Matlab Setup Gpu Agent Skill

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?

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

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

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

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

5. Check driver model (Windows only)

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.

----