AMD recently released ROCm 10.1, with data movement as the release’s organizing theme. The centerpiece is AMD Infinity Storage (AIS), AMD’s umbrella for libraries that move data directly between storage and GPU memory.

ROCm 10.1 extends hipFile, its first AIS library, by adding an asynchronous fast-path backend that performs reads and writes on HIP streams, a batch I/O API, and three tiers of I/O statistics that feed the ROCm profilers. The release also adds NUMA-aware host memory placement to the HIP virtual memory APIs.

AMD is solving a well-understood problem. Model checkpoints, KV caches, and weights now routinely exceed GPU memory capacity, and the default Linux I/O path stages every byte through a host-memory bounce buffer before it reaches the accelerator:

• NVIDIA addressed this years ago with GPUDirect Storage and its cuFile API, and the storage industry-built client support around it.
• AIS is AMD’s answer, and hipFile is sufficiently API-compatible with cuFile that AMD publishes a cuFile-to-hipFile mapping in its HIPIFY porting tool.

The storage ecosystem behind AIS is less mature than the software itself. As of this writing, no storage vendor has announced certified hipFile or AIS support for ROCm 10.1.

AMD’s announcement names VAST Data, Dell PowerScale, HPE, and NetApp as vendors supporting NFS over RDMA, the standard used first by AMD’s companion KV-cache tier, and lists WEKA, Lustre, and IBM GPFS as later backends.

VAST Data is the only storage vendor with a publicly available AMD-specific integration and a public performance claim.

Technical Details

AIS lets the storage device read and write GPU memory via direct memory access, so the CPU submits and completes each request without copying the payload to an intermediate buffer.

hipFile is the user-facing library, providing C and Python APIs for synchronous, asynchronous, and batch operations. hipFile reached general availability in ROCm 7.14 in July 2026, and ROCm 10.1 includes hipFile 0.5.0.

The library is open source in the ROCm systems repository.

What’s new in ROCm 10.1

The ROCm 10.1 changes to AIS fall into three groups, along with a related memory placement change in the HIP runtime:

• Asynchronous fast path: hipFileReadAsync() and hipFileWriteAsync() now use the direct path and execute on a HIP stream, enabling I/O to overlap with kernel execution. AMD reports that this reduces latency for checkpoint engines and KV-cache offload while maintaining sustained bandwidth. 
• Batch I/O API: Applications submit many file requests in a single call, and hipFile distributes them across an internal worker-thread pool to keep NVMe queues full during large-dataset reads.
• Multi-tier statistics: hipFileGetStatsL1, L2, and L3 return progressively finer-grained I/O telemetry, and the counters appear in the ROCm profilers so operators can determine whether a pipeline is bound by compute, memory, or data delivery.
• NUMA-aware host memory: New hipMemLocationTypeHostNuma and hipMemLocationTypeHostNumaCurrent location types place host buffers on a specified NUMA node throughout the virtual-memory lifecycle. Libraries such as RCCL can use these to build socket-local, zero-copy buffers on two-socket servers.

Here’s a high-level of how the data path works:

• hipFile scores each request against two backends.
• The fast path accepts a request only when the file is opened with O_DIRECT, the target is GPU device memory, the offsets and sizes satisfy the file system’s direct I/O alignment requirements, and the file is either a block device or a regular file on ext4 with ordered journaling or on XFS.
• Everything else falls back to POSIX pread and pwrite, combined with hipMemcpy, which preserves correctness while maintaining host-staged performance.
• Administrators can force or prevent fallback through environment variables.
• Any block layer between the file system and the device breaks the direct path, so those configurations fall back to compatibility mode.

AMD publishes setup guides for three topologies, covering local NVMe drives, NVMe-oF disks, and NFS over RDMA shares:

• NFS over RDMA requires setting HIPFILE_UNSUPPORTED_FILE_SYSTEMS to true, which means AMD documents the path as outside the default validated envelope.
• On NVIDIA hardware, hipFile delegates to cuFile, enabling a single codebase to target both vendors.

AMD’s roadmap adds hipObject, an RDMA-based object storage library under the AIS umbrella, in the first half of 2027.

Storage Ecosystem Support

Storage ecosystem support for AIS comes through two mechanisms:

• Protocol-level compatibility means that any system exposing NVMe-oF or NFS over RDMA can serve the fast path without vendor-specific code.
• AMD Infinity Context (AIC), the KV-cache storage tier AMD introduced in July 2026, runs on hipFile and AIS and launches with NFS over RDMA as its first backend.

The table below summarizes the public status of named vendors as of October 12, 2026.

Vendor	Status with AIS / hipFile	Announced	Notes
VAST Data	Active integration via AIC	July 24, 2026 expanded AMD collaboration; KV-cache offload with AMD Infinity Context and ROCm	Claims 9x TTFT and 9.7x throughput on MI355X over 800Gbps NFS over RDMA using LMCache; vendor-run tests
Dell PowerScale	Protocol-compatible	Named by AMD as an NFS over RDMA supporter	No AIS-specific announcement or validation published
HPE	Protocol-compatible	Named by AMD as an NFS over RDMA supporter	No AIS-specific announcement or validation published
NetApp	Protocol-compatible	Named by AMD as an NFS over RDMA supporter	No AIS-specific announcement or validation published
WEKA	Roadmap	WekaFS listed by AMD as a later AIC backend	No date given
Lustre (DDN and others)	Roadmap	Lustre listed by AMD as a later AIC backend	No DDN announcement tied to AIS
IBM Storage Scale (GPFS)	Roadmap	GPFS listed by AMD as a later AIC backend	No IBM announcement tied to AIS
Object stores	Roadmap	hipObject targeted for H1 2027	Object over RDMA support follows hipObject

AMD’s NIC support is broader than its storage list:

• AIC supports AMD Pensando AI-NICs and Broadcom BCM57608 (Thor2) adapters today.
• NVIDIA ConnectX support is “planned.”

