Graphics hardware is complex to say the least, since when we go to buy one we find a list of features on the graphics card and acronyms that pull back the newer users and, therefore, with less knowledge. If this is your case then do not worry about it, since in this article we will teach you to read all those characteristics as an expert.
The process of assembling a PC for the first time on our part is a job that goes beyond assembling the pieces, it also consists of choosing the ones that are suitable. Due to the continuous release cadence we find in stores with graphics cards of different generations and under the same price. So the end user without knowledge can find a worse specification having paid the same.
That is why it is necessary to know what the characteristics of a graphics card mean before making your purchase to know if it is what we need,
Models
One way to identify a graphics card is through the model it is, with this we can know if we are facing a top of the range within its generation or another with more modest specifications.
AMD
If our GPU is from AMD then we must look for those that start as RX, and discard very old models that have acronyms such as R9, R7. And yes, in case you didn’t know, the acronym RX stands for R10, but AMD kept the name for marketing reasons. The next thing you will see are 3 figures if you are between one that has a GPU with GCN architecture or 4 figures if it is RDNA, the latter are better in specifications and in the case of the RX 6 × 00 you will have DirectX 12 Ultimate support.
As for the second figure, the brand in which range the graphics card is, so an RX 6900 will be better than a RX 6800, which at the same time will have better characteristics than a RX 6700. To finish you can find XT models, these are somewhat better specs.
NVIDIA
In the case that your GPUs are from NVIDIA, if you want to have a graphics card that is up to date you have to look for one that starts with the initials RTX instead of GTX, since this will ensure you have full support for DirectX 12 Ultimate. As for the four-digit number, the first two indicate the generation and it grows by tens. So an RTX 30 × 0 will be of a more advanced generation than an RTX 20.
The last two figures indicate the range to which it belongs and, therefore, the power within the same architecture. So the RTX 3090 is more powerful than the RTX 3080. From time to time NVIDIA releases Ti models that have better specs than normal models and also often releases revisions after a while that are given the SUPER surname.
The GPU and its features
At the center of the graphics card is a complex piece of hardware that we know as the GPU or Graphical Processing Unit, a graphic processing unit, and which is responsible for generating the scenes that appear on your monitor in the video memory or VRAM that the surrounds. The end of the process is the sending of said image to the monitor so that our eyes can see it and this is done several dozen times per second.
Today, GPUs are pieces of hardware that have a level of complexity in their architecture at the level of a central processor or CPU, however, it is not necessary to know all the characteristics of the graphics card for purchase. On the other hand, the term is also often used to talk about the entire graphic hardware, which leads to confusion despite the fact that this piece of hardware cannot function without the rest of the elements that surround it.
GPU architecture
With architecture we refer to the surrounding elements of the same and the way in which they are distributed. The explanation is as simple as that more advanced architectures allow adding new visual capabilities, higher performance in higher resolution mode or higher speeds when playing our favorite games.
Both AMD, NVIDIA and Intel with each new architecture improve the specifications of the previous ones. In the case of Intel, since the ARC Alchemist is its first dedicated graphics card, it is not a confusion, on the other hand, in the case of NVIDIA and AMD, several generations of graphics cards have appeared in recent years.
In addition, the name of the graphic architecture is not the same as the commercial name of the graphic card, that is why many times you will find certain graphic cards mentioned not by the name of the box, but by their architecture. In the same way, we can also see what the architecture of the GPU is by looking at the silkscreen printing on its encapsulation, at least in the case of NVIDIA’s.
Size
Because not everyone has the same monitor, or budget, and not the same needs, within the same GPU architecture, several models are launched in the form of different chips. How do these differ? Well, in key points such as:
- The interface with the video memory and, therefore, the bandwidth with it and the video memory in total. Keep in mind that the more bits of bandwidth you can connect more VRAM chips. A good way to see if a graphics card is of a lower end than another is from the amount of memory around the GPU.
- The amount of processing units inside the GPU , as is obvious, the less space inside the chip, less space can be placed inside it and this has an effect on having less computing power.
