Stress test

The stress test pushes your CPU, GPU, or VRAM to maximum load for an extended period. Unlike the performance benchmark, which measures speed, the stress test measures stability: whether your hardware can sustain peak performance without overheating, throttling, or crashing.

Selecting devices to stress

The stress test presents a hierarchical device tree showing all compatible hardware in your system. You can select any combination of:

• CPU Stress: sustains all processor cores at 100% utilization, exposing instability from inadequate cooling, insufficient power delivery, or aggressive overclocks
• GPU Stress: runs an FMA-intensive compute shader that saturates shader cores, memory bandwidth, and power delivery
• VRAM Stress: cycles through five distinct bit patterns across the full device-local VRAM allocation to detect stuck bits, coupling faults, and address errors

If your system has multiple GPUs, each appears as a separate branch in the tree with its own GPU Stress and VRAM Stress options. Select one, several, or all devices depending on what you want to validate.

Duration

The stress test duration is configurable from 1 minute to 8 hours.

Start short and increase only if needed. A few minutes is enough to catch most instability from overclocks or cooling problems. If that passes and you want more confidence, extend to 15–30 minutes. Longer runs of several hours are useful for used hardware validation or diagnosing intermittent issues, but sustained maximum load contributes to component wear and tear.

Sequential vs. parallel execution

The stress test supports two execution modes.

Sequential (default)

Sequential mode tests one device at a time. If you select CPU, GPU, and VRAM, Novabench stresses the CPU first, then the GPU, then the VRAM. Each device runs for the full configured duration.

Sequential mode isolates each component, making it easier to identify which device has a stability issue. This mode is available on all plans.

Parallel (Pro)

Parallel mode stresses all selected devices simultaneously. This creates a worst-case power and thermal load, which is useful for testing system-level stability: power supply capacity, VRM headroom, overall cooling, and system-wide thermal throttling.

Parallel mode requires Pro or higher.

Before you start

When you launch a stress test for the first time, Novabench shows a caution notice reminding you that stress testing deliberately pushes hardware to its limits. Make sure your system has adequate cooling before running extended tests, especially on laptops or passively cooled systems.

This notice appears once and is not shown again on subsequent runs.

During the test

While the stress test runs, a full-screen progress view shows:

• Current device: which component is being stressed (in sequential mode)
• Time remaining: countdown for the current device and overall test
• Component progress: in sequential mode, shows "X of Y" to indicate how many devices have completed

You can cancel the stress test at any time. If a device fails (crashes, triggers a driver reset, or encounters an error), Novabench records the failure point and moves to the next device (sequential) or stops (parallel).

Reading your results

After the stress test completes (or is cancelled), the results screen shows a summary for each tested device.

Pass or fail

Each device receives a Pass or Fail verdict:

• Pass: the device completed the full duration without errors or driver resets
• Fail: the device encountered a problem before the duration expired. The results show the failure timestamp (for example, "Fail @ 2m 30s") indicating how long the device ran before failing.

An overall pass/fail indicator at the top summarizes whether all devices passed.

Per-device metrics

For each device, the results display:

Understanding throttle behavior

The throttle time metric works by finding the peak frequency during the test, then measuring how much the frequency dropped afterward. A large gap between max frequency and min frequency, combined with significant throttle time, indicates that the device could not sustain its peak performance.

For example, a CPU that reaches 5,000 MHz but drops to 4,200 MHz after 30 seconds of sustained load is thermally throttling. Improving cooling (better paste, additional fans, or a different cooler) should reduce throttle time and maintain higher sustained frequencies.

Sensor charts (Plus)

On Plus and above, the results include time-series charts for temperature and frequency throughout the stress test.

• Sequential mode: each component gets its own color-coded region on the chart, so you can see exactly when each device was tested and how temperatures and frequencies changed
• Parallel mode: all sensor data appears on a single merged chart, since all devices run simultaneously

Sensor charts show thermal ramp-up patterns, throttle points, and cooling recovery between sequential tests.

Results history

Past stress test results are saved in a sortable, filterable grid. Each row shows the date, devices tested, pass rate, duration, and peak temperature. You can click any row to view the full results for that test.

Advanced filtering (by device type, pass/fail status, date range, temperature thresholds) requires Plus or above.

Feature availability by plan

Related pages

• CPU benchmark: measure CPU performance under controlled workloads
• GPU benchmark: measure GPU rendering and compute performance
• Sensor monitoring: track temperature, power, and clock speeds over time
• Understanding your scores: how benchmark scores work and what they mean
• Troubleshooting: common issues and tips for reliable results