Two NVIDIA RTX PRO 6000 Blackwell Workstation Edition cards: heat and power

What two 600 W cards do in one tower

Two NVIDIA RTX PRO 6000 Blackwell Workstation Edition cards run well only in a tower designed for 1,200 W of GPU heat that stays inside the case: makers list at most two per machine, and by our estimate the workstation draws about 1.8 kW at the wall. A power limit of 450 W per card, set with nvidia-smi, lowers that to about 1.4 kW at a published cost of a few per cent of throughput, and two Max-Q cards give the same 192 GB at half the card power.

This article is for IT teams and lab managers who plan such a machine or have one under a desk that runs hot. Our page to buy an RTX PRO 6000 Workstation Edition PC covers the configuration and ordering; here the subject is the heat, the power supply, the circuit and the power limit.

Where the heat goes

NVIDIA's family page lists the Workstation Edition with a "double flow-through" cooler at 600 W and describes the card as "engineered for maximum performance in single-GPU workstations". Two axial fans draw air through the heat sink, and the air leaves through the top of the card into the case. Puget Systems, which builds workstations with these cards, describes the consequence for two cards: "the lower card will push hot air into the intake of the upper card". Our own product page says the same in fewer words: the first card's exhaust feeds the second.

The Max-Q is built the other way round. Its blower pulls air from inside the case and pushes it out through the slot bracket at the back, so each card removes most of its own heat from the case; this is what lets four sit side by side. With two Workstation Edition cards, 1,200 W of heat stays in the case until the case fans move it out, and the processor and memory take their cooling air from the same case.

Puget's article of October 2025 on its dual-GPU systems draws the practical conclusion. Cards whose coolers vent back into the case need high-flow chassis fans of the kind used in servers, which are loud, so Puget builds its systems with two RTX PRO 6000 Blackwell Workstation cards in a 5U rackmount chassis with "thicker-than-normal, server-grade chassis fans", designed for server rooms and used remotely. Its parts page for the card calls it "NOT suitable for most multi-GPU configurations".

Published thermal tests and builds

We found no published per-card temperature measurement of exactly two Workstation Edition cards in a closed tower. The closest published evidence comes from integrators and test sites, each with a different setup.

VRLA Tech thermal test, Puget Systems blog and StorageReview enterprise GPU roundup, read on 24 October 2026; GamersNexus figures as reported in our comparison of the three editions. VRLA does not state ambient temperature, test duration or tool.

VRLA's test is the clearest data point on 600 W cards in a tower: with side panels on and full stress load, the hottest of four cards reached 89 to 90 °C, which VRLA describes as essentially no thermal headroom, and the configuration "did not pass our validation". The same page states as its rule that two or more RTX PRO 6000 cards in one chassis should be Max-Q, while its FAQ offers Workstation Edition systems for one or two cards. Two cards make half the heat of four, so the test does not translate directly to two; it shows where the limit lies for this cooler design. The makers' specifications draw the line at two: Dell's configurator offers two 600 W cards for the Pro Precision 9 T6, Lenovo one for the ThinkStation P8, and HP's and Lenovo's four-card towers list only the Max-Q.

Power budget at 600, 450 and 300 W per card

Our estimate: card power plus a 350 W processor (AMD's default TDP for the Threadripper PRO 9975WX), plus 15 per cent for memory, drives, fans and power-supply losses; 1 W = 3.412 BTU/h.

The 300 W row matches two Max-Q cards at full power, which is the comparison that matters for an office. The figures are estimates from rated values. A processor running below its TDP, or a workload that keeps the cards below their limit, draws less; a processor such as Intel's Xeon 698X, rated at 350 W base and 420 W turbo, can draw more.

Power supply, cords and the 230 V circuit

• Power supply. Systems with two 600 W cards use large supplies: StorageReview's launch report gives the Pro Precision 9 T6 a 2,400 W unit, and Puget's dual systems use 2,800 W supplies. Each card has one 12V-2×6 connector, according to Puget's parts page, so the supply needs one native cable per card.
• Input voltage. These supplies reach their rating only on 200 to 240 V. According to StorageReview, Dell's 2,400 W unit delivers 1,500 W at 115 to 199 V and 1,200 W at 100 to 114 V, and Puget writes that its 2,800 W supplies "require 200-240V power circuits". On a European 230 V socket the full rating is available; the lower figures in the makers' tables apply to 100 to 127 V networks.
• Cord. An IEC C13 appliance cord is rated at 10 A, 2.3 kW at 230 V, and a C19 cord at 16 A. StorageReview lists a C19 cord for the 2,200 W unit of Dell's Precision 7960; check the inlet of the supply you order and use the cord the maker ships with it.
• Circuit. We plan a 230 V 16 A circuit at 80 per cent, 2.94 kW. A two-card tower at about 1.8 kW fits, with about 1.1 kW left; two such towers on one circuit, at about 3.6 kW, do not.
• UPS. An uninterruptible power supply for this machine needs headroom above 1.8 kW.

