Tin Drum is a start-up on a mission to create immersive artistic experiences in Mixed Reality (MR).
In February 2019, they showcased “The Life”, the first large scale public Mixed Reality (MR) exhibition in history, at the Serpentine Gallery in London. It was created with the legendary artist Marina Abramović as a representation of her performance work. As Tin Drum pointed out from the outset,
“No one had ever presented a piece in MR as anything other than tech demo, and there was no template from which to draw. We had to make it all up from scratch”.
The rendering and other processes involved in creating good MR content are computationally expensive and time consuming. It also takes time to find the best lighting schemes and fix imperfections on the captured content, so fast iteration is essential to arrive at a version of the exhibit that everyone is happy with.
Exor Technology came on board to assist with accelerating the turnaround time of concept iterations by using the power of the AWS Cloud. When we arrived, it was taking the team anything between a day and three days to produce an iteration of their current project.
Marcus Fielding, COO of Tin Drum, says:
“We wanted to be able to concentrate on the artistic and creative aspects of our projects. We knew the Cloud should be able to help with the compute-intensive processing needed to generate our MR content, and we hoped that it could also speed that process up so that we could experiment creatively at a greater rate. We needed Exor Technology’s help to get us started on AWS”
First, we benchmarked a number of different types of AWS instances on a small sample of content. Interestingly, a bigger machine does not always give you a faster render. There is a sweet spot within various families of AWS EC2 instances, so it is worth trying a few before making a decision.
After that, it was a case of spinning up an instance that had all the software we required to process the frames (in this case it was Blender and some compression tools) and making an image of it, so we could create multiple copies when the time came.
Since EC2 instances are charged by the hour, the most efficient thing to do is estimate how many frames your chosen machine can process in an hour, and then divide the total number of frames by that to give you the number of EC2 instances you will need in your “render farm”.
The original content and the output needs to be put somewhere, so we created an Elastic File Store (EFS) file system that could be mounted by all the EC2 instances as they came to life.
We then wrote and ran a script that divided the rendering job into chunks that could be done in an hour, and for each of those chunks spun up an EC2 instance with instructions to process a given batch.
Then we went to make a cup of tea.
An hour later, it was all done!
So the team went from being able to produce, at best, an iteration a day to an iteration an hour! Total cost per iteration: US$28 (plus EFS storage cost).
It was a win for everyone. At Exor Technology we learnt valuable lessons on how to manage and manipulate MR content. For Tin Drum, Fielding says:
“Exor Tech made sure they understood our MR content pipeline and then developed and benchmarked various options for parallel processing our content on AWS, all the time working with our internal teams to share their knowledge and develop a suitable cloud solution. We were delighted with the outcome which reduced our content turnaround from days to hours.”
And, of course, the lucky ones who got to see it loved it!