Graphics Cards

What does it do?

A graphics card or graphics processor unit (GPU) is very similar to a processor in that both are used to solved arithmetic and logical problems. Where the two components differ is in what types of logical and arithmetic problems they are designed to solve. In terms of hardware, processors usually have anywhere between one to eight cores whiles GPU's can have hundreds of microprocessors. This allows GPU's to carry out more tasks at once but only if those tasks are easy and not too complex. This is where the graphics part of GPU comes in. Typically, rendering a single polygon on a computer is a trivial task in terms of computing power. One microprocessor can handles this task with ease, but what if the computer is told to draw a complex shape? One that consists of multiple polygons?

GPU vs CPU

Lets define a shape with an arbitrary number of vertices, edges, and faces. When a particular computer receives the order to draw the shape, it tries to partition the shape in such a way so that it can be drawn using three dimensional triangles. Lets say the computer determines that two hundred triangles must be drawn to render the shape. This information is relayed to the GPU, which relays the information regarding the specific coordinate location of each vertex, the vertices that are joined by an edge, and the representation of each triangular face to each microprocessor. The work is then divided evenly across one hundred microprocessors meaning each microprocessor is in charge of rendering two simple triangular shapes. Since each microprocessor can work simultaneously among the other microprocessors, the shape is rendered in two iterations. Had the prior example been assigned to a processor with four cores, it would have taken fifty iterations for the shape to be rendered.

So if a GPU can handle more tasks at once, then what is the point of even having a CPU?
A GPU is only more efficient when it comes to executing a large number of simple instructions. On average, a single microprocessor on a GPU is much slower than a core on a multi-core processor. This means if an extremely difficult problem is passed to one microprocessor on a GPU, it will take a large amount of time for that single microprocessor to solve it compared to the faster core on a CPU. In general, CPU's can be thought of as the dedicated device that handles complex, lengthy instructions while GPU handles abundant, simple problems.

What a CPU solves	What a GPU solves
Derive the general solution of the Laplace transform of t^5 * arccosh(at+ qt^b)*sin(ct)*e^dt using a series expansion blindfolded, in a loud room, with a jumbo sized Sharpie, on a 3x5 note card.

More Processes = More Power

While GPU's can be considerably faster than CPU's for some tasks, they almost always take more power to function at an effective speed. With the high end GPU's today like the Titan X, power consumption is upwards of 250w while a high end Intel i7 processor only takes 84w to function. In some cases, people opt to put two GPU's in a single computer and bridge them together in a scalable link interface (SLI). By having two cards, the computer can share the processing load across two GPU's making the computer faster but a lot more power hungry. In particular, as discussed in the web page about processors, heat becomes a major issue in this case because GPU's will not be able to process data as efficiently when they are overheated.

GTX Titan GPU's in Four-Way SLI

Calculating Cost

Lets say a computer running a Titan X is left on for five hours each day for a year, how much does it cost to run the GPU per year?

Heat Expulsion

When dealing with a setup such as the one above, convection currents and fans are essential steps in designing a computer that can remove heat from multiple GPU's effectively. Below is a diagram of a HAF-X case which intakes cool air from the outside via fans located on the top, front, and bottom of the case. This cool air is directed towards the two GPU's in the bottom left hand corner of the computer case. It is then intermixed with hot air blown off by the fans on the GPU thus causing the heat to be transferred from the hot GPU's to the cool air via convection. As a result, the thermal equilibrium temperature reached between the GPU and cool air causes the GPU to transfer heat and therefore run more efficiently. It should be noted that in many setups, it is typical to have as many cold air intakes as possible and only one direction for the hotter outtake air. This setup features a convection zone border which is a fancy name for splitting the cooling process across two sets of intake and outtake fans. This helps ensure the cool air coming into the computer is having minimum heat exchange with the hot air blown out of the back.

Picture Sources:
multiplication table: http://www.math-aids.com/images/multiplication-drills.png
4 way SLI GPUs: https://content.hwigroup.net/images/products_xl/293029/3/nvidia-geforce-gtx-titan-x-sli-4-way.jpg
HAF-X: https://www.dataimage.com/jpeg;base64,/9j/4AAQSkZJRgABAQAAAQABAAD/2wCEAAkGBwgHBgkIBw