matlab-setup-gpu

$npx mdskill add matlab/matlab-agentic-toolkit/matlab-setup-gpu

Detects and validates GPU availability for MATLAB

  • User checks if their system can run GPU code in MATLAB.
  • Depends on MATLAB-native APIs like `canUseGPU` and `gpuDeviceTable`.
  • Decides based on the presence of compatible GPUs detected by `gpuDevice`.
  • Delivers results through `canUseGPU` status and `gpuDeviceCount` properties.

SKILL.md

.github/skills/matlab-setup-gpuView on GitHub ↗
---
name: matlab-setup-gpu
description: >
  Detect and validate GPU availability for MATLAB GPU computing.
  Use when the user can't use their GPU, or is setting up or
  selecting a GPU. Triggers on: gpuDevice, GPU setup, check GPU, GPU
  not found, GPU not working, can't use GPU, GPU not available,
  unable to find a supported GPU device, compatible GPU,
  canUseGPU, validateGPU.
license: MathWorks BSD-3-Clause
compatibility: ">=R2024b"
metadata:
  author: MathWorks
  version: "1.0"
---

# Set Up GPU for MATLAB

Detect whether a compatible GPU is available, validate it works, and diagnose common failure modes.

## When to Use

- User wants to check if their system can run GPU code in MATLAB
- User gets errors when calling `gpuDevice` or `gpuArray`
- User wants to set up or select a specific GPU in a multi-GPU system
- User needs to troubleshoot why GPU computing isn't working
- As a prerequisite check before any GPU-accelerated workflow

## When NOT to Use

- Writing or optimizing GPU-accelerated code — this skill only detects and diagnoses GPU availability
- Monitoring GPU memory or utilization during computation — use `gpuDevice` properties directly or NVIDIA tools
- Distributing work across multiple GPUs with parallel pools — use `parpool` workflows

## Conventions

- Use MATLAB-native APIs (`canUseGPU`, `gpuDeviceTable`, `gpuDevice`) for detection and properties — reserve `nvidia-smi` for information MATLAB cannot provide
- Prefer `canUseGPU` over `gpuDeviceCount == 0` or `isempty(gpuDeviceTable)` for availability checks — `canUseGPU` also verifies licensing and toolbox state
- If recommending a driver update, fetch the **GPU Computing Requirements** documentation page for the user's MATLAB release to check the driver selection policy

## Workflow

### 1. Is a GPU available?

```matlab
canUseGPU
```

Returns `true` if PCT is installed, licensed, GPU support is present, and a device is available. If true, proceed to step 3 (list available devices).

### 2. If `canUseGPU` returns `false` — diagnose

```matlab
validateGPU("all")
```

**If the output contains "requires a newer graphics driver":** this does not guarantee a newer driver exists. Do not state that the GPU is unsupported or end-of-life — direct the user to verify on the **GPU Computing Requirements** page for this MATLAB release before drawing any conclusion about hardware compatibility.

For all other failures, interpret using the **Troubleshooting** table.

### 3. List available devices

```matlab
gpuDeviceTable          % List all GPUs and indices (no memory allocation)
```

Always run `gpuDeviceTable` before selecting a device — it confirms which devices are visible and their MATLAB indices.

### 4. Select device and query properties

```matlab
d = gpuDevice(N);       % Select GPU N
properties(d)           % List available device properties
```

### 5. Check driver model (Windows only)

```matlab
d = gpuDevice;              % returns device already selected
d.DriverModel
d.KernelExecutionTimeout
```

If WDDM: the GPU is shared with the display — long-running kernels may be killed by the OS watchdog, and the GPU may exhibit unstable behaviour when low on memory. If `d.KernelExecutionTimeout` is `true`, kernel timeouts are active.

TCC or MCDM avoid these issues but are only available on select professional and data centre GPUs not connected to a display. If switching is not an option, and user experiences kernel timeouts, advise: increase TDR timeout in Windows Registry, close other GPU-accelerated apps, or break computation into smaller chunks.

### 6. Benchmark GPU performance (optional)

To qualify the GPU's computing power, suggest the user run `gpuBench` (available from the MATLAB File Exchange, not shipped with MATLAB). It benchmarks memory bandwidth, single/double precision compute, compares GPU vs CPU performance, and provides comparison data against other popular GPUs.

## Troubleshooting

| Symptom | Cause | Resolution |
|---------|-------|------------|
| `canUseGPU` returns `false`, no error | PCT not licensed | Check `license('test','Distrib_Computing_Toolbox')`. If false, user needs a PCT license. |
| `validateGPU` not found | PCT not installed or pre-R2024b | Check `ver('parallel')` — if empty, PCT is not installed; if version is below 24.2 (R2024b), `validateGPU` is not available in that release. |
| "requires a newer graphics driver" or validateGPU: platform or driver check FAILED | NVIDIA driver too old, or driver support for the GPU architecture may have ended | Does not guarantee a newer driver exists. Consult the **GPU Computing Requirements** documentation page for the user's release to confirm the GPU is still supported before advising a driver update. |
| validateGPU: "Device available" FAILED | Prohibited or exclusive compute mode | Check the message: if "prohibited" — vGPU graphics-only profile ("B" suffix, e.g. A16-2B), requires compute-enabled profile (Q or C suffix), advise user to escalate to system administrator. If "exclusive" — GPU in use by another process, free it or select a different device. |
| validateGPU: device count is 0 | Device not visible to MATLAB | If running in a container, verify GPU passthrough is configured. Otherwise, consult the **GPU Computing Requirements** documentation page for the user's release to confirm supported hardware. |
| GPU kernel killed mid-computation (Windows) | WDDM driver model with TDR active | See step 5 for driver model implications and remediation |

