Adobe Reveals Two-Thirds of GPUs Sat Idle During AI Training
The disclosure shows that even a major AI developer can lose most of its GPU budget to idle time, and that the fixes were about data plumbing, not model design.
Reporting from 1 source: GIGAZINE.
At GTC 2026, Adobe CTO Ellie Greenfield said that in early training runs, about two-thirds of thousands of GPUs were idle, wasting roughly $600,000 per $1 million spent. Causes included slow data delivery and uneven workload distribution. Adobe fixed this with a high-performance network and fragmented checkpointing.
Adobe's CTO Ellie Greenfield detailed at GTC 2026 how the company built custom large-scale generative AI models, and what the profiler output showed. The data pipeline, not the model, was the bottleneck. Adobe replaced standard Ethernet with a high-performance fabric and split checkpoints into fragments saved across multiple locations, cutting the time spent on continuous write operations. The second problem, uneven workloads, was a data loader issue: equal-sized chunks did not mean equal processing time. Backblaze, which explained the presentation, noted the issue is not unique to Adobe.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.