They usually have a name that indicates the generation along with a number, the higher the number the better the graphics card, at NVIDIA we have for example the AD102 GPU, which has several variants, such as the AD102-225-A1 which is used in the RTX 4070 Ti SUPER, but which has a modified chip, which is the original and belongs to the RTX 4090, the AD102-300-A1.
GPU Cores
As with the GPU, actually trying to call the computing units offered by a graphics card as cores is somewhat erroneous, but we use them every day, although they do not represent the same thing because obviously they are not cores like those found in processors. The GPU of a graphics card has hundreds of small elements grouped in clusters that give rise to the computing units, which we can find with different names depending on the brand, CU (Compute Units) by AMD, Xe-cores in Intel or streaming multiprocessors in NVIDIA.
Memory: VRAM and bandwidth
Both VRAM and bandwidth are probably the easiest specifications to see on current graphics cards, since unlike the other aspects we should look at, the video memory itself is often indicated in the name of the graphics card since we have surely seen the name of the card at some point. 8 GB either 16 GB of some graphics. The type of memory that the graphics card uses indicates how fast and efficient it is, for example a memory GDDR6 is less than one GDDR6X, which in turn will be surpassed by GDDR7.
On the other hand, we have the memory bandwidth, which is usually represented in GB/s, and obviously the higher it is, the better speeds the graphics card will offer. The memory speed together with the bus width (384-bit, 256-bit, 128-bit, etc.) determine the memory bandwidth.
Clock speeds
In this case, it is not too complicated to know what this specification represents, and it is that, just like a CPU, the GPU also has a parameter to measure the clock speed, which is usually represented in MHz. Basically, it determines the speed at which the GPU is capable of processing instructions; there is not much more mystery, since, just like with processors, a greater amount represents greater power.
TGP
If this technical specification sounds familiar to you, it’s because it’s very similar to the TDP (Thermal Design Power) offered by processors, and in this case it refers to the maximum energy that the graphics card is capable of consuming during use. We can take the RTX 4090 as an example, which offers a TGP of 450 W, which implies that this is the maximum it will consume when using it, although there may be peaks that exceed it when using high-performance applications, but they usually do not last more than a few tenths of a second.
As we can imagine, this is an aspect directly linked to temperature, since in the end it not only indicates what it consumes, but also the temperatures that you can expect the component to reach while you are using it.
FP32 Performance
Finally, we have the single-precision floating point format, an aspect that we don’t usually take into account if we don’t know certain technical details of how these types of devices work. Basically, it is a mathematical way of measuring the theoretical performance offered by the GPU, and you may have seen what its unit of measurement is, TFLOPS, which refers to the trillions of operations that the device is capable of performing per second.
As you can imagine, a higher number of TFLOPS indicates that the graphics card has a higher performance. If we have one with 9 TFLOPS and another with 35 TFLOPS (as is the case with the RTX 4070), the one capable of offering us 35 TFLOPS would obviously win.