The world of GPUs can be a scary place fraught with big words, bigger numbers, and lots of confusing nomenclature. Allow us to un-confuse things a bit for you
The amount of memory a GPU has is also called its frame buffer (see below). Most cards these days come with 1GB to 3GB of memory, but some high-end cards like the GTX Titan have 6GB of memory. In the simplest terms, more memory lets you run higher resolutions, but read the Frame Buffer section below for more info.
GPUs nowadays include compartmentalized subsystems that have their own processing cores, called Stream Processors by AMD, and CUDA cores by Nvidia, but both perform the same task. Unlike a CPU, which is designed to handle a wide array of tasks, but only able to execute a handful of threads in parallel at a high clock speed, GPU cores are massively parallel and designed to handle specific tasks such as shader calculations. They can also be used for compute operations, but typically these features are heavily neutered in gaming cards, as the manufacturers want their most demanding clients paying top dollar for expensive workstation cards that offer full support for compute functionality. Since AMD and Nvidia's processor cores are built on different architectures, it's impossible to make direct comparisons between them, so just because one GPU has more cores than another does not automatically make it better.
The memory bus is a crucial pathway between the GPU itself and the card's onboard frame buffer, or memory. The width of the bus and the speed of the memory itself combine to give you a set amount of bandwidth, which equals how much data can be transferred across the bus, usually measured in gigabytes per second. In this respect, and what generally stands with all things PC, more is better. As an example, a GTX 680 with its 6GHz memory (1,500MHz quad-pumped) and 256-bit interface is capable of transferring 192.2GB of data per second, whereas the GTX Titan with the same 6GHz memory but a wider 384-bit interface is capable of transferring 288.4GB per second. Since most modern gaming boards now use 6GHz memory, the width of the interface is the only spec that ever changes, and the wider the better. Lower-end cards like the HD 7790, for example, have a 128-bit memory bus, so as you spend more money you'll find cards with wider buses.
This technology is available in high-end GPUs, and it allows the GPU to dynamically overclock itself when under load for increased performance. GPUs without this technology are locked at one core clock speed all the time.
The frame buffer is composed of DDR memory and is where all the computations are performed to the images before they are output to your display, so you'll need a bigger buffer to run higher resolutions, as the two are directly related to one another. Put simply, if you want to run higher resolutions—as in fill your screen with more pixels—you will need a frame buffer large enough to accommodate all those pixels. The same principle applies if you are running a standard resolution such as 1080p but want to enable super-sampling AA (see below): Since the scene is actually being rendered at a higher resolution and then down-sampled, you'll need a larger frame buffer to handle that higher internal resolution. In general, a 1GB or 2GB buffer is fine for 1080p, but you will need 2GB or 3GB for 2560x1600 at decent frame rates. This is why the GTX Titan has 6GB of memory, as it’s designed to run at the absolute highest resolutions possible, including across three displays at once. Most midrange cards now have 2GB, with 3GB and 4GB frame buffers now commonplace for high-end GPUs.
High resolutions require a lot of RAM, which is embedded in the area around the GPU just like on this 6GB GTX Titan.
All modern GPUs use PCI Express power connectors, either of the 6-pin or 8-pin variety. Small cards require one 6-pin connector, bigger cards require two 6-pin, and the top-shelf cards require one 8-pin and one 6-pin. Flagship boards like the GTX 690 and HD 7990 need two 8-pin connectors. Most high-end cards will draw between 100–200W of power under load, so you'll need around a 500–650W PSU for your entire system. Always give yourself somewhat of a buffer, so when a manufacturer says a 550W PSU is required, go for 650W.
These are what connect your GPU to your display, the most common being DVI, which comes in both single-link and dual-link. Dual-link is needed for resolutions up to 2560x1600, while single-link is fine for up to 1,200 pixels vertically. DisplayPort can go up to 2560x1600, as well. HDMI is another connector you will see: versions 1.0–1.2 support 1080p, 1.3 supports 2560x1600, while 1.4 supports 4K.
PCI Express 3.0
The latest generation of graphics cards from AMD and Nvidia are all PCIe 3.0, which theoretically allows for more bandwidth across the bus compared to PCIe 2.0, but actual in-game improvement will be slim-to-none in most cases, as PCIe 2.0 was never saturated to begin with. Your motherboard chipset and CPU must also support PCIe 3.0, but most Ivy Bridge and older boards do not support it in the chipset, even though the CPU may have the required lanes. In general, every GPU has PCIe 3.0 these days, but if your motherboard only supports version 2.0 you will not suffer a performance hit.
GPU coolers fall into several different categories, including blower, centralized, and water-cooled. The blower type is seen on most "reference" designs, which is what AMD and Nvidia provide to their add-in board partners as the most cost-effective solution typically. It sucks air in from the front of the chassis, then blows it along a heatsink through the back of the card to be exited out the rear of your case. Centralized coolers have one or two fans in the middle that suck air in from anywhere around the card and exhaust it into the same region, creating a pocket of warm air below the card. Water-cooled cards are very rare, of course, but use water to absorb heat contained within a radiator, which is cooled by a fan. Water cooling is usually the most effective (and quiet) way to cool a hot PC component, but its cost and complexity make it less common.
This is Nvidia technology baked into its last few generations of GPUs that allows for hardware-based rendering of physics in games that support it, most notably Borderlands 2, so instead of just a regular explosion, you will see an explosion with particles and volumetric fog and smoke. Typically, AMD card owners will see the PhysX option grayed out in the menus, but the games still look great, so we would not deem this technology a reason to go with Nvidia over AMD at this point in time.
Different GPUs offer different types of antialiasing (AA), which is the smoothing out of jaggies that appear on edges of surfaces in games. Let's look at the most common types:
Full Scene AA (FSAA, or AA):
The most basic type of AA, this is sometimes called super-sampling. It involves rendering a scene at higher resolutions and then down-sampling the final image for a smoother transition between pixels, which appears like softer edges on your screen. If you run 2X AA, the scene will be calculated at double the resolution, and 4X AA renders it at four times the resolution, hence a massive performance hit.
Multi-Sample AA (MSAA):
This is a more efficient form of FSAA, even though scenes are still rendered at higher resolutions, then down-sampled. It achieves this efficiency by only super-sampling pixels that are along edges; by sampling fewer pixels, you don't see as much of a hit as with FSAA.
Fast Approximate AA (FXAA):
This is a shader-based Nvidia creation designed to allow for decent AA with very little to no performance hit. It achieves this by smoothing every pixel onscreen, including those born from pixel shaders, which isn't possible with MSAA.
This is specific to Kepler GPUs and combines MSAA with post-processing to achieve higher-quality antialiasing, but it's not as efficient as FXAA.
Morphological Antialiasing (MLAA):
This is AMD technology that uses GPU-accelerated compute functionality to apply AA as a post-processing effect as opposed to the super-sampling method.
Click the next page to learn more about wi-fi technology, RAM, and PSUs.