Client task
The client required a local computing node to train neural network models and work with large datasets. The main problem with existing solutions was the limited amount of video memory and overheating under prolonged loads.
- Scope of activity: Machine Learning (ML) and Data Science.
- Software used: Python library stack (PyTorch, TensorFlow), NVIDIA CUDA.
Key requirement: The possibility of parallel operation of four flagship GPUs in a single circuit and the availability of RAM with a capacity of 2 TB or more for processing data in RAM.
Selection process and decision
When designing the system, the main emphasis was placed on balancing throughput and efficient heat dissipation.
- Platform: Processor selected Threadripper PRO 7995WX based on the WRX90 chipset. 96 cores allow you to effectively parallelize data preprocessing tasks before feeding them to the GPU.
- Graphics cards: Installation of four RTX 6000 Blackwell required the use of a specialized motherboard with a sufficient number of PCIe 5.0 lanes operating in x16/x16/x16/x16 mode.
- Cooling: Standard air cooling for four server cards in one case is ineffective due to the dense packaging. We implemented a custom liquid cooling system covering the CPU and all GPUs. This made it possible to maintain operating temperatures within 55–60°C under full load.
- Power: The peak power consumption of the system exceeds 2 kW. Power supply installed 3000W, providing the necessary margin on 12V lines and Platinum certification to reduce heat loss.
Result
The designed system allowed the customer to abandon the rental of cloud capacity in favor of local computing. The station operates stably 24/7, providing a total of 384 GB of video memory for the most resource-intensive tasks.
Did you like the project?
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.
Arabic 
English