Short answer: For CUDA-based distributed training, choose an adapter and fabric that your PyTorch and NCCL stack can use reliably—typically InfiniBand, RoCE, or Ethernet/TCP. PyTorch’s documented recommendation is: “Use the NCCL backend for distributed training with CUDA GPU.” That recommendation does not certify every adapter model. The adapter must also match the server’s PCIe layout, GPU-to-NIC topology, firmware, drivers, operating system, and fabric configuration.
Storage controllers solve a different problem. NVMe devices and filesystems feed training data and save checkpoints; they do not perform the inter-node GPU collectives handled by NCCL. If you want direct GPU-to-storage transfers through GPUDirect Storage (GDS), the requirements become stricter: a compatible NVMe device, a GDS-capable filesystem, 4KB alignment, installed GDS software, and a supported GPU and transport stack.
As an Amazon Associate I earn from qualifying purchases.
What “support” means in a distributed-training system
A distributed job has at least two independent data paths:
Recommended Free Tools
- Inter-node communication: GPUs exchange gradients, parameters, and synchronization messages over a network adapter and fabric. NCCL performs the GPU collectives for CUDA workloads.
- Data and checkpoint I/O: Each host reads training data and writes checkpoints through a storage controller, filesystem, and storage device. This path can use ordinary CPU-mediated I/O or, when all prerequisites are met, GPUDirect Storage.
An adapter is therefore “supported” only in the context of a complete software and hardware combination. A card that works in one server may fail to initialize, select the wrong interface, or deliver poor scaling in another because of PCIe slot placement, IOMMU or ACS settings, firmware, driver versions, GPU affinity, or fabric configuration. The official PyTorch and NVIDIA guidance establishes supported technologies and diagnostic areas, not a universal list of approved adapter SKUs.
#1 Best Overall
- Read Before You Buy — No Video Output: These adapters support charging and USB 2.0 data transfer, but cannot transmit video signals. Except for standard USB webcams (which use USB data only), they are not compatible with HDMI/DisplayPort cables, video-capable USB-C hubs, or docking stations with video output.
- Convert USB-A Ports to USB-C: Designed to connect USB-C earphones, cables, flash drives, card readers, and other USB-C accessories to standard USB-A ports. Plug-and-play with no drivers or software required.
- Aluminum Alloy Housing: Built with a sturdy aluminum alloy shell that aids in heat dissipation and protects against daily wear and scratches. Designed to maintain a stable and secure connection.
- Compact & Travel-Friendly: The ultra-compact design allows the adapter to stay plugged into your device without blocking adjacent ports or adding bulk, reducing wear and tear on your original USB ports.
- 12-Month Warranty: Backed by a 12-month manufacturer warranty for peace of mind. Designed to meet strict quality control standards for reliable everyday performance.
Which network interconnects can NCCL use?
PyTorch’s torch.distributed documentation recommends NCCL for distributed training with CUDA GPUs. It discusses NCCL on both InfiniBand and Ethernet GPU hosts. NVIDIA’s NCCL troubleshooting documentation also covers InfiniBand and RoCE diagnostics. These are the practical interconnect categories to evaluate.
| Interconnect | How it fits distributed training | What must be verified |
|---|---|---|
| InfiniBand | NCCL supports InfiniBand and GPUDirect on GPU hosts. It is a purpose-built low-latency fabric for multi-node collectives. | Adapter and switch compatibility, verbs/RDMA drivers, firmware, subnet and fabric health, GPU-to-NIC affinity, and NCCL’s selected interface. |
| RoCE (RDMA over Converged Ethernet) | Ethernet-based RDMA. NVIDIA documents RoCE-specific NCCL troubleshooting, so it is a valid fabric path when the network is engineered for it. | RDMA-capable adapters, lossless or correctly tuned Ethernet configuration, congestion and priority settings, firmware, driver support, and NCCL fabric diagnostics. |
| Standard Ethernet/TCP | PyTorch recommends NCCL for GPU hosts connected by Ethernet. Gloo is documented as a fallback if NCCL has problems. | Correct IP interface selection, sufficient bandwidth, stable routing, MTU consistency where applicable, driver support, and absence of competing interfaces. |
Do not infer a training-throughput number from the interconnect name alone. Actual scaling depends on GPU count, model communication pattern, message sizes, PCIe and switch topology, adapter configuration, and the NCCL build.
InfiniBand adapters
An InfiniBand adapter is a sensible candidate when the cluster already has an InfiniBand fabric and the server can place the adapter close to the GPUs on the PCIe topology. NCCL must be able to access the adapter through the installed RDMA software, and the fabric must pass link, routing, and subnet checks. A nominally fast adapter is not enough if it is attached to a distant CPU root complex, shares constrained PCIe bandwidth, or runs firmware that does not match the driver stack.
