I'm planning an always-on local inference system with two ASUS TUF RTX 5090 OC cards, each roughly 3.6 slots thick, installed in an ASRock X870E Taichi Lite inside a Fractal Design Torrent. The system will also use a Ryzen 9 9950X, Thermalright Phantom Spirit 120 EVO, 1600W power supply, and a 4TB Samsung 990 Pro. I'd power-limit both GPUs to around 450W. My priority is having access to the combined 64GB of VRAM rather than maximizing peak throughput. Has anyone successfully run two cards this large on air cooling in a high-airflow case? Will the four-slot spacing and power limits keep the upper card at reasonable temperatures, or would liquid cooling or a riser-mounted card be necessary? Also, do both slots operate at full bandwidth, or are they limited to PCIe 5.0 x8 when both GPUs are installed?
4 Answers
Power-limiting to 450W will reduce heat, but it won’t eliminate the main issue: the upper card’s cooler will still be starved for fresh air. Professional multi-GPU cards or blower-style designs are better suited to close installations. Before committing, I’d test the exact card dimensions against the lower-slot position and consider an external or riser-mounted arrangement rather than relying on the case fans alone.
Both GPUs won’t run at PCIe 5.0 x16 simultaneously on this platform. With both slots populated, the layout is effectively PCIe 5.0 x8 for each card. That still provides bandwidth comparable to PCIe 4.0 x16 and is generally adequate for inference, but it isn’t full x16 bandwidth. Also verify that your software actually supports pooling or sharing VRAM across two consumer GPUs; having 64GB installed does not automatically create one unified 64GB memory pool.
Check the physical layout carefully before buying. The lower slot on that motherboard sits unusually low, while the Torrent’s bottom fans occupy much of the space beneath the board. Two cards this thick may not physically fit, even though the board advertises four-slot spacing. A riser could solve the clearance and airflow problems, but you’d need to secure the card properly and account for the cable routing.
The upper card is very likely to run hot, even with the power limit. These open-air gaming coolers need space to pull in air, and a thick lower card will block most of the intake area. If there isn’t enough room to get a finger between the cards, expect the top GPU to recirculate hot air. A riser-mounted GPU, an open bench, or water cooling would be much safer than stacking them normally.
Even two thinner Founders Edition-style cards can show a noticeable temperature difference when stacked. Cards that are over three slots thick will make the airflow problem substantially worse.