Build Custom Large-Scale Generative AI Models | NVIDIA GTC
TL;DR
Adobe's CTO explains why the company chose to build proprietary generative AI models from scratch to ensure legal compliance and creative control, then details how they discovered that naive scaling approaches resulted in GPUs sitting idle 60-70% of the time due to coordination bottlenecks.
🎯 Strategic Decision to Build 3 insights
Professional creatives reject prompt roulette
Off-the-shelf models produced random outputs unsuitable for Adobe's customers, who required precise iterative control rather than gambling with text prompts to achieve their specific vision.
Legal liability blocked enterprise adoption
Enterprise legal departments refused available models due to copyright and IP training risks, forcing Adobe to use fully licensed, human-moderated datasets with complete provenance tracing.
Differentiation justified massive investment
Adobe determined that control and legal compliance provided sufficient competitive differentiation to justify building custom frontier models despite requiring millions in GPU infrastructure.
🛠️ Initial Technical Architecture 3 insights
Naive scaling with PyTorch Lightning
The team initially approached large-scale training as simply adding more GPUs to standard loops, using PyTorch Lightning and AWS S3 storage with thousands of NVIDIA A100s.
Simple data parallelism strategy
They implemented straightforward data parallelism that split petabytes of training data across GPUs, where each processed independently before a manager node collated updates.
Early validation masked inefficiency
The first working model validated the technical approach but exhibited typical early-generation artifacts while hiding severe resource underutilization that threatened cost viability.
⚠️ The Utilization Crisis 3 insights
GPUs idle 60-70% of training time
Profiling revealed GPUs sat empty approximately two-thirds of the time waiting for a manager node to gather and merge model updates, effectively wasting $600,000 of every $1 million spent.
Data parallelism fails beyond 16 GPUs
The straightforward data parallelism approach creates exponentially worse coordination bottlenecks when scaling beyond roughly 16 parallel processors, making it unsuitable for frontier model training.
Storage and checkpointing bottlenecks
Loading petabytes from distributed storage and saving massive checkpoints for insurance created constant stalls, compounded by CPU preprocessing delays and unnecessary GPU synchronization.
Bottom Line
Scaling AI training requires fundamental pipeline architecture changes to eliminate coordination overhead and storage bottlenecks, not simply adding more GPUs, as standard data parallelism becomes exponentially inefficient beyond small clusters.
More from NVIDIA AI Podcast
View all
Securing Long-Running AI Agents: From Setup to Sandboxing
NVIDIA details the shift toward autonomous 'long-running' AI agents capable of independent multi-hour execution, introducing the NVIDIA Agent Toolkit featuring open Neotron models, packaged CUDA-X skills, and runtime security to enable scalable enterprise deployment.
How NVIDIA Blackwell and NVIDIA Dynamo Scale AI Agents for Production
NVIDIA Blackwell delivers up to 40x more concurrent AI agents per GPU than Hopper through its rack-scale NVL72 architecture and Dynamo framework, fundamentally shifting AI infrastructure measurement from token throughput to agent concurrency benchmarks.
Build Video Analytics AI Agents with Skills
NVIDIA introduces the Video Search and Summarization (VSS) blueprint for building vision AI agents that process billions of camera streams using vision language models and a new 'skills' framework, enabling deep video search and summarization 60x faster than manual review.
Ask the Experts: Nemotron 3 Nano Omni | Nemotron Labs
NVIDIA researchers detail the development of Nemotron 3 Nano Omni, explaining how they evolved a text-only model into a multimodal system capable of processing vision, audio, and video through progressive training stages while maintaining the hybrid Mamba-Transformer architecture.