As an engineer in a development environment, you've done your research and selected the best Graphics Processing Unit (GPU) or video card for your application.
You've weighed number of cores against core clock frequency and considered the memory bandwidth.
Now, you are headed to product development, and it's time to move your application to a rack mount computer. Depending on the industry you're in, this can be an industrial PC or even a ruggedized military computer system.
Selecting the right GPU or video card to work seamlessly within the rack mount chassis can help ensure the success of your program and vastly enhance computing capacity.
Below are 5 things to consider.
Rule number one: the GPU must fit inside the chassis. PCIe video cards and GPUs typically come in the following dimensions:
In a 1U server or 2U rack mount computer, the cards are typically installed horizontally, although a 2U chassis will accommodate a low-profile card vertically. In a 3U server or higher, the cards are typically installed vertically.
Watch out for the added space required if your GPU or video card requires additional power. The PCIe or EPS power connectors require additional space that can frequently be overlooked, so, pay close attention to the location of the power connectors. A 3U chassis can fit a full-height card, but only if there is no additional power connector or if the power connector is on the rear of the card. If the connector is on the top of the card, it will add about 25mm with the connector and cable bend, causing mechanical interference with the lid of the chassis.
Not all PCIe slots are created equal. PCIe slots can vary in speed (Gen1, Gen2, Gen3), bandwidth (x1, x4, x8, x16), and latency. It's important to check the block diagrams of the motherboard, or the system host board and backplane to see what the attributes are for each slot. Applications with a requirement for low-latency communication with the CPU will work best with the GPU or video card installed in a PCIe slot with direct access to the CPU. Going through a PCIe switch would add delays in communication with the CPU. For applications that need to optimize communications between cards, offloading the communication to a PCIe switch might work best.
Accommodating the ever-increasingly-power-hungry GPUs and video cards can be a challenge, particularly in a rack mount system. According to the PCIe 3.0 specification, the maximum power draw from a PCIe slot is 75W. If additional power is needed, then a six-pin PCIe (75W) or eight-pin PCIe (150W) connector can be used. Some of the latest high-power GPUs are starting to utilize an 8-pin EPS connector, drawing a maximum of 336W. This is just one card, so, keep in mind the number of cards that will be installed and the total system power required. With multiple cards, power delivery becomes more important if the power supplies segment power through separate rails. Drawing all of that power through a single rail can lead to power supply failure, so, make sure you understand how the power will be delivered.
Read further on selecting a power supply for your rackmount computer.
When discussing GPUs in rack mount servers and their power requirements, heat dissipation is never far behind. The greater the power draw of a video card, the greater the heat that is generated, which must be expelled from the chassis. GPUs can generally dissipate heat in two ways. The first is to exhaust heat into the chassis. The second is to exhaust heat out of the rear I/O. If the heat is exhausted into the chassis, it will be important to have proper chassis cooling fans moving the heat quickly out of the chassis. Also, consider slot location to ensure proper airflow between cards. Two cards side by side will have one exhausting heat almost directly onto the other card. If the card utilizes a blower, exhausting the heat outside of the chassis through the rear I/O, you only need to be sure of proper chassis airflow to the blower.
Product life cycle can be critically important, particularly to industrial and military computing applications. It can take months or years to design a product, qualify that product for the application, and then certify it with a government-certifying body. When a part changes, it can be a real pain to re-design, re-qualify, re-certify and have to dump additional resources into an existing product instead of developing the next great product. GPUs or video cards can have a product life cycle as short as 18 months, so, make sure you consider either choosing a longer-life GPU or video card or planning on an upgrade path for card obsolescence.
It is vital that you think about the different system components your GPU may affect. An embedded computer in a rack mount chassis has its limitations with space and airflow; therefore, your graphics card needs to accommodate its environment for optimal performance, not the other way around. Pay close attention to these constraints when selecting a GPU or video card to avoid project delays in procurement and to avoid shortening your computer's life cycle.