In this lively and highly informative discussion Troy Smithells from XENON Systems talks about four innovative examples of how AI is being implemented.
Creating CAD files
He has been working with a startup that is using AI to reproduce an industry’s original proprietary drawings as CAD files. Previous to AI, this has been an extremely time-consuming process and very error prone – to the point that the average amount of errors per conversion makes it easier and cheaper to just start again using CAD.
The startup mentioned that with their pilot customer there were close to 50 staff creating those hand drawn images and it is so time consuming getting them converted to DXF or CAD files. Using OpenAI at first the error rate was about 46% but with his own AI model this process is now 4 times faster and close to 100% accuracy and has saved his pilot customer over $750,000 p.a. since inception.
Pipe inspection
Using physical AI (robotics) to map and inspect pipes. Physical pipes often have a very small circumference and there are frequently 90 degree turns, vertical stacks and odd angles to be navigated. This application of AI is using AI robots to map physical pipes and also do things leak prevention and maintenance.
Troy added to this, saying that when people think about AI they think about ChatGPT but we are now at the stage where more and more we will be seeing AI driven robots doing tasks. In this particular case, the AI robots were used to map and inspect a hospital’s piping system, (going up to one km into the pipeline system) which had never been possible before, with substantial savings in leakage assessment alone. The benefits of now having such a capacity is still being explored.
Predictive and Preventative Maintenance
Troy mentioned a use case where a listening device was put on a conveyor belt. This listening device just listens 24/7 and when the conveyer belt starts to change its sound the device signals that change as an indicator of wear and tear with substantial savings in preventative maintenance.
OH&S on Construction Sites
Troy talked about an application partner that uses Computer Vision, which is a form of AI, to identify whether staff and crew on a construction site are adhering to OH&S regulations. Literally, is this person wearing their helmet, is that person wearing their safety googles. And at a more serious level, is that area safe for humans, is this crew member about to be hit by a crane and send off an alert.
The topic of AI is then further developed into a discussion about fork lifts with AI cameras monitoring for obstacles and people and the costs of having cameras everywhere. Troy cited a customer in the construction industry who said that once he had Computer Vision in place it was very easy and cost effective to have it do other things. This company was working on a maritime dock in Cairns in the tropical North of Australia and were using Computer Vision to do “croc alerts” (monitoring for crocodiles). Once one has the platform in place it is really quite easy to start identifying new and different applications.
In this particular case, the client had started with Cloud but over time was no longer using it due to cost. The cost of running AI at scale on Cloud is exponential. The client repatriated their AI from the Cloud to hardware and this came in at under half the cost of using the Cloud.
Troy mentioned that XENON Systems has software called AI JumpStart which allows a client to try out GPU based hardware to trial their application, work out the bugs and then move to an “on prem” which is a hosted environment with full ownership and not reliant on the Cloud. This enabled the client to trial and then implement, the move of a production environment off the Cloud at minimal cost and seamlessly.
Troy makes the point that AI has been around since the 50’s whereas the Cloud is fairly recent and was mainly designed for spinning things up and getting things up and running. The Cloud is very disparate. There is high-performance compute which powers AI but it also has high-performance networking and high-performance storage. AI likes everything in one rack whereas in the Cloud all these three elements are physically elsewhere and anywhere and this is what makes Cloud very uneconomical when running AI.
For this reason, scale use of AI in the Cloud will never make financial sense. By contrast, XENON’s AI JumpStart is great for proof of concept/proof of value. This is how XENON Systems meets the needs of its customers; by providing an “on-prem” environment or a hosted environment before they make any big investment. Using hyper-scale architecture just doesn’t make any economic sense. Maybe in the future there will be different technologies that address this but not today.
In addition, people don’t want to put their data on the Cloud. If one does then there is not a AI hyper-scale provider that is not US-owned and thus they are governed by the US Cloud Act. And of course, anything that is in the Cloud can be appropriated.
AI hardware is big, heavy and noisy and it draws a lot of power. It also requires space and cooling and a lot of companies just don’t have that capability and so they outsource or “host” their AI in a data center.
Troy gave an example of a startup that wanted to run their AI in an old renovated house in Sydney and had a designated room for their AI using CPU based hardware. He had to explain to the client that GPU’s run hot, fast and loud for 5 years until they break. In addition, the server they wanted to buy was 132kgs and runs at 110 decibels. By comparison, a jet engine runs at 120 decibels. Plus one can’t just use a normal split system to cool this equipment. There is too much moisture.
For all these reasons, data center providers are very important for AI as they provide the power, the storage and the cooling. The client owns the rack and has access to the rack but all those problems are managed by the data center.
Other considerations are that GPU accelerated architecture is not simple. XENON is the only NVIDIA Elite Partner in the country and carries that with pride. Getting customers up to the operating system layer is not easy and XENON leverages 30 years’ experience in getting to the point where AI runs and it does what the customer wants.
Troy mentioned that he often sees clients who have bought all the AI hardware but have no idea how to get it running. Quoting Jensen from NVIDIA, “the five layer cake” the bottom layer being power and energy, next layer is the data center, next being GPU’s, and then models and application layers, Troy said that customers are often comfortable with the last two layers but that getting there is where the problems are and that is where XENON can really assist.
The best time for a company to engage the service a company like XENON is at proof of concept when one is looking at production scale planning and workloads otherwise the learning curve is very problematic and resource (time, staff and money) consuming.
The subject of “clean data” was also discussed and Troy said that he was not an expert in this field but he had asked the question of one of XENON’s AI specialist engineers who had replied that it was no necessary as the data was vectorised into 0’s and 1’s and then any GPU cluster can be put to work to make sense of it. So Troy’s advice on the subject of “you must have clean data” is “not necessarily.
The opportunity cost of going too slow was also discussed and Troy reiterated that having got to proof of concept. His final piece of advice is “to just get started and embrace what is here and now. AI is here and now and it is ripe for the picking” and “ready to go”. It is a case of get on board and “it is not that AI is going to take your job, it is that someone using AI will”.
