The ultimate server for deep learning and neural networks
Project Description
We were approached by a customer who was faced with the task of deploying a local computing station to work with heavy artificial intelligence models and complex 3D rendering. The main challenge was the need to combine extreme multi-threaded performance with the ability to work for long periods under 100% load in a typical office environment.
- Scope of activity: Research in the field of AI (Deep Learning), machine learning and professional visualization.
- Software used: TensorFlow/PyTorch library stack, 3D rendering software (Octane, Redshift), virtualization environments.
Key requirement: Uncompromising reliability for 24/7 operation and efficient heat dissipation from two flagship graphics cards and a 96-core processor without turning your office into a noisy server room.
Upgrade configuration
- Graphics card:2 x Palit GeForce RTX 5090 GameRock [32GB, 21760 CUDA]
- CPU:AMD Ryzen Threadripper PRO 7995WX (up to 5.1GHz, 96 cores)
- Motherboard:ASUS PRO WS WRX90E-SAGE SE [DDR5]
- RAM:256GB (2x128GB) Samsung ECC DDR5
- Drives:3 x 4TB Samsung 9100 PRO
- Cooling:HYPERPC CUSTOM
- Power supply:3000W ASUS PRO WS [80+ Platinum]
- Case:CORSAIR 9000D White
Selection process and decision
To solve such an ambitious task, we abandoned standard consumer solutions in favor of the AMD Ryzen Threadripper PRO platform. The choice of 7995WX is due to the presence of 96 cores and 192 threads - this is the de facto power of a full-fledged server, packaged in a workstation format.
Particular attention was paid to the graphics subsystem. Two NVIDIA GeForce RTX 5090 graphics cards provide a colossal amount of video memory and tensor cores. To avoid overheating and throttling, we integrated a complex custom water cooling system (WCO) that covers both the CPU and both graphics cards.
The CORSAIR 9000D RGB AIRFLOW case was chosen as the “foundation”. Despite the name, the client asked for minimal visual noise, so we customized the lighting in a stark white style to highlight the cleanliness of the build, and installed 18 high-performance Thermalright fans optimized for low noise at high static pressure.
Result
We created not just a computer, but a scientific instrument. Thanks to liquid cooling, the temperature of graphics cards under full load does not exceed 55-60°C, which guarantees the stability of neural network training, which can last for weeks. The customer received the performance of a small server cluster that fits under a desktop and is quieter than a typical gaming PC. The training time for test models was reduced by 4 times compared to the previous solution based on previous generations of GPUs.
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.