RoCE adapters
RoCE uses Ethernet hardware and switches but still depends on RDMA behavior. Treat the adapter, switch configuration, and host networking as one system. NVIDIA’s troubleshooting material specifically calls out RoCE diagnostics; use those checks rather than assuming that an Ethernet link automatically provides working RDMA. Incorrect congestion or priority configuration can produce timeouts or unstable collective performance even when ordinary IP traffic appears healthy.
Rank #2
- 5-in-1 USB-C Hub: Experience comprehensive connectivity featuring a Power Delivery input, two USB-A 2.0 ports, a USB-A 3.0 port, and an HDMI port. (Note: The USB-C power delivery input port is only for connecting an external wall charger to power your laptop and cannot power peripheral devices.)
- 90W Pass-Through Charging: Achieve optimal charging with 90W pass-through power to your laptop, supported by a total input of 100W, with the hub reserving 10W for operational efficiency. (Note: Wall charger not included.)
- Quick Data Transfers: Accelerate your productivity with rapid data transfers using a high-speed 5Gbps USB 3.0 port and two 480Mbps USB 2.0 ports.
- 4K HDMI Display: Enhance your visual experience with a hub capable of delivering 4K resolution at 30Hz in both mirror and extend modes. Please note that this hub is compatible with MacBook (macOS 12 and newer), Windows 10 and 11, ChromeOS, and laptops equipped with DP Alt Mode and Power Delivery. Note: This device is not compatible with Linux.
- What You Get: Anker USB-C Hub (5-in-1, 4K HDMI), welcome guide, 18-month warranty, and our friendly customer service.
Ethernet adapters for TCP
Conventional Ethernet can carry NCCL traffic over TCP. It is often the simplest option operationally, but interface selection matters on hosts with management, storage, and training networks. PyTorch documents NCCL as the preferred backend for CUDA GPU training on Ethernet and identifies Gloo as a fallback when NCCL cannot be made to work.
Configure NCCL to use the intended adapter
NCCL and Gloo try to select a network interface automatically. Automatic selection is convenient on a simple host, but multi-homed training servers should be checked explicitly. PyTorch documents NCCL_SOCKET_IFNAME for NCCL and GLOO_SOCKET_IFNAME for Gloo when you need to override interface selection.
- Use the NCCL backend for CUDA: Configure the distributed process group to use NCCL, as recommended in PyTorch’s distributed-training documentation.
- Inventory interfaces: Identify the IP interface or RDMA device connected to the training fabric. Exclude management and unrelated storage interfaces from the selection decision.
- Override selection when necessary: Set
NCCL_SOCKET_IFNAMEto the intended interface for NCCL jobs. If you test with Gloo, setGLOO_SOCKET_IFNAMEto the corresponding interface. - Check GPU-to-NIC affinity: Map each GPU to the nearest adapter and PCIe root complex. NVIDIA’s troubleshooting guidance highlights GPU-to-NIC topology, IOMMU, ACS, and related platform settings because they can affect reachability and performance.
- Validate the fabric before a long run: Check adapter links, RDMA or IP connectivity, switch paths, firmware, and driver versions on every node. Run the NCCL interface, latency, bandwidth, and InfiniBand or RoCE diagnostics applicable to your fabric.
- Compare all nodes: A job can fail because one host has a different adapter firmware, driver, interface name, PCIe placement, or disabled link. Keep the communication stack consistent across the cluster.
Environment variables are selection controls, not compatibility fixes. If the adapter cannot be opened by the driver, the fabric is unhealthy, or the topology is unsuitable, forcing its interface will not solve the underlying problem.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesAdapter compatibility checklist
Before buying or deploying a network adapter, verify each item against the exact server and software image:
Rank #3
- Sleek 7-in-1 USB-C Hub: Features an HDMI port, two USB-A 3.0 ports, and a USB-C data port, each providing 5Gbps transfer speeds. It also includes a USB-C PD input port for charging up to 100W and dual SD and TF card slots, all in a compact design.
- Flawless 4K@60Hz Video with HDMI: Delivers exceptional clarity and smoothness with its 4K@60Hz HDMI port, making it ideal for high-definition presentations and entertainment. (Note: Only the HDMI port supports video projection; the USB-C port is for data transfer only.)
- Double Up on Efficiency: The two USB-A 3.0 ports and a USB-C port support a fast 5Gbps data rate, significantly boosting your transfer speeds and improving productivity.
- Fast and Reliable 85W Charging: Offers high-capacity, speedy charging for laptops up to 85W, so you spend less time tethered to an outlet and more time being productive.
