to think about AI hardware for deploying Sparrosense which sets the record straight and establishes that – it’s simple and easy to get a handle on.

Quick disclaimer – this post is mainly around Sparrosense AI Supervisor, and may have hints for AI hardware for similar applications but should in no way be expected to be the right hardware for a wide variety of AI applications. For that, you may start from this blogpost.

Architecture:

The architecture that we have zoned onto is simple:

CCTVs are connected to NVRs where the feed is stored. AI GPU server, connected to Sparrosense AI cloud server via the internet, is placed in the local server.

This architecture allows for heavy lifting of Video analytics to happen on the edge server while leveraging advantages of the cloud for accessibility and flexibility.

Let’s go through each of the hardware choices one by one:

CCTVs & NVRs

The biggest misconception is that AI Video Analytics requires specialised & expensive CCTVs. On the contrary, we prefer to use simple and cheap CCTVs which mostly exist anyways on the shop floors for safety & security purposes.

• They are cheap and hence if they get damaged, they can easily be replaced.
• Multiple CCTVs can be installed ensuring that no process gets missed.
• Any CCTV brand works as all brands have the same protocols
• Easy to procure as they can be purchased from the existing vendors that provide CCTVs for security

The biggest advantage of using an existing CCTV setup is that it saves time in procurement. The only additional work that is required is to orient the CCTVs in a way that the required processes are clearly visible – and our team shares some guidelines for the same.

There are often concerns that the environment has dust and smoke, which may reduce the visibility – the standard solution is actually to wipe the CCTV clean every 15 days.

The solution to another big worry of the high temperatures is to place the CCTV sufficiently far away from the source of the heat and use a higher resolution CCTV if required.

More complicated solutions such as cooling jackets, air curtains, etc. are mostly not required.

NVRs

We prefer to use existing NVRs if feasible. Otherwise, a regular NVR 8 to 16 channel is needed as per the number of CCTVs. All the videos are stored only on the NVRs and not on any Sparrosense system.

The required frames are requested from the NVR as required by the AI edge server.

Note: If you are a CCTV solutions provider and wish to be empanelled for Sparrosense – please email Shivangi at shivangi.bhardwaj@sparrosense.com

AI edge server

The AI edge server is where the bulk of the AI work happens. It is essentially just a regular server with a GPU.

If Sparrosense is deployed just to monitor 2-3 CCTVs, then a simple Consumer-grade GPU PC is enough which costs just about INR 1.8L.

Depending upon the number of CCTVs and complexity of the Video Analytics algorithms required we may need server-grade hardware. The standard specs are as follows:

• CPU: 2 X Intel Xeon E5-2690 v3 (12 Cores, 24 Threads)

• GPU: 2 X Nvidia GeForce RTX 2080Ti 11GB

• Motherboard: Intel C621 Chipset

• RAM: 64GB (32GB*2) ECC Reg 2400 MHz DDR4 RAM

• SSD Hard Disk: 512GB SAMSUNG 860 EVO SATA SSD

• Hard Disk: 4TB Enterprise Edition 7200RPM SATA HDD 2U Server Chassis

• Power Supply: 1+1 1600 Watts Redundant Platinum Certified

The total cost of the system is around 4.5 Lakhs.

In theory, it is possible to add more GPUs to the same server, but in our experience, it is often better to have a greater number of these servers, in case more CCTVs are required.

Note: If you are a CCTV solutions provider and wish to be empanelled for Sparrosense – please email Shivangi at shivangi.bhardwaj@sparrosense.com

Bandwidth & Network

Another common misconception is that very high bandwidth is required for AI video Analytics. But, in truth, as no videos are uploaded to the cloud and all processing is happening on the edge server, and only time stamps are sent from the local server to the cloud, the bandwidth requirement is very low.

Sparrosense works perfectly well on bandwidths as low as 2mbps.

Availability of static IP greatly helps and there is a minor requirement of the opening of required ports 554 and 8000 to enable Sparrosense cloud to easily communicate with Sparrosense AI Server.

All communication b/w AI Server and Cloud server is fully encrypted ensuring that all data remains safe and protected.

AI cloud server

Sparrosense leverages AWS and Azure cloud instances to host Sparrosense AI application. We use standard hardware and all cybersecurity protocols designed for low latency, maximum security and rapid scalability.

Sparrosense cloud application sends alerts on Whatsapp and powers various dashboards & reports.

What’s next in AI Video Analytics hardware

Internet is filled with news of the latest advances and often we get questions regarding what's next in AI Video Analytics hardware:

• Smart Cameras: In future, processing of videos might move to smart cameras which have AI processor built right inside the Camera itself. But, we don’t recommend them as on the manufacturing floor, there is a high chance of damaging cameras. The other issue is that AI GPU servers are far more flexible, and in this environment where AI hardware is rapidly advancing, it is economical and practical to keep options for future upgrades open.
• Processing on the cloud: Cloud infrastructure companies are working on reducing the costs of processing AI videos on the cloud and in the future, it may become entirely possible but currently, it remains prohibitively expensive. There are two more issues with cloud processing - Firstly, it is often not feasible to upload videos from NVR to cloud due to bandwidth constraints. Secondly, this may increase latency which may not be desirable for many use cases.
• Storage on the cloud: Storage on the cloud suffers from the same challenges of bandwidth constraints and the cost of video storage on the cloud.
• Alternates to GPUs – ASIC, TPUs: Although TPUs and other specialised hardware have become significantly advanced, but GPUs remain a battle-tested and a cost-effective option. In future, this may change depending upon how AI hardware evolves.

Hope this was helpful in giving readers a quick glimpse of hardware choices for AI Video Analytics and reasons behind the same.

Please do let us know if you have any further questions regarding the hardware for AI video analytics. Would love to understand from other practitioners of AI, as to what hardware and infrastructure choices they make and why.

Ankit Agarwal is the Founder & CTO of Sparrosense and can be reached at ankit.agarwal@sparrosense.com

‍

‍

‍