Since it is the PCI Express connectors that are responsible, on the one hand, for feeding it with the necessary electricity and, on the other, for transmitting the data through it. And this is where we come in with two major problems. The first of them is that the data transfer required between the GPU and the RAM is too high so that its consumption is not recommended in an external interface for peripherals and let’s not talk about the energy consumption, which causes the external GPUs require separate power supplies.
Why are they hard to see?
In reality there are no external graphics cards, but boxes that allow us to place these inside them and that use high-speed external interfaces to connect. Initially, many of them used external PCI Express ports, but their high consumption just to communicate led to the adoption of interfaces such as Thunderbolt. So you should ensure that your PC has these connections. In the specific case of the box, you should make sure that it contains a sufficiently powerful power supply and that its interior allows you to place the model you want to install.
All this becomes a huge extra cost and if we add to this that many of them are models that are only sold by the manufacturer of a specific PC model and being an exclusive piece they charge for it and not precisely for a low price, since that we can find external boxes that can cost as much as a mid-range model and then have to add the cost of the entire computer, which is usually higher than the desktop equivalent. So with these premises, those of you who read this will understand the reason why they are not lavished much.
We also do not see them in mining, since this adds to the amount of money to be recovered and, therefore, they are much less profitable, despite the fact that in this application the performance problems related to rendering such as the speed in which it achieves the DisplayList and graph data batches are not a problem.
Does an external graphics card perform the same as an internal one?
The simple and direct answer to the question is no and there are a number of bottlenecks associated with it, but it is necessary to understand the communication that usually exists between the graphics card and the system processor in order to see through the logic and knowing where the bottleneck is.
For each frame that is generated the GPU needs a screen list that is created before each one by the CPU. This list is not stored in the video RAM, since the processor does not have access to it. Therefore, it is necessary for the graphics card to fetch the information from the RAM and this is done through the PCI Express port, which, unlike other interfaces for peripherals, does not have to go through the IOMMU and this allows it to reduce the latency.
Keep in mind that when rendering a scene every millisecond counts and things don’t start when the scene starts to draw, but much earlier and that’s why getting the DisplayList on an external graphics card reaches the same speed as with an internal one is impossible.
The graphic batch problem
The so-called graphics batches are an important point in gaming performance, by which we mean the huge amount of data that is transmitted from the system RAM to the graphics VRAM through the PCI Express port. It is not only the screen list but all the visual data like textures, models, shader programs that are necessary to build each frame that appears on the screen.
The problem is that the bandwidth for it is much lower and this means that the speed at which the data is transmitted is also lower, which is another negative point in terms of performance.
Is there a solution to the limitations?
The reason why the conventional PCI Express interface is short in length is because if we lengthen the length of the cable, the resistance created would end up greatly increasing power consumption. Therefore, external PCI Express interfaces usually have a much lower number of connection lines for the same consumption. This is less bandwidth for the same power consumption, which is fatal.
That is why an optical interface such as Thunderbolt was chosen, which by its nature does not suffer from the problem of increasing resistance with distance. However, it is limited in terms of bandwidth and this is something that adds additional latency to the communication of the external graphics card. In other words, outsourcing it results in a significant loss of its graphic power.
At the moment there is no solution at the domestic level, although in supercomputers and large servers optical PCI Express interfaces are being used. Which have the bandwidth of an interface of the same name but without the limitations caused by an interconnection by copper cable. That is, being able to maintain the same bandwidth regardless of the distance from the graphics card. Of course, there are no plans to implement this connector in laptops.
