NVIDIA RTX PRO 6000 Blackwell Server Edition power: 600 W, 450 W and the 300 W minimum

600 W, 450 W and 300 W in NVIDIA's brief

The NVIDIA RTX PRO 6000 Blackwell Server Edition draws up to 600 W by default when its 16-pin power cable is strapped for 600 W, and up to 450 W when the cable is strapped for 450 W. In both modes the lowest power limit is 300 W. These figures are in Table 2-1 of NVIDIA's product brief SP-12355-001_v02 of June 2025, and the server maker decides which mode a server uses through the cable, the slot and the firmware.

NVIDIA RTX PRO 6000 Blackwell Server Edition product brief SP-12355-001_v02 (June 2025), Table 2-1, read on 24 October 2026.

Other ranges circulate. NVIDIA's RTX PRO 6000 family page lists the Server Edition at 400 to 600 W, Supermicro at "400-600W", Dell's AI Factory datasheet as "configurable between 450W & 600W", and Lenovo as "600 W (can also be power capped to 450 W to support increased density)". These describe what a page or a server supports. The brief, the more detailed document, sets the card's own floor at 300 W, and the delivered card reports the range it accepts as Min and Max Power Limit in nvidia-smi -q -d POWER.

Power is not the only limit the brief sets. The card is passive and needs air from the server: at least 43 CFM through its heat sink at a 25 °C inlet, rising to 125 CFM at 45 °C, within an ambient range of 10 to 45 °C. A lower power mode reduces the heat to remove; it does not remove the airflow requirement.

How the cable and its sense pins set the ceiling

The 16-pin connector carries two sense pins that tell the card how much power the cable may deliver. NVIDIA's brief states: "If the power level identified by Sense0 and Sense1 is less than the default power cap of the NVIDIA card, the card will not boot." The cable is therefore part of the configuration, as much as the card.

A thread on NVIDIA's developer forum shows the effect. A Supermicro SYS-521GE-TNRT reported Min Power Limit 300 W, Max 450 W and Default 450 W on a Server Edition card. A user replied that the maximum is "adjusted to what the 12VHPWR cable is identifying as", and that the cable shipped with the board, CBL-PWEX-0962Y-30, is "configured for 450W". According to a post of 11 October 2025, Supermicro advised cable CBL-PWEX-1364Y-30 for 600 W. NVIDIA staff did not reply in the thread. nvidia-smi -pl 600 cannot lift a ceiling the cable sets.

For a buyer this means two checks. Ask for the cable part number in the quote, with the power mode it selects, and read the Max Power Limit of every card during acceptance. Our Server, Workstation and Max-Q comparison covers the editions and their cables.

nvidia-smi, SMBPBI and the server's BMC

Below the ceiling, a cap can be set in two ways, and the brief describes both. In band, through the driver with nvidia-smi, "power cap adjustment must be reestablished after each new driver load". Out of band, through SMBPBI, the management interface the server's BMC talks to, "power cap adjustment remains in force across driver loads and system boots". A feed sized for capped cards therefore depends on either a boot-time service that sets the cap again or a BMC that sets it out of band. Our guide to the RTX PRO 6000 power limit with nvidia-smi shows the commands, a systemd unit and DCGM for a fleet.

Server makers handle this differently:

• HPE sells the ProLiant DL385 Gen11 with "450W or 550W power capping enablement kit" for iLO GPU power capping, which "can also be set in iLO 6 GUI for servers already deployed". Its Maximum Performance fan kit is supported only with 600 W GPUs such as the RTX PRO 6000 Server Edition.
• Lenovo offers the card as part 4X67B09287 in two variants: feature CBK8 at 600 W and CHWT, "Slot Capped at 450W", added in July 2026.
• Dell's technical guide for the PowerEdge XE7745 states that the server "does not support user power cap"; its power supply matrix lists the eight cards at 600 W.

What server makers set per configuration

Lenovo Press LP2263 (28 July 2026), HPE QuickSpecs a50004300enw, Dell PowerEdge XE7745 technical guide Rev. A09 (September 2026), Cisco UCS C845A M8 data sheet (24 August 2026), NVIDIA developer forum thread of September 2025 to February 2026, read on 24 October 2026.

The capped modes exist for density. Lenovo's SR650a V4 takes two cards at 600 W or four at 450 W in 2U. HPE's four-card DL385 Gen11 in MLPerf® Inference v6.0 (closed division, Llama 2 70B llama2-70b-99, offline, entry 6.0-0054, retrieved from mlcommons.org on 24 September 2026, result verified by MLCommons Association) ran its cards at 450 W, as our RTX PRO 6000 benchmarks record. We supply the card in servers at 600 W or in capped configurations; send us your rack feed and the card count you plan, and our team tells you which power mode and chassis fit.

kW per server at 600, 450 and 300 W

The card is the largest load in the server but not the only one. For two cards we assume two AMD EPYC 9135 at 200 W each, for four cards two EPYC 9355 at 280 W each, for eight cards two EPYC 9555 at 360 W each, and add 15 per cent for memory, drives, network cards, fans and conversion losses. A measurement on the delivered system replaces these estimates.