- What You Get: Anker USB-C Hub (7-in-1), welcome guide, 18-month warranty, and our friendly customer service.
- Host fit: PCIe generation and lane width, slot height and power, NUMA placement, BIOS support, and available slot bandwidth.
- GPU topology: Distance from each GPU, PCIe switch arrangement, peer-to-peer paths, IOMMU and ACS behavior, and whether the adapter shares a bottlenecked root complex.
- Fabric type: InfiniBand, RoCE, or TCP Ethernet, including switch and cabling requirements.
- Software: Operating-system kernel, adapter driver, RDMA libraries where applicable, NCCL version, and PyTorch version.
- Firmware: Adapter and switch firmware versions that are supported by the driver stack.
- Interface selection: Stable interface names and a documented
NCCL_SOCKET_IFNAMEpolicy on multi-network hosts. - Diagnostics: A repeatable test plan for link state, reachability, latency, bandwidth, and NCCL initialization on every node.
The official PyTorch and NVIDIA documentation does not provide a universal adapter-model matrix, so a model-specific answer requires the GPU type, server model, operating system, fabric, candidate part number, and software versions.
What storage controllers do—and do not do
Storage controllers are not substitutes for network adapters. An NVMe controller can deliver local datasets and checkpoints to a host, but NCCL still uses the network fabric for inter-node collectives. Distributed jobs can therefore have a well-supported network adapter and an entirely separate storage bottleneck.
Ordinary NVMe data and checkpoint I/O
For conventional I/O, select a storage controller and NVMe device that the operating system and filesystem support reliably, then size capacity and endurance for the dataset and checkpoint pattern. Local NVMe can reduce contention with shared storage, but it does not provide direct GPU access unless the GDS requirements below are satisfied.
GPUDirect Storage
PyTorch’s GPUDirect Storage tutorial describes torch.cuda.gds APIs introduced in PyTorch 2.7 and marks the workflow as a prototype. The documented path requires PyTorch 2.7.0 or later, GDS installed according to NVIDIA’s documentation, a filesystem that supports GDS, and 4KB storage alignment.
Rank #4
- Dual Converters, Infinite Potential：Includes 2× USB C male to USB A female adapters and 2× USB A male to USB C female adapters. Perfect for a wide range of uses—tablets with Bluetooth keyboards, expand USB ports on macbook, and more. Two different converters for all your daily needs
- Next-Level 10Gbps & 3A Charging: No more slow 480Mbps, this usb to usb c adapter has a transfer speed of up to 10Gbps, allowing you to do more transferring in less time. This usb adapter fits both USB A and USB C charger, supporting up to 3A fast charging
- Upgraded Exquisite Craftsmanship: With an aluminum alloy housing and metal connector, the usbc to usb adapter is extremely durable and sturdy. Rigorously tested to withstand more than 10,000 times of plugging and unplugging, ensuring long-lasting performance
- Broad Compatible: The usb c to usb adapter widely supports all USB C/ USB A devices like laptops, tablets, cellphones, car chargers, and phone chargers. Such as compatible with MacBook Pro/Air 2023/2022, Thunderbolt 4/3 Devices,Apple MagSafe Watch 9/8/7/SE/Ultra, iPad Pro 2022/2021, Samsung Galaxy S23/S20/S10, and iPhone 17/16/15 Pro. Plug and play
- Please Note: To reach 10Gbps speed, keep the cable under 3.3 ft. For USB A Male to USB C adapters, try flipping the USB C connector. USB C Male to USB A adapters support bidirectional 10Gbps transfer within 3.3 ft
NVIDIA’s GDS guide adds device-level restrictions: the NVMe device must be GDS-compatible and must not have block-device integrity capability. In the relevant block-device section, EXT4 and XFS are supported filesystems; for the described Direct I/O path, EXT4 must be mounted with journaling mode data=ordered. These conditions apply to the particular GDS workflow, not to distributed training in general.
| Storage path | Required components | Typical use |
|---|---|---|
| Standard filesystem I/O | OS-supported controller, NVMe device, filesystem, and adequate CPU, memory, and PCIe capacity. | Loading datasets and saving checkpoints through normal host I/O. |
| GDS direct GPU-to-storage I/O | PyTorch 2.7.0 or later for the documented APIs, installed GDS software, GDS-capable filesystem, 4KB alignment, GDS-compatible NVMe device without block-device integrity capability, and compatible GPU/platform drivers. | Reducing host involvement when reading training data or writing checkpoints directly between GPU memory and storage. |
GDS platform and transport details
NVIDIA lists A100, A40, L4, L40S, H100, H200, B100, B200, and B300 for its NVMe PCI P2PDMA section. This is a specific GDS P2PDMA compatibility list; it is not a general list of GPUs capable of distributed training.
