NVMe SSD Thermal Throttling in Storage Benchmarks

This article explores how thermal throttling in NVMe Solid State Drives (SSDs) manifests during intensive storage benchmarking. When pushed to their limits by synthetic workloads, high-speed NVMe drives generate significant heat, forcing onboard controllers to intentionally reduce performance to prevent permanent silicon damage. Below, we examine the exact indicators of thermal throttling during benchmark runs, including sudden throughput drops, latency spikes, sawtooth performance graphs, and how to distinguish heat issues from normal cache exhaustion.

Sudden Drops in Sequential Throughput

During heavy sequential read and write tests—such as those conducted in CrystalDiskMark, AS SSD, or ATTO—thermal throttling primarily appears as a sudden, drastic reduction in data transfer speeds. A PCIe 4.0 or 5.0 drive operating at 7,000 MB/s to 10,000 MB/s may suddenly plummet to 1,000 MB/s or lower, sometimes dropping below the speeds of older SATA drives.

This performance cliff happens the exact moment the SSD controller reaches its programmed thermal threshold, which typically lies between 70°C and 85°C. To protect the hardware, the firmware drastically cuts the controller's clock speeds and voltage, resulting in an immediate bottleneck during the benchmark run.

The "Sawtooth" Performance Graph

In sustained benchmarks that graph performance over time, such as the AIDA64 Disk Benchmark, HD Tune Pro, or IOMeter, thermal throttling creates a distinct "sawtooth" pattern. This erratic oscillation occurs in four repeating stages:

  1. The drive hits its thermal limit under load and throttles speeds aggressively.
  2. The reduced workload allows the controller to cool down by a few degrees.
  3. Once temperatures drop below a safe threshold, the drive boosts its speed back up.
  4. The renewed high speed quickly generates heat, triggering the throttle again.

Instead of a flat, stable performance line, the benchmark results will show a jagged, up-and-down graph that indicates the cooling solution is insufficient for continuous workloads.

Severe Latency Spikes and IOPS Degradation

Thermal throttling does not just affect raw bandwidth; it heavily impacts drive responsiveness. In benchmarks measuring Random 4K performance and Input/Output Operations Per Second (IOPS), throttling manifests as severe latency spikes and plummeting IOPS scores.

Access times, which are normally measured in fractions of a millisecond, will suddenly spike into tens or hundreds of milliseconds. The drive artificially delays processing read and write queues to reduce the electrical load within the controller. If a benchmark measures "Quality of Service" (QoS) or 99th-percentile latency, a thermally throttled drive will produce abysmal scores due to these artificial delays.

Telemetry and SMART Data Triggers

Alongside visual drops in benchmark scores, system monitoring tools running concurrently (such as HWiNFO64 or CrystalDiskInfo) will explicitly log the throttling event. Modern NVMe drives feature multiple thermal sensors—usually one on the controller and another on the NAND flash memory.

During a throttled benchmark run, the drive's SMART data will show the controller temperature hitting its critical limit. In advanced monitoring software, you will see the "Thermal Throttling" status flag flip from "No" to "Yes," and the "Warning Composite Temperature Time" counter will actively tick upward for every second the benchmark forces the drive to remain in its overheated state.

Distinguishing Thermal Throttling from SLC Cache Exhaustion

When analyzing storage benchmark data, it is vital not to confuse thermal throttling with SLC cache exhaustion, as both result in massive speed drops during testing. You can tell them apart by looking at the workload type and the shape of the performance drop:

  • SLC Cache Exhaustion: This occurs only during sustained write benchmarks. Once the drive's fast SLC buffer fills up, speeds drop to a lower, yet relatively stable, flat line as the drive writes directly to slower TLC or QLC NAND blocks.
  • Thermal Throttling: This occurs during both read and write benchmarks. Because read operations do not utilize the SLC write cache, a massive speed drop during a pure sequential read test is almost guaranteed to be thermal throttling. Furthermore, thermal throttling results in fluctuating, unstable speeds that directly correlate with peak temperature readings, whereas cache exhaustion happens at a predictable data-volume limit regardless of the drive's temperature.