We talk a lot about graphics cards on this channel, because most of the time, it is the most exciting part of a gaming PC, and the biggest contributor to your rig’s gaming performance. But that means we’re talking a lot about technical terms like Core clock, GDDR5, CUDA cores, ROPs, and so on – if you know what all those terms mean, good for you. Gold star. But for those of you who DON’T know, for they laymen out there, this video should help you understand the basics what they mean, and whether you should be actually paying attention to those when shopping for a new GPU. Er, graphics card. The GPU is the actual chip, but sometimes we say that instead of – see, this is why we’re here.
Let’s actually start there – the terms graphics card, video card, and GPU are all used to refer to … one of THESE, but the term GPU, or graphics processing unit, actually only literally means the physical chip on the card’s PCB. See, I told you this was the basics. The reason this distinction is important is that different graphics cards can be based off the same GPU. For example, the two largest desktop GPU makers, AMD and Nvidia, will release a reference version of a new graphics card like the Radeon RX 480 or Geforce GTX 1080. Then, vendors like MSI, ASUS, Gigabyte, and so on, often sell that reference design, but they also design custom versions, with unique coolers, I/O ports, and sometimes, a higher clock speed. But all those unique versions have the same chip at their core. So they’re not really that unique. Are they?
It’s also possible to have 2 GPUs on one graphics card, like AMD’s R9 295X2, Radeon Pro Duo, or Nvidia’s Titan Z. Those can get you tremendous amounts of graphics horsepower, but, since you’re technically running TWO GPUs, the card’s performance in games will depend on how well those games support multi-GPU performance. Nvidia’s name for their multi-GPU connection is SLI, and AMD’s is Crossfire. Sometimes, like in the case of Nvidia’s “70” and “80” cards like the GTX 970, 980, 1070, and 1080, two graphics cards at different tiers will have the same chip, albeit with some modifications. The GTX 970 and 980 are both based off the Maxwell GM204 GPU, while the 1070 and 1080 are based off the Pascal GP104.
Now speaking of Maxwell, and Pascal – you might be aware that those terms refer to the GPU architecture, but.. What is that? A GPU’s architecture is the platform or technology it’s built on. New graphics architectures are developed by AMD and Nvidia around every 1-2 years, and they often shrink the size of the physical components of the processor, which allows them to fit more features and transistors onto the GPU die – “die” meaning the actual silicon in the chip. Architecture changes can also reduce the amount of power required to run cards. For this reason, it becomes problematic to compare different generations of graphics cards based on certain specs alone, as separate architectures use the resources available to them differently.
Speaking of those certain specs, let’s look at the first one. Memory. Video memory, or Video RAM, serves the same function as system RAM, in that it holds whatever data is currently being accessed by the GPU. These are the textures and images that make up what is displayed to you on your screen. VRAM becomes especially important when running at higher resolutions. If you want to know more about how much VRAM you might need in a graphics card, I wrote an entire video for Linus, actually, so you can click HERE for that video. For this video, we just need to know a couple things: Having more VRAM won’t automatically get you more frames per second, but having not enough could cause textures in-game to pop in and out. Right now, it seems like 4GB of VRAM is kind of the sweet spot for 1080p, with 2GB being found on lower end cards that would run with some settings turned down, and 6-8GB allowing the headroom that’s needed for 1440p and 4K. Again, though, that’s assuming the card has the raw power necessary to run those resolutions in the first place. And games will differ from each other in requirements, as well. Some games super needy and some will take whatever you’ve got, just like people.
One of the indicators of that raw power I mentioned is the Core Clock. This refers to the frequency at which the GPU is running, measured in Hz. While this can be used as a general measure to compare the power of different GPUs in one generation, it can’t reliably be used for comparison between different generations, as developments in architecture render those comparisons moot. Graphics cards will often have a Boost Clock, which is what speed the card can boost itself to under a heavy processing load, like a game. However, that speed will drop back down if it gets too hot, which is called thermal throttling, and it’s also why you want a good cooler on your GPU. Similarly, the video memory runs at a certain speed, called the Memory Clock. This is one of a few specs helps us determine the Memory Bandwidth. The others are Memory Bus Width, which is how many bits can travel to and from the VRAM each clock cycle, and the type of memory – either GDDR5, GDDR5X, or HBM, which is vertically stacked RAM that we’ve only seen in AMD’s Fiji GPU so far. Memory Bandwidth is measured in Gigabits per second – think of it as a tube connecting the GPU to its VRAM. The bigger that tube is, the more effectively your GPU can use its VRAM. Having tons and tons of memory won’t do any good if the the Bandwidth is too small to use all of it at once. You can’t squeeze an elephant through a tube sock. Not without some excellent engineering.
Alright – it’s time to speed this up. We’re gonna go down a typical spec list and give quick definitions for each. CUDA cores – CUDA stands for Compute Unified Device Architecture – it’s Nvidia’s custom programming language. CUDA cores are the physical cores in an Nvidia GPU – in AMD GPUs, these cores are called stream processors, and use the OpenCL programming language. We don’t compare these across generations and architectures for the same reason we don’t do it for Core Clocks. Transistor count is just what it sounds like – the number of transistors in the GPU, and you don’t need to worry about it except as an fancy high number.
Now these next specs really don’t matter unless you’re going to do an in-depth technical analysis of a graphics card.
Texture Units basically assist in applying textures to 3D models. ROPs, or Raster Output Pipelines, or render operations units, it doesn’t matter – they’re involved in the final process of outputting pixels to the display, or rendering. They also deal heavily with anti-aliasing.
TDP stands for Thermal Design Power, and it’s the maximum amount of heat the GPU is specified to produce when running normal applications, which is weird, because it’s measured in watts. This doesn’t correlate directly to the actual amount of power consumed, so don’t worry too much about its specific value unless you’re doing in-depth testing, in which case, you wouldn’t need to watch this video.
The ManufacturingProcess of a GPU refers to the “half-pitch” of a memory cell in a processor, or half the distance between identical features. It’s measured in nanometers, and a new process is developed every 2-3 years. The AMD’s RX 480 is made on a 14nm process, and Nvidia’s 10 series GPUs are made on a 16nm process, which is why this generation of GPUs has a pretty big performance leap over last gen’s.
Now, if all these specs don’t really help you when you’re looking to buy a GPU, what should you look for? Benchmarks. The only real way to know how a card is going to run is to.. Run it. Specs can give a general idea of how a card MIGHT perform relative to another, but there’s no way to know for sure until you put it to the test. So before you purchase a graphics card, do your research. Read and watch reviews of that card compared to other cards, and then you can make an informed decision.
Well, I think that’s about enough of that. Hopefully for the uninformed out there, this video helped clear things up a bit. If you want to see more, you can click HERE to watch more videos, follow us on Twitter over here, or like the video and subscribe to NCIX Tech Tips.