Dual GPU Powered by RTX 5090: Power for Neural Networks and Complex Computing
Project Description
We were contacted by a team of developers focused on training deep neural networks (LLM) and high-performance computing. The customer's main problem was that standard gaming solutions could not cope with intense workloads 24/7, and the server racks were too noisy to be placed directly in the office.
- Scope of activity: AI development, Machine Learning, Data Science.
- Software used: PyTorch, TensorFlow, Docker, CUDA-dependent libraries.
Key requirement: Maximum video memory and data bus bandwidth, stable processor operation under 100% load and complete absence of overheating while maintaining a comfortable noise level.
Upgrade configuration
- Graphics card:2 x Palit GeForce RTX 5090 GameRock [32GB, 21760 CUDA]
- CPU:AMD Ryzen Threadripper PRO 7975WX (32 cores, up to 5.3GHz)
- Motherboard:MSI MPG X870E CARBON [DDR5, Wi-Fi 7]
- RAM:256GB Samsung ECC DDR5 4800MHz
- Drives:4TB Samsung 9100 PRO
- Cooling:HYPERPC CUSTOM
- Case:Fractal Design Meshify 3 XL
Selection process and decision
The Threadripper PRO platform based on the WRX90 chipset was chosen to implement the project. This is the only solution that allows you to fully use two flagship RTX 5090 graphics cards at full speed of the PCIe 5.0 interface without a bottleneck.
The main challenge was cooling. The two new generation cards generate enormous amounts of heat. We decided to completely abandon air cooling in favor of a custom liquid circuit. The use of water blocks made it possible to make the system two-slot and compact, while maintaining the temperature of the chips 30–40 degrees lower than in the air. ARCTIC S4028-6K high-speed fans were integrated to cool the VRM and RAM areas, ensuring the stability of the ECC server components.
Result
We have created a true computing center in a desktop PC form factor. Thanks to custom water cooling and fine-tuning fan curves, the station remains quiet even when training the neural network for several days. Result: The customer received a 2.5 times increase in model training speed compared to the previous generation of systems, and the use of ECC memory and an industrial power supply eliminated the risk of data loss during critical calculations.
Did you like the upgrade?
Want to order a similar computer or create your own unique project? Contact us for a consultation - we will discuss every detail of your future PC.