AI PC s 2x RTX 3080

AI training server with 2X NVIDIA RTX 3080

We’ve been building compute servers for neural network computing in industry since 2018. In 2021, we built the first Training server that is built on two NVIDIA RTX 3080 graphics cards. This PC was developed for our major partner, the Czech company Smart View s.r.o., which specializes in the field of machine vision and automation. Smart View is an official partner of COGNEX and Ricoh Machin Vision and also a domestic manufacturer of machine vision lighting.

Our task was to build a very powerful computing server for neural network computing and AI. The main requirements from the customer were to deliver a server with two NVIDIA RTX 3080 graphics cards along with a COGNEX grabber card and a PoE IEEE 802.3af iEi card in a 4U RACK chassis. Initially this task may have sounded simple, but there were a few complications in the design that we had to resolve. We had to change the originally requested placement in the RACK chassis to a gaming chassis, specifically a Fractal Desingn Define, to give us enough space to mount all the cards and also to have room to address cooling for all the devices. We chose the motherboard to meet the requirements for enough PCIe slots and, along with the Intel® Core™ i7-9800X processor, to have enough PCIe lines to run all the cards.

We chose the graphics cards mainly based on their size so that we could fit additional cards to the PC. So we chose the smallest vented RTX 3080 on the market, the EVGA GeForce RTX 3080 XC3 Ultra Gaming. These cards offer ample performance when oveclocked, and there’s room to keep them cool when two cards are plugged in side by side.

Incidentally, the cooling gave us a bit of a workout too – the build is cooled by ambient air, which is drawn into the chassis by three powerful 120mm fans through the front vent. For better access of cool air to the graphics cards, we also installed one powerful 120mm fan to optimally ventilate the space between the graphics cards. The air exhaust from the assembly is provided by three 140mm fans which remove the hot air sufficiently.

Powering the entire assembly is a 1200W, 92% efficient, platinum certified power supply with an extreme +12V single rail design. When designing a suitable power supply, we calculated that the assembly should consistently draw around 850 W . We weren’t far off the mark – at full load, the sustained consumption was around 870 W, but the surprise was the fluctuations, where we almost reached the limit of the power supply – 1190 W! Long term testing and a burn-in process convinced us that the power supply is designed adequately.

The result of our work is a machine designed to run the CPU continuously at 100% load along with full load of both graphics cards, including overclocking from the nominal frequency of 1440 MHz to a stable 1700 MHz at a constant ambient temperature of 25 °C.

We firmly believe that this machine will serve its main purpose in the long run. We also believe that the end customer will be very satisfied with its extreme performance for neural network computation and processing.

Other similar types of these PCs are already slowly in the pipeline. If you are interested in supplying similar computing machines, please do not hesitate to contact us.