All GPUs within the same architecture, despite having different sizes, share all the common elements that define it, but at different power levels. The fact that one model has a GPU size and less memory than another also translates into a price range for all needs.
Shader units and “cores”
One marketing pitfall that graphics card manufacturers perform is to call what are 32-bit floating-point computing units as cores. Realistically, a kernel is the minimum unit that can execute a complete instruction in a program, so an execution unit that is only a part of the kernel cannot be considered a kernel in itself.
Shader units in AMD’s case are called Compute Units or Workgroup Processors depending on the architecture we are talking about, instead, NVIDIA has been using the term Streaming Multiprocessors or SM for some time. However, when they speak of cores in their specifications, they generally refer to units capable of doing 32-bit floating point arithmetic calculations under the name of CUDA, NVIDIA, or Stream Processors, AMD cores.
In fact, since these calculation units are within what are shader units, the fact that a GPU has more CUDA cores or more Stream Processors means that it is more powerful than another that has less.
GPU clock speeds
In the characteristics of the graphics card you are going to find various clock speeds that the GPU can reach at certain times. It must be taken into account that most of the time the GPU will remain at a relatively low clock speed in order to consume as little as possible, not generate heat and not turn on the fans of its cooling system.
However, when we start a graphically heavy application such as rendering a 3D scene with Blender or when playing a video game then it will start up at its base clock speed marked in its specifications, so your system cooling will start up and the GPU will operate all the time at that frequency, at least until we need it.
In some configurations the graphics card remains dormant, since the CPU usually includes an integrated GPU inside it for functions as simple as moving the Windows desktop.
Boost frequency
In some GPUs you will see a higher clock speed, this is the Boost frequency, which works the same as in a CPU. So for a limited time the clock speed will go up to speed up the rendering of the scene or to make the necessary calculations as fast as possible. The Boost frequency, however, cannot be reached for long, since the higher the consumption then the more temperature released and that is dangerous for the GPU in the long term. In the case of AMD models, there is a lighter Boost frequency called Game Clock, which has a longer duration in time as it is not as large a rise as that of Boost.
Many manufacturers when giving the figures of their technical specifications use the speed rates from the Boost frequency as if it were active all the time. While it is true that such performance can be achieved, this can only be done temporarily.
Fixed function units
They are so called due to the fact that they always perform the same function, so it is not programmable, but is fixed 100% of the time. The function of these units is to carry out a series of processes that are repeated continuously in each frame when rendering a scene recursively.
Fill rate and texturing
Fill rate is the number of pixels that a GPU writes onto video memory to create the frame over it. This does not coincide with those sent to the screen, since due to the way a 3D scene is generated, a pixel in VRAM can be written several times. For this, a type of unit called ROPS is used, Raster Outputs, which receive the already processed pixel and copy it into the video memory.
Texturing rate, on the other hand, has to do with texture units and they work in reverse. While the ROPS writes to memory, the texture unit reads from the stored texture maps of the VRAM to know the base color value of that pixel in that particular object.
The fill rate is read in pixels per second, while the texturing rate is read in texels per second. Accompanied by a letter in front of the term, where M stands for millions and G for billions.
The power in the features of a graphics card
The most common way to represent the computing power of a GPU is through the number of 32-bit floating point operations it can perform. Well, in reality it is not like that, since we are talking about the rate achieved at the time of carrying out a complete instruction that is the addition with multiplication or FMA that consists of two operations per clock cycle.
So how is computing power calculated? Well, very simple, with the following formula:
Clock speed * number of Compute Units or Stream Multiprocessing * CUDA cores or Stream Processors per unit * 2 = FP32 computing power.
Being the most simplified form:
Clock speed * CUDA cores or Stream Processors per unit across the entire GPU * 2 = FP32 computing power.
Resolution
Another of the characteristics of a graphics card that we have to look at is the supported target resolution, at the level of moving desktop applications any GPU can move them without problems, however in terms of gaming it is different and we find that within of each architecture these are distributed according to the resolution for which they are intended.
Currently the high-end ones are for 4K and, therefore, a resolution of 3840 x 2160 pixels, the mid-range is QHD with 1440p resolution and to finish the low-end is in Full HD or 1080p. In each range the power required to render quasi doubles, but in reality this is not the case and today low-end cards are more prepared to move scenes in Full HD than high-end cards, those with four times the resolution.
