Versity demonstrates that tape libraries can serve as a cost-efficient data source for GPU-driven AI workloads, enabling cold-storage tiering.

Versity has added a high-speed S3 over RDMA capability to its ScoutAM archive gateway software, enabling tape archives to pump data directly into a GPU's high-bandwidth memory. The open-source Versity gateway supports S3 over RDMA (Remote Direct Memory Access) through NVIDIA's cuObject framework, which accelerates data transfer between the archive's S3 gateway and the S3 client. Behind this fast front end, Scale Out Archive Manager (ScoutAM) manages the tape library, writing data to tape and reading it back at scale. The result is a GPUDirect Tape Gateway that pairs full-speed RDMA at the front end with the capacity and durability of tape archive on the back end.

Versity CEO and co-founder Bruce Gilpin explains the shift in thinking around tape infrastructure: "With the cost of NVMe and memory skyrocketing and a 2U server now back-ordered for months, people are getting desperate to build out AI infrastructure with unconventional and unconstrained storage systems. Tape has not been used much in AI because the players were financially unconstrained. But that is changing, partly due to budget issues but more due to availability. Tape drives stream data faster than people realize and you can run them in parallel. 50 drives can generate a 20 GBps stream to feed GPUs and it's dirt cheap. One big player is experimenting and I think we will see others. Lucky for us, you need a super capable software stack to orchestrate all of this but the base storage system is a bunch of cheap tapes and relatively cheap drives and libraries."

The Versity Gateway is a stateless, modular, open-source S3 gateway that translates standard AWS S3 commands into operations against a range of backend storage systems, including POSIX filesystems, ScoutFS/ScoutAM (supporting tape), and cloud storage. The new architecture adds an RDMA-based data path alongside the gateway's existing TCP control path. It scales by distributing many GPU or host client systems across any number of gateway server instances, each pair establishing its own direct RDMA path between client memory and gateway memory.

RDMA allows data to move directly between the memory of one system and another, without routing through the host CPU. NVIDIA's cuObject framework, part of the GPUDirect Storage ecosystem, applies this to S3-compatible object storage, giving applications a way to move data directly between GPU or system memory and object storage services using RDMA transports such as InfiniBand or RoCE. The cuObject architecture splits an S3 request into two separate channels: session setup, authentication, bucket verification, and object metadata flow in-band through the standard S3 API, while the object data itself moves out-of-band over an RDMA connection that runs alongside the S3 session. The RDMA data channel never passes through the client's CPU and never passes through the S3 protocol stack, avoiding the overhead that the protocol would otherwise add to the transfer.

For workloads involving GPUs specifically, data can travel directly between the network card and GPU memory, skipping the CPU almost entirely. A GPU host system—comprising GPU, CPU, and NIC—connects over RDMA (RoCE/IB) directly to a Versity Gateway server system's NIC and into gateway memory, with backend storage reached through the gateway's normal path. For PUT requests, the object is uploaded to the gateway's memory via RDMA, which the gateway then streams to backend storage. For GET requests, data is first staged into gateway memory from backend storage and then transferred to the client over RDMA, enabling client RDMA access to any of the gateway's supported backends.

ScoutAM's GPUDirect Tape Gateway addresses a problem that has become more pressing as datasets grow: how to work with enormous volumes of data on cost-effective, durable media without turning storage operations into a slow-down drag on GPU-driven pipelines. Archiving data to tape is likely to remain a steady, ongoing ingest process, but RDMA changes the equation on retrieval. AI training pipelines do not read a dataset once; they reread it every time a model needs to be retrained or a dataset needs to be reprocessed. ScoutAM can reorder large data set requests to match optimal tape offset locations, minimize tape mounts, and maximize streaming data performance. Combined with the new capability to utilize RDMA transfer to GPU, recalling massive datasets from tape becomes feasible for model training and other large-scale data processing pipelines.

Organizations running AI training pipelines, large-scale machine learning jobs, or HPC applications often find their storage layer, not their compute layer, is the limiting factor. Because performance gains are most pronounced on the read side, the workloads that benefit most immediately are those defined by repeated, large-scale reads: loading and reloading training data into GPU memory, feeding checkpoints back into resumed training runs, and rereading existing datasets for new analysis. For tape specifically, this changes what's practical to keep in an active archive rather than treating it as cold and slow storage. Training datasets, checkpoints, and simulation outputs that might previously have seemed too slow to recall efficiently can now be pulled back at RDMA speed when needed again. Organizations in fields such as genomics, media and entertainment, scientific research, and large-scale AI development—all of which routinely generate data volumes measured in petabytes—get a faster path both to long-term retention and back out again.

Data still arrives and is retrieved as standard S3 objects; POSIX filesystems and ScoutFS continue to work as before, and ScoutAM still manages tape as it always has. Because the Versity Gateway is open source and includes RDMA support as part of the same project rather than as a separate paid add-on, teams can add this capability to existing S3 workflows without adopting a new proprietary gateway or license.

MinIO's AIStor, announced in late 2024, also supports fast S3. It extended MinIO's Enterprise Object Store software with the S3 API, PromptObject, and support for S3 over RDMA to provide fast object storage for AI training and inferencing. MinIO states that data moves directly from AIStor to GPU memory over RDMA, bypassing the CPU and system memory entirely in a zero-copy, kernel-bypass approach. However, AIStor is a paid and proprietary product, unlike Versity's open-source GPUDirect Tape Gateway, which ships as part of the Versity S3 Gateway, with RDMA and cuObject support included without a separate paid tier.

In June, BeeGFS parallel filesystem supplier ThinkParQ and tape archive supplier GRAU DATA formed a strategic partnership through which the latter's XtreemStor tape archive system was integrated with BeeGFS. The two companies stated that this integration provides a storage infrastructure capable of handling the high performance demands of data-intensive AI, HPC, and big data workloads alongside the long-term storage of data needed for and generated by such applications.