Free tools Windows power users keep installed
One-click scans. No signup required.
Setting up NVLink on Tesla V100 hardware is primarily a compatibility and topology task, not a software switch. First identify the exact V100 form factor and whether the GPUs run on bare metal, through PCIe passthrough, or as vGPUs. Confirm that the host actually provides a supported direct NVLink path, inspect it with nvidia-smi topo -m, and query peer-to-peer capability with nvidia-smi topo -p2p n. Only then configure your inference framework’s multi-GPU mode and benchmark the real model. A topology report alone does not prove that every link is active or predict inference performance.
What NVLink changes in a V100 inference system
NVLink is a physical interconnect between compatible GPU components. It can provide a direct path for GPU-to-GPU communication, but it does not automatically distribute an inference model, select devices, or turn on a framework’s model-parallel mode. Those are separate software configuration tasks.
As an Amazon Associate I earn from qualifying purchases.
The first decision is therefore whether the installed hardware can have NVLink at all. A generic Tesla V100 name is not enough to establish the connection. The board form factor, server design, placement of the GPUs, bridge or backplane arrangement, and deployment mode all matter. NVIDIA’s validated vGPU guidance specifically lists Tesla V100 SXM2 and Tesla V100 SXM2 32GB configurations for NVLink peer-to-peer support. That listing should not be treated as a universal compatibility statement for every V100 card or every server.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Think of the setup as four independent checks:
- Physical compatibility: the exact V100 variant and host must provide a supported NVLink path.
- System visibility: the driver must expose the installed GPUs and their reported relationships.
- Deployment support: a vGPU assignment has additional profile, hypervisor, and guest-OS requirements.
- Application behavior: the inference framework must be configured for the desired multi-GPU strategy and tested with the target workload.
Failure at any one of these layers can leave you with a working CUDA device that is not using NVLink for inference.
#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.
Check the V100 form factor and host before changing software
Record the physical and virtual details of the system before installing packages or changing framework settings. This prevents a common mistake: trying to solve a missing physical connection with a driver option.
| Item to record | Why it matters | What is established |
|---|---|---|
| V100 form factor | Different board and module designs do not have identical connectivity. | NVIDIA’s cited vGPU table names V100 SXM2 and V100 SXM2 32GB for the documented NVLink peer-to-peer path. |
| Host or server model | The chassis, baseboard, carrier, and GPU placement determine whether a direct link is physically possible. | A generic V100 model number does not establish host-level fit. |
| Deployment mode | Bare metal, PCIe passthrough, and vGPU expose different support boundaries. | vGPU requires validation of the assigned profile, hypervisor, and guest OS together. |
| GPU assignment | Inference can only use the devices visible to the process or virtual machine. | The exact assignment is system-specific and is not defined by topology commands alone. |
| Driver and software versions | Support and reporting depend on the installed stack. | No universal driver, CUDA, or framework version is specified here; use the applicable platform documentation. |
For a bare-metal build, consult the server vendor’s documentation for the exact board and slot arrangement. For passthrough, verify that the guest receives the intended physical GPUs and that the host configuration preserves the supported peer path. For vGPU, do not assume that a physical NVLink connection is exposed identically inside every virtual machine.
Understand the three deployment cases
Bare metal
On bare metal, the operating system and CUDA stack see the physical GPUs directly. Your checklist is the physical V100 variant, the host topology, driver state, and the inference framework’s peer-access and placement behavior. NVLink does not replace framework configuration: the application still needs model parallelism, data parallelism, or another supported distribution method.
PCIe passthrough
Passthrough assigns physical devices to a guest rather than exposing a managed vGPU profile. The practical question is whether the guest and hypervisor preserve the direct path your hardware provides. Validate the physical topology on the host and the devices visible in the guest, then run the topology and peer-capability checks in the environment where inference will execute.
vGPU
vGPU is a separate support case. NVIDIA’s cited validated-platform guidance describes a documented V100 NVLink peer-to-peer path for named V100 SXM2 profiles, Linux guests, and direct connections. The same guidance states limitations for that path, including no NVSwitch and no PCIe peer-to-peer support. These restrictions apply to the cited vGPU configuration; they should not be generalized to bare-metal CUDA without checking the relevant documentation for that stack.
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.
For a vGPU deployment, record the physical GPU model, vGPU profile, hypervisor release, guest operating system, and connection type as one compatibility set. A supported physical host does not by itself prove that a particular profile and guest combination supports peer access.