For NVMe over Fabrics, NVIDIA describes additional storage transport packages. The guide recommends DOCA for new GDS installations and describes MLNX_OFED for applicable existing installations. NVMe-oF therefore adds another compatibility layer: the storage transport, target, initiator, controller, filesystem, and GDS release must all align.
Free tools Windows power users keep installed
One-click scans. No signup required.
Putting the network and storage paths together
- Define the workload: Record GPU model and count per node, number of nodes, expected dataset access pattern, checkpoint frequency, and whether direct GPU-to-storage I/O is actually needed.
- Choose the communication fabric: Select InfiniBand, RoCE, or TCP Ethernet based on existing infrastructure and operational capability, not on the adapter label alone.
- Map topology: Document GPU, NIC, PCIe switch, NUMA, and storage-controller placement on every server.
- Standardize software: Align PyTorch, NCCL, kernel, adapter drivers, firmware, RDMA packages where applicable, filesystem, and GDS components.
- Configure interface selection: Confirm automatic NCCL selection or set
NCCL_SOCKET_IFNAME; configureGLOO_SOCKET_IFNAMEonly when using Gloo tests or fallback jobs. - Test communication independently: Run NCCL fabric and topology diagnostics before introducing dataset and checkpoint traffic.
- Test storage independently: Verify filesystem behavior, alignment, NVMe compatibility, and checkpoint read/write reliability. Enable GDS only after every documented prerequisite is met.
- Run a small multi-node job: Confirm that all ranks initialize, collectives complete, and checkpoint files are readable before scaling to a production run.
Troubleshooting common failures
Ranks cannot initialize
Check that every node sees the intended interface, that NCCL_SOCKET_IFNAME is not selecting a management network, and that firewall or routing rules permit node-to-node traffic. For InfiniBand or RoCE, verify link state, RDMA device visibility, firmware, and fabric diagnostics.
Best Value
- 5-in-1 Connectivity: Equipped with a 4K HDMI port, a 5 Gbps USB-C data port, two 5 Gbps USB-A ports, and a USB C 100W PD-IN port. Note: The USB C 100W PD-IN port supports only charging and does not support data transfer devices such as headphones or speakers.
- Powerful Pass-Through Charging: Supports up to 85W pass-through charging so you can power up your laptop while you use the hub. Note: Pass-through charging requires a charger (not included). Note: To achieve full power for iPad, we recommend using a 45W wall charger.
- Transfer Files in Seconds: Move files to and from your laptop at speeds of up to 5 Gbps via the USB-C and USB-A data ports. Note: The USB C 5Gbps Data port does not support video output.
- HD Display: Connect to the HDMI port to stream or mirror content to an external monitor in resolutions of up to 4K@30Hz. Note: The USB-C ports do not support video output.
- What You Get: Anker 332 USB-C Hub (5-in-1), welcome guide, our worry-free 18-month warranty, and friendly customer service.
NCCL works but scaling is unexpectedly poor
Inspect GPU-to-NIC and PCIe affinity, NUMA placement, IOMMU and ACS settings, switch paths, congestion, and whether one node has a slower or differently configured adapter. A successful connection does not prove that the topology is optimal.
RoCE jobs time out
Use the RoCE-specific diagnostics in NVIDIA’s NCCL troubleshooting guidance. Check the Ethernet fabric’s congestion and priority configuration, adapter and switch firmware, MTU consistency, and RDMA driver state instead of treating the link as ordinary TCP Ethernet.
GDS initialization fails
Verify the PyTorch version, installed GDS software, filesystem support, 4KB alignment, NVMe-device compatibility, and absence of block-device integrity capability. Confirm that the GPU and platform are covered by the applicable GDS documentation. If any condition is uncertain, use standard filesystem I/O rather than claiming the device is GDS-compatible.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11CUDA versus ROCm considerations
The recommendations above are scoped to CUDA, PyTorch, NCCL, and NVIDIA GDS. PyTorch’s ROCm documentation says ROCm distributed currently supports NCCL and Gloo and describes hipFile as its GDS implementation. Do not apply CUDA/cuFile instructions to a ROCm deployment without checking the ROCm-specific documentation and compatibility requirements.
The practical answer
There is no single “distributed-AI adapter” or “training storage controller.” For CUDA, start with NCCL and select an InfiniBand, RoCE, or Ethernet adapter that the complete host, driver, firmware, topology, and fabric stack supports. Use storage controllers separately for datasets and checkpoints. Choose GDS only when the exact NVMe device, filesystem, alignment, GPU/platform, and software transport requirements have been verified. Without those details, a part-number recommendation would be guesswork rather than a reliable compatibility answer.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