AMD said storage partner engagements would run through the third quarter of 2026 and that AIC would launch its Infera DI platform in the fourth quarter, so additional vendor announcements are likely to be made around that time.

Analysis

ROCm 10.1 continues AMD’s strategy of closing gaps in the CUDA ecosystem, one subsystem at a time, on a roughly six-week release cadence. 

Direct GPU storage access is a prerequisite for credible large-scale training and for disaggregated inference, and AMD now has a shipping, open-source implementation.

Pairing AIS with the AIC KV-cache tier focuses the work on inference economics, where AMD competes most effectively on memory capacity and cost.

Some notes:

• Open, upstream-first approach: AMD contributes to LMCache, NIXL, and vLLM without a proprietary certification program, lowering the barrier for storage vendors and aligning with AMD’s open-ecosystem message.
• Standards-based storage attach: Leading with NFS over RDMA lets customers reuse existing enterprise NAS, provided it supports RDMA, thereby broadening the addressable base.
• Positioning risk: Relying on protocol compatibility instead of vendor certification leaves customers to validate performance themselves, and enterprise buyers usually expect named, tested configurations before committing.

Practitioner Impact

For teams running checkpoint-heavy training or long-context inference on AMD Instinct GPUs, AIS removes a structural disadvantage against NVIDIA platforms:

• The async fast path and batch API map directly onto checkpoint engines and KV-cache managers.
• cuFile compatibility lowers the porting cost for code already written against GPUDirect Storage.

The operational picture does come with some caveats:

• Storage layout constraints: Many production servers place NVMe behind LVM, MD RAID, or dm-crypt. These layouts fall back to host-staged I/O without an error, so teams must validate with AMD’s ais-check tool and the new statistics counters before assuming direct-path performance.
• Shared storage is in the early stages: The fast path is validated for local ext4 and XFS. Networked file systems depend on NFS over RDMA, enabled via an override flag, and parallel file systems such as WekaFS, Lustre, and Storage Scale are on the roadmap.
• Framework integration: The benefit reaches most users through AMD’s upstream contributions to LMCache, NIXL, and vLLM integrations. Teams on SGLang or other serving stacks wait for later phases.
• Observability gains: The tiered statistics are immediately useful, independent of storage vendor, because they indicate whether a pipeline is data-starved.

Competitive Impact

NVIDIA GPUDirect Storage is the reference point, with years of production use and native client support from nearly every AI storage vendor. NVIDIA has since extended the model to rack-scale KV-cache storage through its STX reference architecture and BlueField-4 DPUs, which a wide set of storage vendors committed to support at GTC 2026.

AMD enters with comparable core plumbing and a much shorter partner list.

The table below compares the main alternatives:

Alternative	Model / Approach	Ecosystem Maturity	Compared to AMD Infinity Storage
AMD Infinity Storage (hipFile)	Open-source direct GPU I/O library; fast path for ext4/XFS on NVMe and NVMe-oF, NFS over RDMA by override	GA July 2026; one public storage integration	Subject offering; cuFile-compatible API and runs on NVIDIA by delegating to cuFile
NVIDIA GPUDirect Storage (cuFile)	Kernel driver plus cuFile API for direct DMA from local and networked storage	Mature; broad parallel FS and NAS client support	Same architectural model with a far larger certified vendor base
NVIDIA STX / BlueField-4	DPU-fronted, rack-scale storage reference architecture for KV-cache extension	Announced 2026; many storage vendors committed	Targets the same KV-cache problem as AIC with hardware offload AMD does not yet match
Vendor client stacks (WEKA, DDN, VAST)	Proprietary clients and protocols with GPU-direct paths	Mature on NVIDIA; AMD support varies	Deliver optimized paths today; AMD coverage depends on each vendor adopting hipFile
Host-staged POSIX I/O	Standard read path through a host bounce buffer	Universal	hipFile’s fallback; works everywhere but carries the bandwidth and CPU cost AIS removes

AMD’s differentiation is strongest in openness and portability because hipFile runs on both vendors’ GPUs and AIC avoids a proprietary storage protocol.

Its weakest point is ecosystem depth, because a buyer evaluating AMD for a large inference cluster today finds only one storage vendor with published AIS results and a roadmap for the parallel file systems that dominate HPC and training deployments.

Final Thoughts

ROCm 10.1 matures AMD Infinity Storage from a basic direct-I/O library into a stream-aware, batch-capable, observable data path that aligns with checkpointing and KV-cache offload. The engineering is solid, the cuFile mapping is pragmatic, and the open, upstream-first model provides storage vendors with a low-friction way to participate.

For customers evaluating AMD Instinct for inference and training at scale, the readiness of the underlying storage is now the gating factor in the AMD platform decision.

AMD has delivered the software layer, and the fourth-quarter AIC launch is where Dell, HPE, NetApp, WEKA, DDN, and IBM must translate protocol compatibility into named, validated configurations so AIS can compete on equal footing with GPUDirect Storage.