What an 8x RTX PRO 6000 server really draws: power and temperatures we measured

We run one server with eight NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs in a data centre near Sofia. Spec sheets give the maximum power of each card. They do not tell you what the whole machine pulls from the wall, how hot the cards get in a 4U chassis, or which BMC setting quietly cuts your performance in half. These are the numbers we measured on our own server between July and October 2026.

The numbers at a glance

The rest of this article explains each line and what it means if you plan to put a server like this in a rack.

Power: what the wall sees

NVIDIA lists the Server Edition at "up to 600 W (configurable)". Eight cards at 600 W are 4,800 W for the GPUs alone. Under full GPU load our server draws about 6.5 kW at the wall. The difference, about 1.7 kW, is everything else: two server CPUs, the system memory, the fans and the losses in the power supplies.

Two other numbers matter as much as the peak:

• Idle is not zero. With no workload the machine still draws about 1.1 kW. Eight GPUs at rest, the CPUs, the memory and the fans add up. Over a 30-day month that is roughly 790 kWh before anyone runs a single job.
• The synthetic peak is higher than real work. With a stress test on the GPUs, CPUs and memory at the same time we saw 6,960 W. Real AI workloads rarely push the CPUs that hard while the GPUs are saturated, so 6.5 kW is the more useful planning number. The 7 kW figure is the one your breaker and your power contract must survive.

Our server has four power supplies in a 3+1 redundant setup: three carry the load and one is spare. If you size the supplies yourself, size them for the synthetic peak with one supply failed, not for the average.

Do not trust the BMC power reading

The easiest way to read power on a server is the BMC, through IPMI and its DCMI power reading. On our machine that number is wrong by about half. The BMC sees two of the four power supplies, and the load is shared evenly across all four, so DCMI shows roughly 50% of the real draw.

We found this only by comparing the BMC with the data centre's own electricity meter over several weeks. The lesson is simple: measure at the wall or at the rack PDU, not from the BMC. If your colocation provider bills by metered kWh, ask for the meter data and compare it with what your monitoring says during the first month.

Temperatures: 87°C at full power

The Server Edition has no fans of its own. NVIDIA describes it as a passive card that relies on the server's airflow. In a 4U chassis with eight cards side by side, that airflow is the whole cooling system.

We tested with gpu-burn, an open-source tool that keeps every GPU at full load. With the chassis fans at full speed and each card at its 600 W limit, the hottest card peaked at 87°C and kept the full 600 W without throttling. With the limit lowered to 525 W per card, the peak dropped to 82°C. During a real customer job on 17 October 2026 the cards ran at 85°C and 599 W.

So the chassis can hold 8 × 600 W, but with little margin. Even with the fans at maximum, the cards sit in the high 80s. A warmer room or a blocked filter would push them into thermal slowdown, which nvidia-smi reports as the GPU running above its "Max Operating Temp".

The fan setting that cost us 38% of the power

The most expensive mistake is a default. Out of the box, our BMC used the "Optimal" fan mode. Under a pure GPU load the cards throttled to 372 W at 88°C, about 62% of their rated power.

The reason: the BMC fan curve follows the CPU and system sensors, not the GPUs. A GPU job leaves the CPUs cool, so the BMC sees no reason to spin the fans up while the GPUs overheat. Switching the fan mode to "Full Speed" in the BMC fixed it at once, and the setting survives reboots.

If you build or buy a multi-GPU server with passive cards, check this before anything else. Run a full GPU load for 30 minutes and watch the power draw of each card in nvidia-smi. If it falls well below the limit while the temperature sits near 88°C, the fans are not following the GPUs.

Power caps: trading watts for margin

nvidia-smi -pl sets a power limit per card, within the range the card reports; it needs root. A lower limit gives you thermal margin at the cost of some speed. In our test, 525 W per card gave 5°C more headroom than 600 W.

Two practical notes:

• A cap set with nvidia-smi -pl should be treated as runtime state. If you rely on it, apply it at every boot, for example with a systemd service, and check it after driver updates.
• We run the full 600 W, because the cards hold it without throttling with the fans at full speed. If the data centre gets warmer, a cap of 525 to 550 W is the first lever, not a new chassis.

What to plan for in colocation

If you are putting a server like this into a colocation rack, here is what our measurements translate into:

• Power per rack. Plan for about 7 kW per server at peak and 6.5 kW sustained. Uptime Institute's 2025 survey describes most new racks in the 10 to 30 kW range and few above 30 kW. One server fits a 10 kW rack. Two need about 14 kW at peak. Ask what your rack is allowed to draw before you buy the second machine.
• Cooling. Nearly all of that power ends up as heat: 6,960 W is about 23,750 BTU per hour. The room must remove it continuously, not just at the average.
• Fans at full speed. Expect them to run at maximum all the time. It is loud and it uses some power at idle, but it is what keeps eight passive cards at 600 W.
• Metering. Use the provider's meter or the rack PDU as the truth, and calibrate your own monitoring against it.
• Billing. Idle draw is real money every month even when the GPUs earn nothing. Our AI server cost guide shows how power fits next to the purchase price.

Colocate your own server or rent a whole one

Our own machine is the one measured above: eight RTX PRO 6000 Server Edition cards, 768 GB of GPU memory, root access and the whole server for one customer. Pricing starts at $1.70 per GPU-hour, about $9,930 a month for the whole machine, power and cooling included. Details are on the pricing page.

Questions

How much power does one RTX PRO 6000 Server Edition use? NVIDIA rates it at up to 600 W, and the limit is configurable. Under a sustained full load with enough airflow our cards stayed at 599 to 600 W.

Why does my multi-GPU server throttle when the GPUs are not at their limit? Check the BMC fan mode first. On our server the factory "Optimal" mode followed CPU temperatures and let the GPUs throttle to 372 W at 88°C. "Full Speed" fixed it.

Can I read server power from IPMI? You can, but verify it. On our machine the DCMI reading covers two of four power supplies and shows about half the real draw. A wall meter or rack PDU is the reliable source.

How much heat does an 8-GPU server produce? Almost all of the electrical power becomes heat. At our measured 6,960 W peak that is about 23,750 BTU per hour.

Sources

All pages were checked on 20 October 2026.

• NEOXIS measurements on our 8 × RTX PRO 6000 Blackwell Server Edition server, July to October 2026: data centre power meter (idle and full GPU load), wall measurement under synthetic stress, gpu-burn tests on 9 July 2026, customer workload on 17 October 2026.
• NVIDIA: RTX PRO 6000 Blackwell Server Edition, "up to 600 W (configurable)", 96 GB GDDR7, 1,597 GB/s.
• NVIDIA: RTX PRO 6000 Blackwell Server Edition datasheet, hosted by Akamai, passive thermal design.
• NVIDIA: nvidia-smi documentation, the -pl power limit option and the clock event reasons for thermal slowdown.
• gpu-burn, open-source multi-GPU stress test.
• Uptime Institute: 15th annual Global Data Center Survey, press release of 30 July 2025, on rack densities.
• NEOXIS: pricing and server specifications.