Our article on GPU server power on 16 A, 32 A and three-phase covers larger feeds, and Max-Q power on 230 V the same arithmetic for Max-Q builds.

We build two-card workstations with the power supply, cables and slots chosen for 1,200 W of cards and check the room before we quote; tell us the circuit rating and where the machine will stand, and our team replies with a configuration.

Power limiting with nvidia-smi

A lower power limit reduces power and heat on both cards without changing the hardware. On Linux, nvidia-smi -q -d POWER shows each card's Min, Max and Default Power Limit, and nvidia-smi -i 0 -pl 450, run as root, sets the first card to 450 W. NVIDIA's documentation says the value "needs to be between Min and Max Power Limit as reported by nvidia-smi". NVIDIA gives only the 600 W rating for the Workstation Edition, so the Min Power Limit that the delivered card reports is the lowest value it accepts.

The limit does not survive a reboot or a driver reload. NVIDIA's driver persistence documentation counts the software power cap among the settings that last only while the driver keeps the GPU initialised, so the workstation needs the persistence daemon and a small systemd unit that sets the limit at every boot. Our guide to the RTX PRO 6000 power limit gives the unit step by step and addresses each card by its UUID, because the index order can change between reboots.

The cost of a limit depends on the work. StorageReview capped a single Workstation Edition at 450 W in a Dell tower and measured 2 to 5 per cent less vLLM throughput at short context and 4 to 9 per cent at long context, with no change in its FP8 matrix-multiply result; its output tokens per second per rated watt rose from 3.8 to 4.8 on Ministral 3 14B in FP8 at 32 streams (October 2026, vLLM 0.27.1, driver 595.71.05). Token generation for few users waits mainly on memory and loses least; prompt processing, large batches and training wait on compute and lose more. While the limit is active, nvidia-smi -q -d PERFORMANCE reports SW Power Cap among the clocks event reasons.

A limit lowers the heat, and the airflow pattern stays the same. At 450 W each card still releases its heat into the case and the lower card still feeds the upper one, now with 900 W in total instead of 1,200 W.

Monitoring temperatures and throttling

nvidia-smi shows whether the cooling keeps up. nvidia-smi -q -d TEMPERATURE lists GPU Current Temp, GPU Slowdown Temp, GPU Shutdown Temp and, on supported cards, Memory Current Temp. nvidia-smi -q -d PERFORMANCE lists the clocks event reasons: SW Thermal Slowdown means the software algorithm is reducing clocks because of temperature, and HW Thermal Slowdown, in NVIDIA's words, is "reducing the core clocks by a factor of 2 or more due to temperature being too high". On a two-card tower, compare the two cards after an hour of sustained load with the side panels closed, the condition under which VRLA validates its systems.

We assemble every workstation and test it before delivery (power-on and basic load test). The checks in our GPU server acceptance testing guide apply to a workstation as well, and our article on GPU monitoring with DCGM covers alerting when a card throttles.

Room, noise and placement

All of the 1.8 kW becomes heat in the room, about 6,100 BTU per hour. A closed office without cooling warms up over a working day, and the workstation's own intake temperature rises with it, which leaves the upper card even less margin. No maker publishes an acoustic figure for a two-card Workstation Edition tower under GPU load; for one card, GamersNexus measured about 32.5 dBA. Puget advises against using its dual-GPU rackmount systems at a desk because of their power needs and loud fans, and installs them in server rooms. Our guide to a GPU server in an office covers rooms shared with people.

When two Max-Q or a server is the better choice

• Two Max-Q. The same 192 GB and 1,792 GB/s per card at 300 W each, with blowers that exhaust out of the case, about 1.1 kW at the wall by our estimate, and a tower that can take four later. Puget measured the Max-Q 5 to 14 per cent behind the Workstation Edition in content-creation applications. Our guide to what a workstation with two Max-Q runs has the models and users.
• A rack server. For a service that runs continuously, a server with two RTX PRO 6000 Server Edition cards keeps the heat and noise in a server room, with redundant power supplies and remote management; our page on the 2× RTX PRO 6000 Server Edition server covers it, and our article on Server Edition power at 600, 450 and 300 W the power modes.
• The Server Edition in the tower. A passive card needs ducted server airflow, which a tower does not provide; our article on the Server Edition in a workstation explains why we recommend the Workstation Edition or the Max-Q instead.

We supply all three editions and build workstations and rack servers with them to order; describe the workloads and the room, and we recommend the edition and the power limit before we quote.

FAQ