Australian media production house Kennedy Miller Mitchell (KMM), known for its work on action film Furiosa: A Mad Max Saga, has replaced its HPE storage and compute infrastructure with Dell hardware based around PowerScale NAS arrays.
The move comes as it leverages generative artificial intelligence (GenAI) that allows it to pre-visualise movie scenes before they are shot, and looks forward to a future in filmmaking where productivity sky-rockets.
Before production work on Furiosa began, KMM – which has worked on the Mad Max films since the beginning – replaced all of its HPE infrastructure with storage arrays, workstations, monitors and switches from Dell.
Besides simple infrastructure modernisation requirements, choosing Dell was also governed by one particular requirement: to deploy hardware well-adapted to GenAI. That’s because key to KMM’s work now is pre-visualisation of film scenes, said chief information officer Yan Chen.
“In other words, we direct and edit the film in 3D before we even shoot it,” he said. “That allows us in advance to evaluate which scenes we will actually shoot and those we might waste our time on because we would not keep them during editing. It’s a very economical approach. And generative AI is becoming increasingly relevant for this preparatory work.
“We combine LLM models, image generation models – notably Stability AI’s Stable Diffusion – and models generated from voice that we have personalised for our own needs,” said Chen. “Then, we combine them to generate our pre-visualisation rushes. We really have set up and AI factory.
“Generative AI is playing an increasingly important role in filmmaking, even in the images you see ultimately on the screen,” he said. “That goes from deepfakes to voice replacement, via rotoscoping [transformation of a filmed character to a drawn one] and compositing [fusing different images]. Generative AI is useful to accelerate a number of production stages from the beginning to the end.”
PowerScale provides performance
Furiosa represented 20PB of data that had to be processed by GPUs deployed on workstations. KMM said it will eventually replace these with AI servers.
KMM’s Dell Precision workstations are equipped with Nvidia GeForce RTX 4090 cards. These cards use the same Ada Lovelace circuits found in the L40 and L4 GPUs that Nvidia sells in servers aimed at inference workloads. These have the fourth-generation Tensor cores also present in the Nvidia H100 GPU used on supercomputers.
Such configurations can be found in HPE’s catalogue, but it’s the PowerScale NAS array that convinced Chen to change supplier. PowerScale is the modern iteration of Isilon scale-out NAS, and offers storage I/O at speeds of up to 200Gbps via Ethernet to workstations.
A key advantage of PowerScale is its high availability. The 20PB of data are shared across several SSD nodes that offer a direct connection to the workstation even when several users have access to the same file (which may be stored on different nodes).
On the latest version of OneFS, a PowerScale NAS can transmit data via RDMA-over-converged Ethernet (RoCE). This extended Ethernet protocol allows files to reside directly on RAM and so bypasses the potential bottleneck at the processor.
Yan Chen didn’t confirm whether RoCE is in use at KMM, but said the company had deployed new Dell Ethernet switches to guarantee an optimal throughput between the PowerScale array and the Precision workstations. In fact, RoCE isn’t routable except using converged Ethernet-compatible switches.
“The only part of our infrastructure that we haven’t replaced with Dell equipment are the switches from Ubiquiti that allow the workstations to communicate with each other,” he said.
Chen recognises that replacing everything has translated to a particularly costly IT migration. “But we have benefited from an exceptional budget of AU$350m [US$233m] for Furiosa: A Mad Max Saga,” he said. “It is the biggest budget of all time for an Australian movie. It allows us to make this investment, which will prove very profitable in years to come.”
Chen said KMM’s next investments will be in servers, with likely targets those equipped when the new B100 and B200 GPUs are available.
“This generative AI infrastructure represents a real industrial revolution in cinema,” he said, as he contemplated a radical change in the mode of filmmaking and distribution. “In Hollywood, producers estimate generative AI will allow them to make 300 or 400 films a year, as opposed to 30-something now. Nevertheless, that poses an additional problem. Will people go to the cinema to see so many films?”