Seven practices, one engineering team. We design, build, benchmark and operate the compute layer beneath your hardest problems — on-premise, hybrid or sovereign cloud.
A genomics pipeline and a trading engine are not the same machine. We start from your regulatory perimeter, your data gravity and your deadlines — then design backwards.
Industries · Media & Entertainment
A facility does not lose a show to a slow render. It loses it to ninety artists each losing forty minutes a day to a scene that will not open, a cache that re-cooks, a queue that reordered itself overnight. We design the pipeline around that number, then measure it every week of the production.
The constraint we design around first
Farm utilisation charts look healthy while a show slips. Utilisation measures whether the hardware is busy. It says nothing about whether the artist who needs the result is sitting on their hands. We instrument the second thing.
In the facility audits we run, the wait an artist experiences is usually dominated by file open and cache behaviour rather than by frame time. The renders are generally fine. Getting to the point of submitting one is not.
None of these are visible on a GPU utilisation dashboard. All of them show up in a per-artist timing log, which is the first thing we ask to be allowed to collect.
Render farm and queue management
Every facility eventually has the same meeting: two shows, one farm, one week to deliver. Whether that meeting is calm depends entirely on whether the priority and preemption model was agreed while nobody was under pressure.
01
Deadline and Tractor dominate for a reason: they understand DCC applications, licence-aware limits and per-task dependencies out of the box. OpenCue suits facilities that want to own their scheduler and have the engineers to do it. Slurm is the right answer when the same cluster also runs simulation and machine learning work and you want one accounting model. We will tell you when your existing scheduler is fine and the problem is elsewhere.
02
Node pools carved by GPU memory, host memory and licence entitlement, so a task that needs 48 GB of VRAM never lands on a 24 GB card and fails at frame 400. Limits are expressed against the scarce thing — plugin licences, simulation threads, storage connections — rather than against slot counts that stopped being meaningful when nodes got bigger.
03
Three bands, not fifteen: interactive and lookdev work that an artist is waiting on, show-critical batch, and background bake. Preemption applies only downward and only to work that checkpoints or is cheap to restart. A 14-hour simulation is never preempted, which is why it belongs in its own pool with its own reserved capacity rather than competing on priority.
04
The compute is the easy part. Renderer and plugin licensing is what makes cloud burst uneconomic, because floating licences served from your facility do not follow the work, and cloud-metered entitlements are billed per hour of wall clock including the hour you spent syncing textures. We model burst cost with the licence line, the egress line and the data-staging window included. Sometimes the answer is that buying twelve more nodes is cheaper than bursting twice.
05
Simulate, cache, render, comp, slap-comp, review. When one stage is patched, only the affected downstream tasks should re-run. Getting that right means job definitions carry real input hashes rather than timestamps, which is unglamorous work that removes a large share of wasted farm hours.
06
Frame hours, licence hours, storage occupancy and power attributed per show and per sequence, exported weekly. This is what lets a producer argue for more capacity with evidence, and what stops the farm being treated as a free resource with no owner.
Storage for media
Media storage is sold on sustained sequential throughput because that demo looks impressive. The workload that decides whether artists are productive is millions of small reads against textures, caches, geometry and OpenUSD layers. Those two profiles want different things from the same box.
Streaming 4K DPX at 2 GB/s is a solved problem. Serving 40 000 texture stats per second while doing it is not. We size the metadata tier and the client-side cache before we argue about spindles, because that is what scene open time responds to.
A show's total data is not the number that matters. The number that matters is how much of it is touched in a week. We derive that from access telemetry rather than from a producer's estimate, then buy fast capacity for the working set and cheap capacity for the rest. Getting this wrong in either direction is expensive: undersized and everything queues, oversized and you bought performance the show cannot consume.
| Tier | What lives here | Dominant IO profile | What it is sized against |
|---|---|---|---|
| Node-local NVMe | Frame scratch, sim caches in flight, texture cache | Mixed, burst, single-client | Largest single task's peak footprint |
| Hot project tier | Active sequences, published caches, textures, USD layers | Metadata-heavy small reads, high concurrency | Weekly working set plus 40% headroom |
| Editorial and plates | Camera originals, conform media, dailies masters | Sustained sequential, few clients | Concurrent scrub and conform streams |
| Nearline project | Sequences not in active work this fortnight | Occasional large reads on recall | Recall time an artist will tolerate |
| Archive | Delivered shows, source masters, full project bundles | Write once, read rarely, verify regularly | Restore window agreed with production |
GPU rendering, realtime and the volume
A batch render that misses its estimate costs money. A volume that drops frames costs a shooting day with the cast standing in it. The engineering discipline is closer to broadcast than to a render farm, and the budget should reflect that.
Moving a show to GPU rendering is a memory question before it is a speed question. We profile scenes for VRAM footprint, out-of-core behaviour and denoiser cost, then specify cards against the heaviest sequence rather than the average. Where a show will not fit, we say so and keep a CPU partition for those shots instead of forcing a migration mid-production.