Our estimate: board power from NVIDIA's brief at 600, 450 and 300 W; processor TDP from AMD's EPYC 9135, 9355 and 9555 pages; plus 15 per cent for the rest of the server.

Dell's datasheet for its eight-card PowerEdge XE7740 and XE7745 gives "~8 kW" per node at full load, more than our 6.3 kW, so plan the room for the higher figure until a measurement on the delivered system shows otherwise. We plan a 230 V 16 A feed at 2.94 kW, 32 A at 5.89 kW and 400 V three-phase 16 A at 8.87 kW, 80 per cent of the rating, as in our guide to GPU rack power and cooling. With redundant supplies on A and B feeds, each side carries the whole server.

Read against those feeds, the cap changes the installation in two places. A four-card server at 600 W, about 3.4 kW, is above one 16 A feed; at 450 W, about 2.7 kW, it stays below by our estimate. An eight-card server at 600 W needs three-phase power; at 450 W, about 5.0 kW, it fits a 32 A single-phase feed by our estimate, though not with Dell's 8 kW figure. Our article on GPU server power: 16 A, 32 A and three-phase covers plugs and PDUs.

Power supplies follow the same arithmetic. Dell's 3,200 W supplies for the XE7745 deliver 3,200 W at 220.1 to 240 V and 2,900 W at 200 to 220 V, so European 230 V mains gets the full rating. With eight RTX PRO 6000 Server Edition at 600 W, the XE7745's matrix supports 4+4 A/B grid redundancy only with 3,200 W supplies, not with 2,400 W. HPE's DL380a Gen12 requires 2,400 or 3,200 W supplies with RTX PRO 6000 Server Edition cards, five for two or four double-wide GPUs, on C19 or C19-C20 cords only.

What the 450 W cap costs in throughput

Two published measurements give figures, with different methods.

StorageReview's enterprise GPU roundup of 14 October 2026 ran the RTX PRO 6000 Workstation Edition at 600 W and at 450 W with NVIDIA's NGC vLLM 26.08 container (vLLM 0.27.1, CUDA 13.4.1) and driver 595.71.05. At 450 W, "the card gave up 2 to 5 percent at short context and 4 to 9 percent at long context", and delivered 4.8 output tokens per second per rated watt on a 14B model against 3.8 at 600 W; its FP8 matrix-multiply result "did not move". The card sat in a Dell Pro Max Tower T2; NVIDIA rates both editions at 600 W, and StorageReview writes that they "profile identically" in its testing.

Delta Computer published a power sweep of the Server Edition on 11 March 2026 with two compute tests. Its page does not describe the method, the host or the ambient temperature, and the values below are paired by their position in its lists.

Delta Computer blog "RTX PRO 6000 Server Edition als Allrounder", 11 March 2026, read on 24 October 2026; temperatures from the tensor test. Method, tool, host and ambient not stated.

By our arithmetic, the tensor test keeps 86 per cent of its result at 450 W and 54 per cent at 300 W. The two sources fit together once the workload is named. Pure compute falls with the power limit, because the card lowers its clocks. Token generation for a moderate number of users waits mainly on memory bandwidth, which the cap does not change, so it loses only a few per cent. Prompt processing of long contexts and large batches lean on compute, which is where StorageReview found the larger loss. HPE's MLPerf® runs point the same way: by our division of each result by its card count, the four cards at 450 W (entry 6.0-0054) delivered 3 per cent less per card offline than two cards at 600 W (entry 6.0-0052), on different driver builds.

Choosing the mode follows from this. Run 600 W where feed and chassis allow it, and for training, fine-tuning and long-context prefill. Use 450 W where it decides the feed or the chassis, at a cost of a few per cent for interactive inference. 300 W is the floor; at about half the tensor throughput it suits a room with a hard heat limit more than a production server. Our guide to GPU energy per token explains how to measure the trade-off on your own model. We size feed, supplies and power mode before we quote; tell us the feed per rack position and your inlet temperature, and we return the configuration with the power per feed.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build AI servers to order with two to eight RTX PRO 6000 Server Edition cards at 600 W or in capped configurations, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component. Quantities and lead time are confirmed in writing with the quote. Our chooser for 2, 4 or 8 Server Edition cards and the pages on four cards and eight cards go into each size.

FAQ