Your GPUs are idle. The data didn't arrive.
An AI factory is only as fast as the data feeding it. A GPU that waits on data still bills by the hour — and a few petabytes of training data can take weeks to land while the cluster sits waiting. The zx Appliance feeds the factory at line rate, moving checkpoints, weights, and datasets ~10× faster than typical movers — so your accelerators stay at work.
Compute is abundant. Moving the data isn't.
- Idle GPUs are among the most expensive line items in the AI factory — and every stall on data movement burns that budget.
- Datasets stall during GPU data staging from primary storage, other clusters, or the cloud, leaving accelerators waiting.
- Multi-terabyte checkpoints and weights have to move fast, often across sites, on every training run.
- rsync and scp are single-stream; Globus leaves throughput to the endpoints you tune — and hand-tuning DTNs to saturate a high-speed link takes months, often without reaching line rate.
A dedicated data-movement layer for AI infrastructure.
Saturate the wire
zx runs at line rate and scales out with no software ceiling — keeping accelerators continuously fed instead of waiting on storage.
Move models fast
Shuttle multi-terabyte checkpoints and weights between storage, clusters, and regions at ~10× the speed of typical movers — so a save or restore is no longer a coffee break.
Burst to cloud
Move file and object across edge, hybrid, and cloud. A co-designed path avoids the 30–50% cloud-path penalty, so you stage data where the GPUs are at full speed.
Day-one performance
A pretuned, system-engineered appliance — hardware matched to zx, no transfer-tuning project, no science experiment. Rack it and move data at line rate from day one.
A wire-speed data path that bypasses the cloud stack.
zx embeds directly in NVIDIA BlueField DPUs — moving data at line rate while offloading it from the host CPU. That sidesteps the 30–50% penalty of standard cloud and host data paths, and does it with fewer servers and less power. Featured on the NVIDIA blog; demonstrated at SC21 and SC22.
Free the host CPU
Data movement runs on the DPU, not the server — so host cores stay on the work that feeds your GPUs.
Skip the cloud-stack penalty
A co-designed, DPU-resident path avoids the 30–50% overhead of HTTP/REST cloud tooling — line rate, in and out.
Fewer servers, less power
More throughput per watt and per rack unit — the efficiency and TCO win that compounds across your fleet.
NVIDIA's blog highlighted that Zettar's BlueField-3 data-migration and storage-offload solution consolidates into about 4U of rack space — versus roughly 13U for an x86-based equivalent.
Zettar moved an actual petabyte over a 5,000-mile network loop in 29 hours — encrypted and checksummed — at 96% bandwidth utilization.
That run was capped at 80 Gbps to spare the shared network — on a full 100 Gbps link, it's a petabyte a day.
Where does GPU time actually go to data — and how do you get it back? How to keep GPUs fed →
Put a number on it — and make your case.
Every hour a GPU waits on data is an hour you’ve already paid for. Delay doesn’t pause the bill — it just stops the work.
Common questions
How do I fix a data bottleneck that is starving my GPUs?
The bottleneck is usually data movement, not compute — datasets, checkpoints, and weights arriving too slowly. Move them in parallel at line rate with a scale-out data mover so the pipeline keeps up; the Zettar zx Appliance feeds clusters at wire speed with no software ceiling.
How do you keep GPUs from sitting idle?
Idle GPUs usually stall on data that has not arrived. Zettar moves datasets, checkpoints, and model weights at wire speed so accelerators stay fed and keep earning.
What is AI data movement or AI data transfer?
Moving the large datasets, model checkpoints, and weights that AI training and inference depend on — across storage, sites, and clouds — fast enough to keep GPU clusters busy.
How fast can Zettar feed a GPU cluster?
At line rate — zx fills the bandwidth between your storage and the cluster, and scales out with no software ceiling. Proven at 1 PB in 29 hours with SLAC and ESnet.
Will it keep up as our GPU cluster grows?
Yes. zx scales out — add servers and aggregate throughput grows, with no software ceiling. Your environment sets the limit, not the data mover, so the appliance grows with the cluster instead of becoming the next bottleneck.
Does it work across hybrid and multi-cloud?
Yes — file and object movement across on-premises, hybrid, and public cloud, bypassing the 30–50% cloud-stack penalty.
Can Zettar run on NVIDIA BlueField DPUs?
Yes. zx runs on NVIDIA BlueField, offloading the host CPU — roughly 4U versus about 13U for an equivalent x86 setup, per NVIDIA.
Stop paying for idle GPUs.
See the zx Appliance keep your AI infrastructure fed at wire speed.
