The domain of expertise required is High-Performance Computing (HPC) Systems Integration and Thermal Engineering. I will summarize this material from the perspective of a Senior Systems Integrator.
Abstract
This submission outlines the critical thermal remediation and integration process for a high-density, dual-GPU workstation leveraging two NVIDIA RTX 5090s and an AMD Threadripper 9980X 64-core processor. The system, initially air-cooled, suffered from acute thermal throttling, with the top GPU reaching 87°C under load, necessitating a complete custom liquid cooling deployment.
The mitigation strategy involved full disassembly of the highly complex 5090 Founders Edition coolers, installation of high-coverage water blocks (Heatkiller for CPU, Bixie for GPUs), and subsequent relocation into the McPro Apollo X chassis to accommodate extreme radiator volume. Key engineering challenges included integrating an 86mm thick 360mm front radiator and designing a bespoke, 3D-printed, toolless sliding rail mechanism for mounting the central pump/reservoir distribution block.
The project successfully resolved the thermal limitations. Post-integration metrics demonstrated a massive reduction in operating temperatures, with GPUs stabilizing below 50°C at full load. This thermal headroom enabled significant performance gains, including a reported boost clock increase of nearly 200 MHz and substantial power efficiency improvements (GPUs drawing approximately 400W versus a stock range of 575-600W).
Custom Liquid Cooling Deployment in High-Density HPC Workstation
0:00 Initial Thermal Constraint: The system, configured with dual RTX 5090 GPUs and a 64-core Threadripper 9980X, was rendered "unusable" in its air-cooled state due to the top GPU reaching a critical temperature of 87°C within minutes of rendering, leading to thermal throttling and fan noise.
0:46 CPU Block Selection: The Threadripper CPU was integrated into the custom loop using a Heatkiller water block, noting that although the 64-core chip requires less cooling than anticipated out of the box, full-coverage liquid cooling was adopted for system consistency.
1:13 GPU Disassembly Complexity: The disassembly of the RTX 5090 Founders Edition cards was intricate, highlighting the multi-component construction, including a separate PCIe connector assembly (1:57) secured by six screws, multiple ribbon and two-pin cables requiring careful removal (2:26), and a retention bracket design that relies solely on pre-bent tension without springs (3:06).
3:36 Founders Edition PCB Analysis: The 5090 PCB is noted for its extremely dense, small form factor, which facilitates unobstructed airflow in the stock configuration. The use of liquid metal as the factory Thermal Interface Material (TIM) was also confirmed (3:48).
4:21 Water Block Integration: The GPUs were fitted with single-slot Bixie water blocks. The blocks are compact (234mm in length), with approximately half the block’s length consisting of empty space designed to route the internal IO cable assembly (4:48).
5:04 Die Protection: The 5090 die features a three-barrier gasket system surrounding the chip, intended to prevent the liquid metal TIM from migrating onto the surrounding circuitry.
5:44 Chassis Inadequacy & Migration: The initial Phanteks case proved unsuitable for the proposed extreme liquid cooling components due to limited mounting space for large radiators.
6:16 New Chassis Specification: The system was re-housed in the McPro Apollo X case, selected for its capacity to handle large radiator volumes, including an 86mm thick Alphacool 360mm "monster" radiator at the front, complemented by a secondary 240mm radiator at the bottom (7:25).
8:48 Pump/Reservoir Mounting Solution: Due to severe space constraints in front of the motherboard, the Steel Key UNI240D reservoir and Alphacool Apex D5 pump required a custom mounting solution. A sliding rail mechanism was designed via 3D scanning and modeled in Fusion, then printed in resin on a Formlabs Form 4 to ensure tight tolerances (9:58).
12:28 Validated Thermal Performance: Under full load (Blender rendering), the liquid-cooled 5090 GPUs maintained temperatures below 50°C. Maximum reported load temperature, even with side panels installed and the Threadripper running, was below 60°C.
12:50 Performance and Efficiency Gains: The reduction in thermal stress resulted in boost clock increases of nearly 200 MHz. Furthermore, the GPUs exhibited significantly lower power consumption, drawing approximately 400-450W under load compared to the standard air-cooled TDP of 575-600W (13:18).