12 November 2026 · Jakub M.

Two hours at full load: how hot does the Mac Studio M5 Ultra get and when does it slow down?

We loaded the processor and graphics of a Mac Studio M5 Ultra to the maximum for two hours and measured temperatures, power and performance. Processor alone: no slowdown. Graphics alone: 3% slower. Both at once: macOS kept the graphics at over 90% and cut the processor to about 60%.

In our earlier tests the Mac ran real AI workloads. Sixteen parallel text generators for an hour drew 73 W and the chip did not slow down at all (one-hour load test), and even the giant Qwen3.5 397B stayed at about 104 W (Qwen3.5 397B test). This time we wanted to find the limit: two hours of synthetic load that pushes the processor and graphics as hard as they go, while we watched temperatures, power and the moment the Mac starts to slow down.

How we measured

• Hardware: Mac Studio with Apple M5 Ultra (36-core CPU, 80-core GPU) and 256 GB of unified memory, macOS 27.2.1, default settings.
• Processor load: our own small program doing vector floating-point multiply-add on all 36 cores. Every 10 seconds it reports how many billion operations per second (GFLOPS) it managed.
• Graphics load: repeated multiplication of 8,192 × 8,192 matrices in 16-bit precision with Apple’s MLX 0.32.4, reported as TFLOPS every 10 seconds.
• Temperatures: the 14 chip sensors that macOS reports as PMU tdie, read every 5 seconds. We give their average and the hottest one.
• Power, clock speeds and thermal pressure: Apple’s powermetrics tool, every 5 seconds. Thermal pressure is macOS’s own assessment (Nominal, Moderate, Heavy).
• Schedule: 5 minutes idle, 10 minutes processor only, 3 minutes pause, 10 minutes graphics only, 3 minutes pause, 80 minutes processor and graphics together, 10 minutes cooling down.

Results

Mac Studio M5 Ultra 36/80, 256 GB, macOS 27.2.1, measured on 12 November 2026. Temperatures in °C from the 14 chip sensors macOS reports as PMU tdie: average of all and the hottest one. Power from powermetrics (chip only). GPU power in brackets: peak in the first two minutes. For CPU + GPU the temperatures are from the last 10 minutes; performance and power are the range from minute 10 to the end.

Chip temperature (top) and performance relative to the best solo result (bottom) over the whole test. Shaded: periods under load.

Processor alone: no problem

With all 36 cores at full load the processor drew 124 W and the chip never went above 47.8 °C. Performance stayed at 3,049 GFLOPS from the first minute to the last and macOS reported normal thermal pressure throughout. The cooler handles the processor alone with plenty to spare.

Graphics alone: hot, but barely slower

The graphics drew up to 178 W and the hottest point of the chip reached 94.7 °C within two minutes. macOS then switched thermal pressure to Heavy and slightly lowered the graphics clock, from 1,620 MHz to about 1,555 MHz. Power settled at about 155 W and the hottest point at about 88 °C. Performance fell from 104.8 to 101.5 TFLOPS, so by about 3%.

Both at once: macOS sacrifices the processor

In the first minute the chip drew 268 W (124 W processor, 144 W graphics), with a short peak of 287 W. After two and a half minutes the hottest point reached 96.3 °C, the highest value of the whole test. macOS reacted by lowering the processor clocks from 4.2–4.3 GHz to 2.3–3.0 GHz, and processor power fell from 124 W to 36–40 W. The graphics lost much less.

For the remaining 75 minutes the Mac ran steadily in this state: the processor at 60–64% of its solo performance, the graphics at 91–94% of its best result, the chip at 77–79 °C on average and 84–87 °C at the hottest point, with total chip power of 160–178 W. Nothing crashed, no computation failed, and after the test macOS (pmset) reported no recorded thermal warning. In the last half hour performance rose slightly (processor from about 1,840 to 1,960 GFLOPS) along with power and temperature. We do not know why; it may have been a change in room temperature.

Our reading: in a normal room this cooler carries about 160–180 W of chip power over the long term. Above that macOS slows the chip down, and it protects the graphics first.

Cooling down

The Mac cools quickly. Thirty seconds after the load stopped the hottest point was at 56 °C, after two minutes at 40 °C and after ten minutes at 31 °C. At rest before the test it was 23–24 °C.

What this means if you rent the Mac

• Running language models mainly loads the graphics and memory and draws much less power than this test: 73 W and about 104 W in our earlier measurements. That is far from the limits we hit here.
• Long computations on the graphics (training, fine-tuning, image generation) will run about 3% slower once the chip is warm.
• Full processor and graphics load at the same time cuts the processor to roughly 60%. If speed matters, run heavy processor jobs (compiling, data preparation) separately from the graphics workload.

Limits of the test

• We measured one Mac. It is the same model and configuration as our second one, but we did not run the full two hours on it.
• The load is synthetic and harsher than normal work. Real applications rarely load the processor and graphics to the maximum at the same time.
• Apple does not document the temperature sensors, so take the values as indicative. They are not directly comparable with the averages from the macmon tool in our earlier test.
• We did not measure room temperature, fan speed or noise. Power is for the chip only, not the whole computer.

Try it yourself

If you want to see what the Mac Studio M5 Ultra can do with your own workload, you can rent it from one day.

Request a rental Performance

← All articles