Workstation Thermals and Power in Indian Offices: Planning for 600 W Cards

Overview

Workstation GPU power has climbed steadily, and current top-tier cards draw in the region of 600 W each. Combined with a workstation CPU, drives and losses, a single-card machine can exceed a kilowatt at the wall and a dual-card machine considerably more. That would be manageable if the load were bursty, as it is for rendering or gaming, but AI workloads are sustained — a fine-tuning run or a serving process holds the card near its power limit for hours. In an Indian office, that combination is a real planning problem and it is usually discovered after purchase.

Why sustained load is different

Traditional workstation thermal design assumes duty cycles. A render spikes and finishes; a simulation runs then stops; a game varies with the scene. Cooling systems and the thermal mass of the chassis absorb peaks, and average temperatures stay well below limits. AI inference and training break that assumption entirely: a serving process at steady load holds the GPU near its power limit indefinitely, and a fine-tuning run does the same for hours or days.

Under sustained load the system reaches thermal equilibrium, and equilibrium is determined by intake air temperature and airflow, not by thermal mass. A machine that benchmarks beautifully for two minutes may settle 15 to 20 percent below that figure after twenty minutes, and that settled figure is the performance you actually own.

The room is part of the system

Essentially all electrical power entering a workstation leaves as heat. A machine drawing 1,000 W adds 1,000 W of heat to its room, continuously, which is comparable to a small room heater running permanently. Four such machines in a modest cabin add four kilowatts — enough to overwhelm the cooling in most Indian office rooms, particularly outside working hours when building air conditioning is often reduced or off.

That last point causes a specific and common failure: overnight training runs. The machine is fine during the day when the office is cooled and fails or throttles badly at night when cooling is switched off, producing results that are inconsistent for reasons nobody connects to the building management system. If machines run overnight, the room’s cooling must too.

Electrical provisioning

Two Indian specifics compound this. Supply voltage variation is common and a power supply operating at the low end of its input range runs hotter and less efficiently, adding heat. And UPS sizing is frequently wrong: a UPS specified for a conventional desktop will not carry a kilowatt-class workstation, and discovering this during a power cut mid-training is expensive. If the work matters, size the UPS for the actual sustained draw with margin.

Detecting throttling

The reason thermal problems persist is that they do not announce themselves. The machine keeps computing, just more slowly, and unless someone is watching clock speeds nobody attributes a gradually slower workflow to heat. Establishing a baseline is cheap and worth doing at commissioning.

Monitor four values during a sustained load test: GPU core clock, GPU temperature, power draw and the throttle reason flags the driver exposes. Run a representative workload for at least thirty minutes and record the settled values. If clocks fall materially from their initial figures and the throttle reason indicates thermal or power limits, you have a problem to fix before it becomes a habit. Repeat the test in May and in December, because the answer differs.

What to fix, in order

Intake air temperature first, because it sets the floor everything else works against. A machine drawing cool air performs better than any amount of case fan tuning against warm intake. Position machines away from other heat sources and away from being tucked into enclosed desk cabinets, which is the single most common installation error.

Second, chassis airflow: a straight front-to-back path with adequate intake area, not a decorative case with restricted mesh. Third, dust filtration and a cleaning schedule, since Indian office and industrial dust loads are high and a clogged filter changes thermal behaviour within months. Fourth, only then consider undervolting or power limiting, which trades a little peak performance for substantially lower heat and often produces higher sustained throughput — counterintuitive but frequently true. The tier context is in the 96 GB desk-side tier.

When the office is the wrong place

Three signals that the machines belong in a server room rather than at desks. Persistent throttling that survives the fixes above, which means the room simply cannot reject the heat. More than two or three high-power machines in one space, at which point you are operating a small data hall without the infrastructure. And any requirement for overnight or weekend running, which conflicts with normal building cooling schedules.

Consolidating into a properly cooled room — or a server — usually improves performance and reduces noise and heat complaints simultaneously. It also changes the machines from personal equipment into shared infrastructure, which is generally the right direction once a team is running sustained AI workloads. The transition is covered in when a team outgrows AI workstations, and the cost comparison in AI workstation TCO in India.

Frequently asked questions