Inspect the installed GPU topology
Use NVIDIA’s topology matrix
Run the following command from the host or from the guest when the deployment documentation tells you to inspect the virtual topology:
nvidia-smi topo -mNVIDIA documents this command as the way to determine the topology relationships among GPUs. Read the GPU-to-GPU entries rather than looking only for a successful command exit. The matrix shows how the devices are reported as connected, allowing you to compare the expected direct path with the path the system actually exposes.
Capture the complete output with the machine identity and software versions. This makes it possible to distinguish a topology change from a change in GPU assignment or driver state. If the expected GPUs are missing, resolve device visibility first; an inference framework cannot use a device that the execution environment does not expose.
Query peer-to-peer capability
Run:
nvidia-smi topo -p2p nThis queries NVLink peer-to-peer capability between GPUs. Treat the result as a capability check, not as a bandwidth test. It tells you whether the reported pair is eligible for the queried peer relationship in that environment; it does not configure the inference framework or guarantee that a particular model transfer will use the link.
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.
Run both commands after the final GPU assignment is in place. A host matrix and a guest matrix can describe different visible devices, especially in virtualized deployments.
Apply the support conditions for vGPU deployments
If the system uses vGPU, verify all of the following before interpreting a positive topology result as usable inference support:
- The physical accelerator is one of the V100 configurations covered by the applicable NVIDIA support table; the cited table names V100 SXM2 and V100 SXM2 32GB for its documented NVLink peer-to-peer path.
- The selected vGPU profile is one of the profiles covered by that documentation.
- The hypervisor release is supported for the profile and feature combination.
- The guest operating system meets the documented requirement; the cited validated path specifies Linux guest support.
- The GPUs have the direct connection required by that path.
- You are not relying on NVSwitch or PCIe peer-to-peer for a path whose stated limitations exclude them.
NVIDIA’s Virtual GPU Software User Guide instructs readers to determine the NVLink topology between physical GPUs in a host or vGPUs assigned to a virtual machine by running nvidia-smi topo -m from the host or VM. Use that command in the supported context, then verify the profile and platform conditions separately. A command that returns a matrix cannot override an unsupported profile or guest configuration.
Configure multi-GPU inference after topology is validated
NVLink topology does not select GPUs for an inference process. Configure the framework you actually use according to its supported distribution model. The exact setting depends on whether the application uses model parallelism, pipeline parallelism, data parallelism, or a framework-specific inference service.
Model-parallel inference
Model parallelism places different portions of one model on different GPUs. In this arrangement, communication between stages can be sensitive to placement and transfer patterns. Define the device map or equivalent placement in the framework, then confirm that every intended GPU is visible to the process. Keep the placement fixed while comparing a topology with and without the direct path.
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
Data-parallel inference
Data parallelism keeps a copy of the model on multiple GPUs and divides requests or batches among them. The communication pattern differs from model parallelism, so a topology that looks favorable for one workload may not deliver the same result for another. Configure the framework’s supported worker or process layout and measure request throughput and latency under the same load.
Mixed or service-level distribution
Some inference services assign complete requests to separate GPU workers instead of splitting one model. In that case, NVLink may be less central than GPU placement, memory capacity, and request scheduling. Do not assume that installing a bridge or exposing a topology entry changes the service’s scheduling policy.
Because the applicable framework, model, precision, and software versions are unspecified, there is no single framework command that can be prescribed safely here. Use the framework’s documented multi-GPU method, then verify device placement in its own logs or diagnostics.
Benchmark the workload instead of promising a speedup
No universal V100 inference speedup figure is established for NVLink. Performance depends on the model’s communication volume, batch size, precision, memory transfers, GPU placement, framework implementation, driver and CUDA versions, and whether the workload is latency- or throughput-bound.
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 & 11Create a controlled comparison. Keep the following variables constant:
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.
- Model weights and model version.
- Input shape and sequence or image dimensions.
- Batch size and request concurrency.
- Precision and any quantization settings.
- Number of GPUs and their process or device placement.
- Driver, CUDA, framework, and application versions.
- Warm-up procedure and measurement interval.
Record at least the following outcomes:
| Measurement | Use |
|---|---|
| End-to-end latency | Shows the time a caller experiences for an individual request or defined batch. |
| Throughput | Shows completed requests, samples, or tokens per unit of time under a stated load. |
| GPU memory use | Helps identify whether placement or replication, rather than interconnect traffic, is limiting the run. |
| GPU utilization over time | Shows whether devices are progressing together or waiting during communication. |
| Request concurrency | Provides the load context needed to interpret latency and throughput. |
Run enough repetitions to separate startup effects from steady-state behavior, and report the exact configuration with the result. A topology matrix is evidence about connectivity relationships; it is not a benchmark and cannot substitute for this comparison.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Diagnose an unexpected topology or disappointing result
The GPUs do not appear as expected
- Confirm that the intended GPUs are visible in the environment where inference runs.
