Operations

Running the Sparks

Four NVIDIA DGX Spark in our own infrastructure in Austria. This page shows the operating data as it comes: daily values from yesterday and 90-day histories. Including the unflattering ones.

Per Spark

Spark 1

vLLM · Nemotron 3.5 Bolt and Qwen 3.8 27B

Days without an automatic restart: 2
Automatic restarts, 90 days: 5
GPU yesterday min / mean / max: 48 / 51 / 85 °C

GPU temperature per day over 90 days: band from minimum to maximum, line for the mean, dashed 85 °C. On 1 days the maximum reached 85 °C or more.

Spark 2

vLLM · Qwen 3.8 Flash-Next, internal only

Days without an automatic restart: ≥ 2
Automatic restarts, 90 days: 0
GPU yesterday min / mean / max: 47 / 53 / 79 °C

GPU temperature per day over 90 days: band from minimum to maximum, line for the mean, dashed 85 °C. On 0 days the maximum reached 85 °C or more.

Spark 3

vLLM · Qwen 3.6 35B and Nemotron 3.5 Lightning

Days without an automatic restart: 26
Automatic restarts, 90 days: 14
GPU yesterday min / mean / max: 44 / 46 / 74 °C

GPU temperature per day over 90 days: band from minimum to maximum, line for the mean, dashed 85 °C. On 1 days the maximum reached 85 °C or more.

Spark 4

Ollama · models on demand and Clef

Days without an automatic restart: 68
Automatic restarts, 90 days: 1
GPU yesterday min / mean / max: 43 / 45 / 60 °C

GPU temperature per day over 90 days: band from minimum to maximum, line for the mean, dashed 85 °C. On 0 days the maximum reached 85 °C or more.

An automatic restart means systemd brought a service back up by itself after a crash or a failed start. Manual restarts, for example for updates, do not count. Several restarts within three hours count as one event, for example a start loop while setting up a new model. "≥" means: no restart since the series began.

Speed per model

Weekly median of the nightly measurement, daily values for models that have only been running for a few weeks: single-stream decode rate, the same method as in our engineering articles. Since 21 Oct 2026; before that the measurement used a different prompt and is not comparable with the newer values (vertical line).

Qwen 3.8 27B

Spark 1 · latest 38.0 t/s

Nemotron 3.5 Bolt

Spark 1 · latest 74.3 t/s

Qwen 3.8 Flash-Next

Spark 2 · latest 36.5 t/s

Qwen 3.6 35B

Spark 3 · latest 79.3 t/s

Nemotron 3.5 Lightning

Spark 3 · latest 110.1 t/s

Dips stay in. The drop on Spark 1 and 2 in October came from the kernel's memory compaction; how we found it is described in the article on two models on one DGX Spark.

What is not shown here

No host names, addresses or software versions, no minute-level data and no request volume. No energy figures either: the Sparks only report GPU power, not power at the wall, and a number derived from it would be wrong. Availability over the last 30 days is on the status page.