Outpost VFX Accelerates Model Training with AWS
For ML engineers scaling video or image-based model fine-tuning: this Outpost VFX and AWS case study is one of the more technically detailed published examples of PyTorch DDP on AWS P5 instances (NVIDIA H100/NVLink) for production VFX workloads. Reported: Outpost VFX, a UK-headquartered VFX studio with operations in Canada and India, worked with the AWS Generative AI Innovation Center over a 6-week advisory to convert their face swap model to multi-GPU DDP training on P5 instances. Baseline was single-GPU training on RTX 3090 workstations (fine-tune: 1-2 weeks). Result: 'up to 8x improvement in face replacement model learning speeds,' per the co-authored AWS blog. v001 client delivery now takes 2 days versus the previous 1-2 week cycle. The technical lever: NVLink-connected H100s on P5 provide substantially higher gradient synchronization bandwidth than PCIe-connected G5 GPUs.
For practitioners migrating video or image ML workloads to cloud GPU clusters, this case study from Outpost VFX and AWS provides verified specifics that vendor write-ups rarely include: the exact distributed strategy (PyTorch DDP), the instance class (P5/H100), the interconnect advantage (NVLink over PCIe), and a concrete before/after delivery timeline.
What happened
According to an AWS Machine Learning Blog post co-written with Tim Chauncey (CTO) and Dheeraj Bhadani (Lead Software Architect) of Outpost VFX, the studio moved its face replacement model fine-tuning from local RTX 3090 GPU workstations to AWS P5 instances with NVIDIA H100 GPUs. The collaboration was structured as a 6-week advisory with the AWS Generative AI Innovation Center. The result, per the post: "up to 8x improvement in face replacement model learning speeds." v001 delivery to clients for initial director review now takes 2 days, versus the previous cycle of 1-2 weeks per fine-tune. Traditional non-AI compositing for beauty and de-aging approvals previously required over 5 days, per the blog.
Technical implementation The engineering approach centered on PyTorch Distributed Data Parallel (DDP): model weights are copied to each GPU, enabling the system to process more images per training batch. This is a model-parallelism-on-data strategy, not pipeline or tensor parallelism - effective for fine-tuning when the model fits in a single GPU's VRAM but throughput is the bottleneck.
The instance choice - AWS P5 with NVIDIA H100 GPUs and NVLink interconnects - is the architectural decision driving the speedup. NVLink provides significantly higher bandwidth for gradient synchronization across GPUs compared to PCIe used in G-series instances. The H100 also brings 14,592 CUDA cores and 80GB of HBM3 memory, a substantial upgrade from the RTX 3090 baseline. The implementation ran within a segregated, secure cloud environment, a requirement for handling sensitive production footage (unpublished on-set material).
Measured results and quotes The performance baseline was one GPU on a G5 instance versus multiple H100s on P5. The blog reports up to 8x improvement in training speed to reach a defined loss threshold.
Tim Chauncey, CTO of Outpost VFX, is quoted: "We are now able to iterate much faster thanks to our parallelized workflow and the ability to harness multiple top-end GPUs at once. Speed of iteration is critical to VFX work, and this architecture provides more robust and scalable capabilities for future development."
Dheeraj Bhadani, Lead Software Architect, is quoted: "What excites me most is that these models are no longer research experiments; they are becoming an integral part of the modern VFX pipeline."
Practitioner implications
This case study is more actionable than typical vendor posts because it names the distributed strategy, instance class, and interconnect. For engineers evaluating similar migrations: the NVLink/P5 advantage over G-series/PCIe is specifically a gradient synchronization bandwidth benefit - relevant when training batch size is the limiting factor. The 6-week advisory structure suggests that GPU migration at this level still requires significant engineering work beyond selecting the right instance type. Teams should plan for distributed code conversion, not just a lift-and-shift.
The blog does not publish cost figures or per-image-resolution dataset details. Cost per training run and data staging I/O remain practitioner-measured variables.
What to watch
Outpost VFX and AWS identify Amazon SageMaker AI (managed training, model versioning, hosted inference) as a planned next step for further streamlining model development across global studios. Published results from that expansion - including cost and orchestration overhead - would make this case study substantially more replicable.
Key Points
- 1PyTorch DDP on AWS P5 (NVIDIA H100/NVLink) cut Outpost VFX face-swap fine-tuning from 1-2 weeks to 2-day v001 delivery, an up-to-8x speedup.
- 2NVLink-connected H100s on P5 outperform PCIe-connected G5s specifically because gradient synchronization bandwidth is the distributed training bottleneck.
- 3Vendor case study names exact framework (DDP), instance (P5), and interconnect; cost per run and data staging remain practitioner-measured gaps.
Scoring Rationale
A technically detailed vendor case study with named spokespeople, specific distributed training architecture (PyTorch DDP on P5/H100/NVLink), and a concrete before/after timeline (1-2 weeks to 2-day delivery). Score reflects practical value for ML engineers evaluating cloud GPU migration, offset slightly by single-source vendor co-authorship and absence of cost metrics.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