- Check whether a VM was assigned the expected devices and profile.
- Recheck the physical board and server documentation for the exact V100 form factor.
- Review driver state and restart only after recording the current topology output.
If the hardware path is absent, no inference setting can create NVLink. A different framework configuration may change communication behavior, but it cannot add a missing physical interconnect.
The matrix shows a path, but peer access is unavailable
Run nvidia-smi topo -p2p n and compare its result with the matrix. In a vGPU deployment, recheck the profile, hypervisor, Linux guest requirement, and direct-connection limitation together. Do not infer that a relationship shown by one view automatically grants every peer operation to the application.
Topology looks correct, but inference is not faster
First verify that the framework is actually using multiple GPUs in the intended mode. Then compare model placement, batch size, precision, and concurrency with the baseline. A workload with little inter-GPU communication may gain little from NVLink, while a poorly placed model may spend time elsewhere. The absence of a measured improvement is not proof that the link is inactive.
A link appears inactive or unstable
Use NVIDIA’s separate link-status and diagnostic information rather than relying on the topology matrix alone. NVIDIA’s Data Center GPU Manager documentation cautions that topology records known relationships but does not itself prove active link state or measure bandwidth. Check the host’s physical connection, device seating or carrier arrangement, and driver condition according to the platform vendor’s procedures. Preserve diagnostic output before making changes so that a hardware or software state can be compared afterward.
Use this deployment decision checklist
| Question | If yes | If no |
|---|---|---|
| Is the exact V100 form factor documented for the intended NVLink path? | Continue to host-topology checks. | Stop and verify compatibility; do not assume the card supports the path. |
Does nvidia-smi topo -m show the expected GPU relationships? |
Run the peer-capability query. | Investigate visibility, physical layout, and deployment assignment. |
Does nvidia-smi topo -p2p n report the needed capability? |
Configure the framework’s multi-GPU mode. | Review support conditions and driver or virtualization state. |
| Is the deployment vGPU? | Validate profile, hypervisor, Linux guest, and direct-connection requirements. | Apply the bare-metal or passthrough platform documentation. |
| Has the target model been benchmarked with fixed variables? | Use the measured latency and throughput for the capacity decision. | Do not claim a speedup from topology alone. |
What to document for a reproducible setup
Keep a short record that another administrator can reproduce:
- Exact V100 model and form factor, including whether it is an SXM2 or SXM2 32GB configuration.
- Server or host model and the physical GPU arrangement.
- Bare-metal, passthrough, or vGPU deployment mode.
- For vGPU: profile, hypervisor, guest OS, and direct-connection assumptions.
- Driver, CUDA, framework, and inference-service versions.
- Full output from
nvidia-smi topo -mandnvidia-smi topo -p2p n. - Framework device placement and multi-GPU strategy.
- Model, precision, batch size, concurrency, warm-up method, latency, throughput, and memory observations.
- Any separate link-status or diagnostic results used to investigate activity.
This record separates a physical compatibility issue from a virtual assignment problem, a framework placement issue, and a workload characteristic. It also prevents a one-time topology observation from being mistaken for a guaranteed performance property.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteQuick Recap
Common mistakes to avoid
- Assuming every V100 is equivalent: form factor and host topology are decisive.
- Buying a generic accelerator and expecting NVLink: the card, server, and connection must be matched as a supported configuration.
- Treating
topo -mas a speed test: it reports relationships, not application bandwidth or inference results. - Skipping the peer-capability query: a visible relationship does not replace an explicit P2P check.
- Applying vGPU rules to bare metal: vGPU profile, hypervisor, and guest restrictions are deployment-specific.
- Changing several variables at once: keep model, batch, precision, placement, and software versions fixed when benchmarking.
- Promising a universal speedup: no general V100 inference improvement is established; measure the target workload.
Final setup sequence
- Identify the V100 form factor, host, GPU arrangement, and deployment mode.
- Confirm the exact platform and, for vGPU, profile, hypervisor, and Linux guest support conditions.
- Run
nvidia-smi topo -min the execution environment and save the output. - Run
nvidia-smi topo -p2p nto query peer capability. - Resolve unexpected visibility or support results before configuring the model.
- Configure the inference framework’s documented multi-GPU strategy and device placement.
- Check separate link-status or diagnostic information when activity is in doubt.
- Benchmark the actual model with fixed variables and use those measurements—not topology alone—to decide whether the setup meets its latency or throughput goal.
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.
