GPU Overview
The One Constant in GPU Computing is Change
You have heard the line “constant is change” a million times, but it is true. GPU computing is driven by the e-games market where rendering lifelike scenes and quick trigger responses push technology and competitive spirit to buy the latest tech for the home rig. It is a fast-moving market and what’s hot today might not see life in 12 months.
The GPU market is also influenced by activities such as crypto currency mining. When prices are up, hundreds (even thousands) of GPUs are joined solving complex math problems to receive a valuable “coin.” Although not as much of an influence today, the practice can affect the availability and price of certain GPU classifications.
Some manufacturers have been responding to the needs of industrial computing and the explosion of AI (Artificial Intelligence) applications. If your build might be repeated in volume, and you need a consistent product, consider seeking out an industrial model which should have a longer life cycle.
If you are looking for a GPU to bundle into your custom build, you already know for what purpose. Small low power GPUs are a great addition to add a bit more sophistication to a high-end display and return some compute power to the CPU. Some vision processing tasks are a good match for medium class GPUs and need some auxiliary power. These GPUs are also well used for the execution of AI inference activities. Complex high-end GPUs are the choice for crunching high data streams and learning algorithms. These high energy use GPUs are often used in tandem for the most demanding industrial applications.
Size, compute power, and electricity consumption are key considerations and GPUs can come in various configurations. Consider the footprint of your chosen computer and compare the length and width of the card. Width is measured in “slots” so if you are considering a computer with multiple PCIe (PCI Express) slots, you might lose one or more due to the overhang of the card.
More powerful cards require more electricity, although there are some pleasant exceptions to the rule. If your card’s power consumption exceeds 75W, then axillary power must be available in the computer. When thinking about energy use, this is a good time to consider a proper power supply. Underpowering a system can cause many problems.
Power consumption is related to heat generation which is an important consideration to industrial computing. There are a variety of fanless, and cassette style configurations specifically designed for certain high-performance GPUs. Although fans are the most common way to move ambient air for cooling, there are configurations available that will facilitate GPU and heat dissipation under water.
Your application and the amount of computer power you need might not always be coordinated. Desktop and rack style systems are a common setup for the office environment. Industrial shoebox configurations that keep the compute section fanless and isolated can be as large as 21 liters, single GPU shoebox types are 12-15 liters. Compact rugged computer with a cassette for the GPU are 4-5 liters. For sub compact and edge CPU/Accelerated computing, check out the Jetson platform and Hailo cards.