Layered USD composition lets departments work concurrently without a check-out ritual, but it moves the cost onto the filesystem: stage composition is a metadata storm. We tune the storage path and the asset resolver together, and cache resolved stages where the pipeline allows it. Hydra delegates make renderer choice a runtime decision rather than a rebuild.
The render cluster behind a volume has a hard frame budget and no retries. Design work concentrates on genlock and PTP distribution to every node and processor, nDisplay-style frustum partitioning across render nodes, deterministic network paths with no shared traffic, and colour transforms verified on the wall itself. Spare nodes are racked, configured and idle, because a failover on the day cannot involve a build.
A lookdev artist with a capable local GPU does not submit to the farm for every iteration. That is often the cheapest idle-time reduction available, and it is frequently overlooked because workstations sit in a different budget line from the cluster. We size both together and show the trade in one model.
The question is never whether the farm is fast. It is whether an artist can try an idea, see it, and try the next one before they lose the thought.
Colour and review workflow
Colour management is an infrastructure responsibility, not a per-artist preference. One OCIO configuration, one distribution mechanism, one calibration schedule, and a review path whose latency you have actually measured rather than assumed.
The ACES or show-specific OCIO config lives in version control and is deployed to workstations, render nodes and review rooms by the same mechanism. Nobody edits a local copy. When a show changes its display transform, the change is a commit with a date, which matters when a delivery is questioned six months later.
Content security
Content security controls are not optional overhead on a show with a marquee title. They are the reason you are allowed to bid. Building them in is far cheaper than retrofitting them the week a studio sends its questionnaire.
Separate storage namespaces, separate project groups, separate render pools where a studio requires it, and network policy that prevents one show's assets being reachable from another show's workstation. Enforced by the platform rather than by a naming convention and good intentions.
Access control, logging, media handling, physical security, patching, personnel screening and incident response, each mapped to a control reference with an artefact behind it. The goal is an assessment where the assessor reads documents rather than interviewing your engineers for two days.
Forensic and visible watermarking applied at the review and dailies layer, keyed per viewer and per session, with the key mapping retained long enough to be useful. It does not prevent a leak. It changes the behaviour of the person considering one, and it makes attribution possible afterwards.
Pixel-streamed sessions with no local media, managed endpoints, clipboard and peripheral policy, and egress restricted to the review path. If a studio will not accept remote work on a title, the architecture should make it easy to run that show on-premise only rather than forcing a policy fight.
We keep the questionnaire response, the network diagrams and the control evidence as maintained artefacts that update when the estate changes. Most facilities rewrite this from memory for every bid, which is where inconsistencies that trigger findings come from.
Questions facilities actually ask
If your situation is not covered here, ask. We would rather say no early than discover a mismatch in week three of a show.
It depends on how many weeks a year your farm is saturated. A facility with two predictable crunch windows is usually better off buying for the steady state and bursting for those windows. A facility that is saturated nine months a year is paying a rental premium on work it could own. The three lines that decide it are licence cost in the cloud, egress and staging time for textures and caches, and whether your pipeline can address remote storage without rewriting paths. We build the model with your real shot list and show you the crossover point.
That is the most common call we take. Sequential throughput and small-file metadata performance are almost independent, and media storage is marketed on the first. Ask your vendor for metadata operations per second at your concurrency, measured with a file-size distribution taken from your own project tree. If the answer is a shrug, the bottleneck is probably there. Fixing it is sometimes a configuration change and a client-side cache rather than new hardware.
That is the normal arrangement. Pipeline TDs know the show, the tools and the artists better than we ever will. We take responsibility for the compute, storage, network and scheduler layers and for measuring what the pipeline experiences. Where we think a pipeline decision is costing artist time, we bring a measurement rather than an opinion.
Sometimes, with a narrow scope. Mid-show we will look at scheduler configuration, storage tuning, cache policy and obvious requeue waste, because those can move inside a week without touching the pipeline. We will not re-architect storage or migrate a renderer during a production. If the only real fix is structural, we will tell you that and help you plan it for after delivery.
A fixed number of render nodes with matched hardware, a deterministic network with nothing else on it, genlock and PTP to every node and processor, a colour pipeline verified on the panels rather than on a monitor, configured hot spares, and an operator runbook written for the person who will be there at 5 a.m. rather than for the engineer who built it. The compute is rarely the risk. Timing, colour and the failover plan are.
Below roughly twenty artists, a well-chosen commercial NAS, a straightforward Deadline setup and good workstations will serve you better than anything we would design, and we will say so. Our work pays for itself when the number of people waiting is large enough that an hour of idle time per person per day is a real line in the budget.
Let's Talk
A week of scheduler logs and a representative scene are enough for us to model your farm and storage path and put a number on recoverable idle time. If the answer is that your current estate is adequate, we will say that instead of quoting.
Sydney · Melbourne · Canberra. Australian owned, Australian staffed.