RDMA speed meets tape-scale capacity. S3 over RDMA through NVIDIA’s cuObject framework moves data directly between GPU memory and the archive — and with ScoutAM behind it, that speed now carries all the way to tape.
Zero-copy data path. No CPU, no protocol stack, no staging buffers
First GPUDirect performance brought to a tape archive system
Open source with Apache 2.0 license in Versity S3 Gateway
Most S3 compatible storage moves data over TCP/IP. TCP is reliable and well understood, but every transfer pulls in the host CPU — copying between memory buffers, managing context switches, processing packets in the kernel. As network and accelerator speeds increase, that gap widens and expensive compute sits idle waiting on data.
Every hop is CPU-mediated. The bottleneck recurs through the whole data lifecycle, and worst of all on large dataset recall from mass storage.
Data moves directly between client memory and gateway memory. No temporary files and for GPU workloads no CPU in the data path at all.
One S3 request, split into two channels
Control stays in-band
Session setup, authentication, bucket verification, and object metadata flow through the standard S3 API, exactly as for any other request.
Data moves out-of-band
The object data travels over an RDMA connection running alongside the S3 session. Never through the client CPU, never through the S3 protocol stack.
The path is zero-copy
No staging through intermediate buffers or temporary files. For GPU workloads, data travels between the NIC and GPU memory directly.
An RDMA data path alongside the existing control path.
The Versity S3 Gateway is a stateless, modular S3 gateway that connects applications other storage systems. With GPUDirect support, it adds a separate RDMA data path alongside the existing S3 control path. Authentication, bucket verification, and metadata continue through the standard S3 API, while object data moves directly between client memory and gateway memory over RDMA. From there, the gateway reads from or writes to the selected backend using its existing storage path.
Recall at RDMA speed makes tape practical for active AI pipelines
Archiving to tape stays what it has always been — a steady, ongoing ingest. RDMA changes the equation on the way back out. AI training pipelines don’t read a dataset once; they reread it every time a model is retrained or a dataset reprocessed.
Re-ordered recalls
ScoutAM re-orders large data set requests to match optimal tape offset locations.
Fewer mounts, more streaming
Minimized tape mounts and maximized streaming performance across the library.
Straight into GPU memory
RDMA transfer to GPU makes massive dataset recall feasible for model training.
Open source, Apache 2.0. License
The GPUDirect Tape Gateway, including support for RDMA and cuObject, are available in the open source Versity S3 Gateway. Any team already running or considering versitygw can adopt GPUDirect acceleration.
Built for workflows that need archived data back in motion
AI & ML training
Loading and reloading training data and checkpoints into GPU memory.
HPC & research
Simulation outputs rereadable at speed for new analysis runs.
Genomics
Petabyte-scale sequence archives kept active rather than cold.
Media & entertainment
High-resolution masters recalled on demand from long-term storage.
Bring us your workload, and we’ll show you how data moves from GPU or host memory over RDMA, through the Versity S3 Gateway, into ScoutAM-managed tape, and back again. An engineer will walk through a personalized solution tailored to your hardware and archive environment.