## Key Functions

| Function | Purpose |
|----------|---------|
| `canUseGPU` | Quick availability check — true if PCT licensed and a usable GPU device exists |
| `validateGPU` | Structured diagnostic — tests each layer and reports results |
| `gpuDeviceTable` | List all visible GPUs with indices, names, and status (no memory allocation) |
| `gpuDevice` | Select a GPU by index, or return the current device (no argument) |

----

Copyright 2026 The MathWorks, Inc.

----

More from matlab/matlab-agentic-toolkit

SkillDescription
matlab-access-datafeed>
matlab-add-awgnRead BEFORE writing any code that adds Additive White Gaussian Noise (AWGN) to signals and converts between SNR, Eb/No, Es/No, and per-subcarrier SNR for communications simulations, using awgn(), convertSNR(), berawgn(). The default MATLAB patterns for AWGN (e.g., 'measured' option, manual SNR formulas) produce subtly incorrect results. This skill specifies the correct calling conventions, required function usage, and critical anti-patterns that must be avoided.
matlab-analyze-ams-waveformAnalyze AMS waveform data using Mixed-Signal Blockset utilities: phase noise measurement, clock jitter, anti-aliased resampling, timing measurements, lock time, INL/DNL, ADC/DAC calibration, HSpice import. Use when analyzing time-domain voltage from PLL/VCO/clock simulations, measuring phase noise from variable-step solver output, computing jitter, or resampling non-uniform data.
matlab-analyze-dataAnalyze data using MATLAB. Use when the task involves tables, timetables, time-series data, numeric arrays, sensor matrices, or gridded data — including but not limited to exploring, filtering, sorting, cleaning, transforming, aggregating, smoothing, padding, trimming, and answering questions about data. MATLAB provides extensive, easy-to-use built-in functions for these workflows with no additional products required.
matlab-analyze-dependenciesAnalyze the effective toolbox file set to produce a Dependency Manifest — classify all transitive dependencies as included, product, add-on, or external-unresolved, then present resolution options with tradeoffs. Use after matlab-define-toolbox-api when the spec is approved.
matlab-analyze-emS-parameters, insertion loss, fields, currents, mesh control, and solver selection for RF PCB performance validation. TRIGGER: user asks to compute S-parameters, analyze insertion/return loss, extract fields or currents, compare MoM vs FEM, or control mesh for any RF PCB component. Invoke BEFORE writing sparameters() or solver code — API is non-obvious. SKIP: designing or creating components (use the specific matlab-design-pcb-* skill), material/stackup setup only (use matlab-manage-pcb-material), optimization sweeps (use matlab-optimize-pcb-design), PDN/IR-drop analysis (use matlab-analyze-pcb-pdn).
matlab-analyze-installed-antennaAnalyze antennas installed on electrically large conducting platforms using MATLAB Antenna Toolbox. Loads platform geometry from STL/STEP/IGES, installs antenna elements, selects electromagnetic solvers (MoM-PO, FMM, MoM), and computes patterns, impedance, coupling, and efficiency. Use when the user wants to model an antenna on a vehicle, aircraft, ship, satellite, or other large structure.
matlab-analyze-pcb-pdnPDN DC voltage/current analysis, IR drop, design rule checking, and multi-net batch analysis on imported PCB layouts. TRIGGER: user asks about power integrity, PDN analysis, IR drop, voltage distribution, current density, power nets, or design rule checking on a PCB. Invoke BEFORE writing code — the PDN API chain is specialized and non-obvious. SKIP: importing a PCB file (use matlab-read-pcb-layout), EM field/S-parameter extraction (use matlab-analyze-em), material/stackup setup only (use matlab-manage-pcb-material), transmission line design (use matlab-design-pcb-txline).
matlab-analyze-rcsCalculate and visualize monostatic and bistatic radar cross section (RCS) using MATLAB Antenna Toolbox. Computes RCS of platforms, antennas, and arrays with PO, MoM, and FMM solvers, supporting HH/VV/HV/VH polarization, GPU acceleration, and near-field observation. Use when the user wants to compute, plot, or analyze radar cross section.
matlab-analyze-rf-propagationAnalyze RF propagation and plan wireless sites using MATLAB Antenna Toolbox. Creates transmitter/receiver sites, computes signal strength, coverage maps, SINR, line-of-sight, and ray tracing in geographic or indoor environments. Supports multiple propagation models (free-space, close-in, Longley-Rice, ray tracing, rain/gas/fog), custom terrain, building data, and directional antennas. Use when the user wants to compute coverage, signal strength, path loss, SINR, ray tracing, or plan a wireless network.