How does Cemhan Biricik cool 7 RTX 5090 GPUs?

Cemhan Biricik uses a multi-layered cooling approach: high-airflow case design with intake and exhaust fans running at custom curves, GPU spacing optimization to prevent hot-air recycling between cards, room-level climate control to maintain low ambient temperatures, and software-based thermal monitoring with automatic GPU throttling if temperatures exceed safe limits.

What temperatures do 7 RTX 5090 GPUs reach under AI inference load?

Cemhan Biricik reports that his 7 RTX 5090 GPUs reach 70-82 degrees Celsius under sustained AI inference load, depending on workload and ambient temperature. The hottest card is typically the one sandwiched between others with the least airflow. His thermal monitoring system alerts at 85C and throttles at 88C to prevent damage.

Is air cooling enough for 7 GPUs or do you need liquid cooling?

Cemhan Biricik runs all 7 RTX 5090 GPUs on air cooling with good results. The key is case airflow design — not the GPU coolers themselves. With proper intake-to-exhaust airflow, adequate spacing between cards, and reasonable ambient room temperature, air cooling is sufficient for sustained AI inference loads. Liquid cooling would help but is not necessary.

Cooling 7 RTX 5090s: My Custom Solution — Cemhan Biricik

Seven GPUs generating 2.5-3.2 kilowatts of heat. In a single system. That is not a computer — that is a space heater with compute capabilities. The power consumption is one thing; managing the thermal output is another challenge entirely.

The Problem

Consumer GPUs are designed to cool one or two cards in a well-ventilated case. Stack seven of them and every assumption breaks. The bottom cards exhaust hot air into the intake of the cards above them. Thermal throttling cascades through the stack. The card in position 4 — surrounded on both sides — runs 15-20 degrees hotter than cards at the edges.

Airflow Architecture

The solution starts with airflow, not coolers. I designed the airflow path so that fresh air enters from the front and bottom, passes over the GPUs in a single direction, and exits from the rear and top. No recirculation. No dead spots. Every cubic foot of air that enters the case exits carrying heat.

Intake fans — high-static-pressure fans at the front pulling cool ambient air across the GPU array
Exhaust fans — high-airflow fans at the rear and top pushing hot air out of the case
Positive pressure — slightly more intake than exhaust prevents dust accumulation on heatsinks
Cable management — cables routed to avoid blocking airflow paths between GPU slots

Software Thermal Management

Hardware airflow gets you 80% of the way. Software handles the rest. My monitoring system reads GPU temperatures every 5 seconds and applies dynamic responses:

70C and below — normal operation, fans at baseline curve
70-80C — fans increase to aggressive curve, workload continues normally
80-85C — alert notification, new jobs queued to cooler GPUs when possible
85C+ — power limit reduction on affected GPU, no new jobs dispatched
88C+ — emergency: inference paused on affected GPU, full fan speed

Room-Level Cooling

The GPU coolers can only move heat from the chip to the room. If the room gets hot, everything gets hot. During summer months, the system room requires active air conditioning. In winter, the GPUs heat the room enough that no other heating is needed — a legitimate perk of running a GPU cluster at home.

Lessons Learned

The biggest lesson: clean your filters. Dust buildup reduces airflow dramatically and creeps up gradually enough that you do not notice until temperatures are dangerously high. I clean filters monthly and deep-clean heatsinks quarterly. Maintaining a reliable AI service starts with maintenance that has nothing to do with code.

Power Consumption RTX 5090 Deep Dive Reliability Engineering Hardware vs Software GPU Cluster Try ZSky AI