Custom Case for 8-13 GPU Rendering and Learning Servers

Most of you have heard exciting stuff happening in the High-Performance Computing (HPC) world, and you probably also heard that deep learning, AI and rendering requires a lot of hardware. Powerful hardware that regular computer cases and fans are not designed to handle. Numerous people have asked us what kind of computer case and fan are best for building a personal 8 to 13 GPU beast, and with this blog, I hope to answer them.

GPUs are well known in the gaming world where ultra-fast, graphic intensive rendering is essential. These high power computing components are designed specifically to handle mathematically intensive tasks. That makes them perfect in the area of AI analytics, big data performance, and 3D rendering. The problem is, with high performance comes excess heat. A lot of heat.

Powerful GPUs dissipate up to 300 Watts of heat, each. And most GPUs will decrease their speed to prevent overheating as soon as a card hits 78-82 °C range. The result is a decreased performance which can be significant for multiple GPUs (10-25%) where each GPU heats up the GPUs next to itself. This thermal throttling is a bottleneck that reduces performance more than poor hardware choices do.

Typical computer cases and fans are badly designed for multiple GPU server/workstation. You may be fine with standard computer cases and fans for a light-duty desktop PC, but for a heavy-duty GPU server/workstation you will need to make special considerations. Specifically, you need a custom case optimized for massive cooling and airflow capacity.

At Nanoxia, we are always looking into novel hardware for scientific computing. We make fully custom chassis that uses GPUs from the ground up. If you are looking to unleash the full power of 8-13 triple-or-quadruple slot GPUs without building something the size of a football stadium next to a river for hydropower and cooling, the Hydra 6, 7, and 8 chassis might be suitable for you.

Of these three cases, the Hydra 6 is the largest (almost a meter wide!) and designed with massive GPU thermal capacity in mind. Coupled with the optional 14 x Hydra 120mm high speed server fans, the Hydra 6 chassis supports an insane amount of airflow (2,590 CFM) and cooling capacity (201,600mm²). This dwarfs the amount of cooling air required by rack-mounted servers (160 cubic feet per meet per kilowatt) and blade servers (120 CFM per kW).

 
Pre-built GPU machines are certainly powerful, but not everyone can afford a Microway Octoputer, a SuperMicro SuperServer, or a Nvidia DGX System. These perform exceptionally well at numerical simulations related to processor intensive tasks (i.e. astrophysical research) but are quite expensive. As a value proposition, you can build 8 to 13 Radeon Vega 56 cards with HBM (or HBM2) high bandwidth memory. You can also build 8 to 13 Intel Phi 7220 Coprocessor cards without worrying about being thermally-constrained. All these alternative solutions offer significant value compared to the Nvidia Tesla / DGX line.

You might be wondering, can I mine cryptocurrency with this? Yes, of course, that's when Hydra 8 comes out! What if you prioritize power and portability? That would be Hydra 7.

Now go power everything from Adobe Premier and 3D visual rendering to AI analytics and astrophysical research into black holes.