The VRAM of a graphics card and its characteristics
VRAM is the memory that the GPU uses to work, it is a local type and therefore it is generally out of range of the CPU. That is, it cannot be used by the central processor of the PC to execute programs and, therefore, adding more VRAM through the graphics card does not mean expanding the RAM memory of the system.
Currently there are two types of VRAM memory that we can find in the characteristics of a graphics card. On the one hand, the GDDR where they are already in the sixth generation and have a form of several chips mounted on the same board as the GPU. The other type of VRAM that we have is the HBM, which can be identified by being mounted on the GPU on a small base, which we call interposer, which is located above the card plate.
Due to the specifications of both types of video memory, we can find them in different markets, currently GDDR5, GDDR6 and GDDR6X are the most common in the domestic market and HBM2 graphics card memories for much more professional environments.
VRAM bandwidth
If we have a GDDR type memory then we have to count the number of chips in the graphics card and multiply them by 32 to find out the number of data pins between the VRAM and the GPU. Only the one-sided ones, since there is the possibility of making two memory chips share an interface. That is the case with the RTX 3090 Ti, which has 24 memory chips, but only 384 bus bits, 12 times 32 bits. In the case that we have an HBM memory, what we will do is count the number of chips and multiply them by 1024.
The second point is to look at the amount of Gbps of memory specifications, this is the amount of information measured in bits per second, not bytes, that is transmitted in one second for each data pin. So to have the bandwidth of the VRAM we will only have to multiply the Gbps by the number that you have previously obtained.
Since GPUs are processors that depend on data throughput for their performance, lower bandwidth on one model with the same GPU compared to another means lower performance.
Amount of VRAM memory
The last thing we have to know is the storage capacity of the video RAM, currently the most used memory is GDDR6, which can have 1 or 2 GB capacity chips. In PC it is not normal to see asymmetric configurations of chips in capacity, so a graphics card with 8 memory chips can have 8 GB or 16 GB. On the other hand, unlike the interface, the GDDR6 chips on both sides do count when it comes to obtaining the capacity.
As for the HBM memory, each chip is several memories stacked and assembled as one, so you should look at the specifications for the amount of memory available.
Other features of a graphics card
Apart from the GPU and the video memory, there are other characteristics in a graphics card that we should not overlook and therefore are also important.
PCI Express
The graphics cards are connected to the PC using the 16-pin PCI Express interface, which allows communication between and the rest of the PC, specifically with the CPU and RAM of the system. However, despite the fact that the connector is generalized in shape, new versions of the standard have been appearing with increasing speed. So if you are going to buy a graphics card with a PCI Express 4.0 interface and your motherboard supports a 3.0 type, you will be able to use it, but you will lose performance by not having enough bandwidth for communication.
The support of one version of the PCI Express standard or another also helps us to mark the age of one graphics card over another, so keep this detail in mind so that you do not get an outdated graphics card in its specifications.
Feeding
Graphics cards require electrical power to operate and the PCI Express port can deliver up to 75W of power. However, there are graphics cards that need additional power in order to function and usually come with an additional connector that gives it extra power apart from that provided by the communication interface. The 6-pin connector has the capacity to deliver 75 W of additional power, the 8-pin 150 W.
Some graphics cards have a 12-pin connector, which can give up to 500 W of additional power to a GPU and exceed the limit of 375 W. In any case, the more consumption then more power, so it is a way for users levels to know if they are facing a graphics card with high-end specifications or not.
Video outputs
A graphics card needs to send the images it creates to the monitor, however not all graphics cards have video outputs and although this may result in a contradiction, in reality it is not, since there are graphics cards whose specifications can be seen they do not have video outputs, this is because they are used for scientific computing or mining.
The most used video outputs today are HDMI and DisplayPort, so if you find yourself with a graphics card with DVI or VGA outputs then it will show that you are facing a very old model and that, therefore, not it will reach the minimum specifications. This does not usually happen, but be careful that they do not strain you in that regard.
