GPU Utilization: Adobe's GPUs Sat Two-Thirds Idle

Low GPU utilization at Adobe: GPU idle two-thirds of the time

GPU utilization at Adobe was low in the early training runs. Backblaze reports that Adobe's CTO, Ely Greenfield, told the audience at NVIDIA's GTC about the company's three-year effort to build frontier generative AI models from scratch. Backblaze adds that the early runs showed roughly two-thirds of GPU time spent simply waiting for data.

Greenfield put the cost in dollars. Of every million dollars going into training, $600,000 was burned on GPUs sitting and doing nothing. Backblaze restates that as roughly sixty cents wasted per dollar of GPU compute.

What the early runs looked like, per Backblaze

Measure	Value
GPU time spent waiting	Roughly two-thirds
Wasted per $1M put into training	$600,000
Wasted per dollar of GPU compute	Roughly sixty cents

All the Adobe figures come from Backblaze's post of August 25, 2026, not from the talk itself. The session is listed on NVIDIA On-Demand as S82197.

AI training bottleneck one: storage throughput and checkpointing

The first AI training bottleneck was how fast data could reach the GPUs. Backblaze says Greenfield's team traced the waste to two causes, neither unique to Adobe: whether data reaches the GPUs fast enough, and whether the work is spread evenly once it arrives.

Adobe's training data lived in petabytes of distributed cloud storage. That data was shipped constantly to thousands of GPUs over standard Ethernet.

Checkpointing added to the bill. A checkpoint is a saved copy of a training run's state. Backblaze reports that most of Adobe's checkpoints were never needed again, yet writing and reading them still consumed real GPU time, which cost a lot of money whether they were used or not.

Adobe's fix was to replace standard Ethernet with a high-performance networking fabric designed for the petabyte-scale traffic that distributed AI training generates. For a related effort on training network traffic, see MetaRoCE: Meta's RDMA over Ethernet for AI, No PFC.

The Backblaze post runs long, with the Adobe account and the product section in one page. To keep it as notes with its headings intact, the URL to Markdown tool converts a web page into Markdown from its address.

GPU bottleneck two: data parallelism with a balanced data loader

The second GPU bottleneck was how the work was divided. Adobe's pipeline used data parallelism in its simplest form: it sliced the training data into equal-sized chunks and assigned one chunk to each GPU in the cluster.

Equal-sized chunks, Backblaze notes, were not equal work. Some GPUs received simple, low-resolution assets that processed in seconds, while others received large, complex files that took minutes. The GPUs with light chunks finished early and then waited.

Adobe's answer was a custom-built balanced data loader. It understood the processing cost of each asset and distributed work so that every GPU finished at roughly the same time.

GPU utilization limit: Adobe reached about 80 percent

After both fixes, Adobe's GPUs ran at roughly 80% utilization. Backblaze calls that close to the practical maximum, because thousands of machines running together carry some coordination overhead. Its section title for this result, "From 40 to 80 cents", frames it per dollar of GPU compute: useful work rose from about 40 cents to about 80 cents.

GPU utilization before and after, per Backblaze

State	GPU utilization
Early training runs	More time waiting than computing
After both fixes	About 80%

How to increase low GPU utilization: fix storage, then the loader

Backblaze puts storage first and the data loader second. Its wording is that the loader optimizes what happens once data is there, but storage determines whether it is there at all.

Backblaze adds that building something like Adobe's balanced data loader takes real engineering investment. That includes time spent profiling workloads, testing distribution strategies, and tuning until every GPU finishes at roughly the same pace. Teams still fighting storage bottlenecks rarely get to that work.

Adobe GPU idle fix: the Backblaze post is also a storage pitch

Backblaze sells cloud storage, and the post ends with a section titled "The Storage Side of the Solution" that promotes its own service. The Adobe figures above come from the first part of the post, not from that closing section.

• The Adobe account in the first half is Backblaze's report of a conference talk: the two-thirds figure, the two causes, the fixes and the roughly 80% result.
• The closing section describes Backblaze's own product.