RTX PRO 6000 power limit: running the 600 W card at 300 to 450 W with nvidia-smi on Linux

How to set the RTX PRO 6000 power limit with nvidia-smi

On Linux, the RTX PRO 6000 power limit is set with nvidia-smi, the tool that comes with the NVIDIA driver. nvidia-smi -q -d POWER shows the Min, Max and Default Power Limit of every card, and nvidia-smi -i 0 -pl 450, run as root, lowers the first GPU to 450 W. NVIDIA’s documentation says the value “needs to be between Min and Max Power Limit as reported by nvidia-smi” and that the option “Requires root.” A limit set this way does not survive a reboot or a driver reload, so a server needs a unit that sets it again at every boot, and a fleet can hand the job to DCGM.

Of the options in this article, -pl, -pm (persistence mode) and -lgc (locked clocks) carry the “Requires root” note in NVIDIA’s documentation; the query options -q, -d and -L do not. -i selects one GPU by index, board serial number, UUID or PCI bus ID. NVIDIA recommends the UUID or the bus ID for consistency, because the enumeration order can change between reboots.

A lower limit is expected to cost token generation for one or a few users less than prompt processing and training, because generation waits mainly on memory and the other two on compute.

Power limits NVIDIA documents for each card

NVIDIA publishes the maximum for every card and a configurable range only for the passive server cards in the table. Its product brief SP-12355-001_v02 for the RTX PRO 6000 Server Edition, dated 27 June 2025, gives “600 W maximum (default)” and “300 W minimum” when the 16-pin power cable is strapped for 600 W, and a 450 W maximum with the same 300 W minimum when it is strapped for 450 W. NVIDIA’s RTX PRO 6000 family page gives the Server Edition’s power consumption as 400 to 600 W. The table follows the brief, the more detailed document, and the delivered card reports the range it accepts as Min and Max Power Limit.

NVIDIA RTX PRO 6000 family page and RTX PRO 4000 SFF page, read on 16 October 2026; NVIDIA product briefs SP-12355-001_v02 (27 June 2025) and PB-12128-001_v01 (11 April 2025); DGX Spark User Guide, last updated 10 September 2026.

The cable sets a ceiling of its own. The Server Edition reads the sense pins of its 12V-2×6 connector, and on a cable strapped for 450 W it reports a Max Power Limit of 450 W, so nvidia-smi -pl 600 changes nothing; our comparison of the Workstation, Max-Q and Server Edition covers that case and the cable in the order. For the Workstation Edition and the Max-Q, NVIDIA gives only the rating on the pages we read, so the Min Power Limit that nvidia-smi reports on the delivered card is the lowest value it accepts.

What a lower power limit does to LLM inference and training

NVIDIA’s technical blog on LLM inference optimisation, published on 17 November 2023, describes the two phases of a request. Prefill, the processing of the prompt, is “a matrix-matrix operation that’s highly parallelized” and “effectively saturates GPU utilization.” In decode, the generation of each new token, the time spent moving weights and cache from memory “dominates the latency”. At its limit, nvidia-smi reports SW Power Cap: the power algorithm “is reducing the clocks below requested clocks because the GPU is consuming too much power”, and lower clocks slow compute-bound work most.

Phases from NVIDIA’s technical blog “Mastering LLM Techniques: Inference Optimization” (17 November 2023); SW Power Cap from the nvidia-smi documentation, read on 16 October 2026. Direction only, without percentages.

We found no measurement of an RTX PRO 6000 at lower power limits from NVIDIA or from a test publication that states its date and conditions, so the table shows the direction, not a figure to plan with. A RAG assistant with long retrieved passages spends more of each request in prefill than a chat does, so its users notice a cap in the time to the first token. NVIDIA’s blog adds that larger batches use more of the available compute, which moves a busy server towards the compute-bound case. Measure throughput and energy per token at the default and at one or two lower limits, as our guide to GPU energy per token describes.

Why cap: circuits, rack feeds and eight cards per server

Eight Server Edition cards draw up to 4,800 W at 600 W, 3,600 W at 450 W and 2,400 W at 300 W, before processors, memory and fans. Two such servers in one rack need 9.6 kW for the cards alone at the default limit, and 7.2 kW at 450 W. Some server models take more cards only when they are capped: Lenovo’s product guide for the SR650a V4, updated on 5 October 2026, lists two Server Edition cards at 600 W, or four in slots capped at 450 W. Feeds and airflow per rack are in our guide to GPU rack power and cooling.

In an office or a small server room the circuit decides. A 16 A circuit at 230 V carries 3,680 W, and we plan at 80 per cent of a feed. A tower with two Workstation Edition cards has 1,200 W of card power at 600 W and 900 W at 450 W, which also takes 300 W of heat out of the room, as our article on a GPU server in an office or small server room works out.

A feed sized for capped cards depends on the cap holding after every boot. NVIDIA’s briefs for the Server Edition and the H200 NVL say an nvidia-smi cap “must be reestablished after each new driver load”, while a cap set out-of-band through SMBPBI “remains in force across driver loads and system boots.” The Server Edition brief adds that SMBPBI “requires that the NVIDIA driver is loaded for full functionality”. Where the server maker’s BMC offers no such setting, size the breakers for the uncapped rating.

We build AI servers with up to eight RTX PRO 6000 Server Edition or H200 NVL cards and check the rack, power and airflow before we quote. Send us the feed per rack position and the card count you plan through the form below.

Making the power limit persistent at boot

NVIDIA’s driver persistence documentation counts the software power capping limit among the settings that last only while the driver keeps the GPU initialised, and states that “The GPU is typically deinitialized after the last client exits.” Without persistence, a limit can fall back between two jobs, not only at a reboot. NVIDIA steers users to the persistence daemon, nvidia-persistenced, and writes that the legacy persistence mode will eventually be “deprecated and removed”; its sample init scripts, one of them for systemd, ship with the driver documentation and are not run by the installer.

The procedure below is our example for a Linux server; the nvidia-smi commands in it are as NVIDIA documents them, and every step runs as root.

1. Read each card’s Min, Max and Default Power Limit with nvidia-smi -q -d POWER, and note each GPU’s UUID with nvidia-smi -L, since the index order can change between reboots.
2. Make sure nvidia-persistenced starts at boot, from your distribution’s driver package or NVIDIA’s sample systemd script, and check that nvidia-smi -q reports Persistence Mode as Enabled.
3. Set the limit by hand on one card, for example nvidia-smi -i 0 -pl 450, and confirm it under load with nvidia-smi -q -d POWER.
4. Create /etc/systemd/system/gpu-power-limit.service with an After= line under [Unit] naming the unit that starts nvidia-persistenced, and under [Service] Type=oneshot, RemainAfterExit=yes and one ExecStart= line per card, such as /usr/bin/nvidia-smi -i <UUID> -pl 450.
5. Add WantedBy=multi-user.target under [Install], run systemctl daemon-reload and then systemctl enable --now gpu-power-limit.
6. Reboot and read the limits again; after a driver update or a driver reload without a reboot, run systemctl restart gpu-power-limit and check once more.

For an RTX PRO 6000 undervolt, the nvidia-smi documentation lists no option to set the GPU voltage; its documented controls are the power limit and locked clocks with -lgc, which is “Supported on Volta+. Requires root.” For a server, keep to the power limit, which the driver enforces and reports.

DCGM power limits for a fleet of GPU servers

NVIDIA’s Data Center GPU Manager suits a fleet. Its feature overview, updated on 17 June 2026, lists Power Limit among the settings it manages, to “Set the maximum allowed power consumption”, and says: “To define a target configuration for a group, use the dcgmi config subcommand.” A group is created with dcgmi group -c GPU_Group. DCGM then “will enforce that configuration, for example across driver restarts, GPU resets or at job start”, which covers the reloads that a boot unit misses.

Its policy feature can notify on “Excursions above specified board power threshold”. DCGM’s feature table, which still uses NVIDIA’s old brand names, marks configuration management for all four product lines but policy notifications only for Tesla, the data-centre line. For monitoring, export the enforced limit, DCGM_FI_DEV_BOARD_POWER_LIMIT_ENFORCED_WATTS, beside board power, and alert when it is lower than planned and no one set a cap. Other fields and clock event reasons are in our guide to GPU server monitoring with DCGM.

Compact and small-form-factor builds

Where the chassis sets the power budget, a card designed for it fits better than a capped 600 W card. The RTX PRO 6000 Max-Q has the same GPU and 96 GB at 300 W, with a blower, and NVIDIA states up to four per workstation. A Workstation Edition capped at 300 W keeps its flow-through cooler and its size. A tower with four Max-Q cards draws about 1.8 to 2 kW at the wall by our estimate, within one 16 A circuit.

For short-depth servers and branch sites, NVIDIA lists the RTX PRO 4000 SFF at 70 W in a low-profile, dual-slot card with 24 GB of GDDR7, and the RTX PRO 2000 at 70 W with 16 GB. At that power the budget is set by design, and the chassis maker’s list of qualified cards decides which one fits.

We supply the RTX PRO 6000 Max-Q, the RTX PRO 4000 SFF and the RTX PRO 2000 as cards and in workstations and short-depth servers built to order. Describe the chassis, its power supply and the workload in the form below.

DGX Spark power limit

NVIDIA’s DGX Spark User Guide, last updated on 10 September 2026, lists a “240W external power supply (included)” and a GB10 TDP of 140 W, with 100 W available for the other components, such as the ConnectX-7 network, the SSD and USB-C. Its known-issues page warns that a different adapter “may reduce performance, prevent boot, or cause unexpected shutdowns.” We found no section of the user guide that describes setting a power limit on DGX Spark; the supplied adapter bounds the draw, so four Sparks need 960 W of adapter rating.

What we supply

We supply the RTX PRO 6000 Workstation Edition, Max-Q and Server Edition, the H200 NVL, the RTX PRO 4000 SFF and the RTX PRO 2000 as cards or in AI servers built to order, and DGX Spark, with manufacturer warranty on one EU contract and invoice. Each server is assembled and burn-in tested, and we check the rack, power and airflow before we quote. Operating system, drivers, CUDA and a container runtime are installed on request, and the driver includes nvidia-smi for reading power and setting limits. The maximum power of each card we supply is on our professional NVIDIA GPU page.

FAQ