Video transcoding is a very resource intensive task. Sometimes it is quite expensive to execute it at the expense of the CPU, and graphics accelerators are used to save resources. This is especially true in OTT broadcasting, when there are many profiles per channel.
In this article, we will look at the main advantages and disadvantages of these solutions using the example of Intel’s QuickSync technology and Nvidia’s NVENC technology. Interestingly, being formally competitors in the field of coding, both companies are developing in parallel and even cooperating in the production of new chips.
First, let’s decide on the graphics that we will compare. We will consider only stable solutions for 24/7 work: it is impossible to do otherwise in TV broadcasting. Intel is simple: Take the latest generation of Coffee Lake Xeon processors with integrated Intel® UHD Graphics P630, Intel® Xeon® E-2246G processor. At Nvidia, the choice of a video card for transcoding is a little more complicated. We chose Quadro RTX 4000 (8GB) – this is the server counterpart of the GeForce RTX 2070 Super (8GB). Unlike RTX, it has no official restrictions on processing more than 3 streams simultaneously. This limitation can be removed by installing a patch created by the craftsmen, but, as we have already noted, we will consider only verified and official solutions. Earlier versions of video cards shook off at once: they lose when working with HEVC, since they do not have the ability to encode B-frames.
Maximum number of transcoded channels
So let’s move on to the numbers. Let’s carry out a load test for the maximum possible number of transcoded channels (fastest mode) on one server.
In this comparison, Nvidia has twice the performance of Intel in AVC transcoding and is practically on par with HEVC encoding.
Price for 1 channel including server costs
Now we know the maximum possible number of FHD (1920 * 1080) resolution channels per server with integrated Intel graphics and Nvidia graphics card, which means we can calculate the price for 1 FHD channel.
It turns out that there is no difference in price for AVC. In the case of HEVC, Nvidia is much more expensive for the price per channel per platform, if you calculate the maximum number of channels (that is, use the fastest encoding algorithms, sacrificing quality).
At this point, we will interrupt the calculations and move on to the issue of quality, since it will be much more honest to compare the equally acceptable quality, and not obtained in fast modes.
Output stream quality compared to original
Let’s consider another important issue – video compression quality. After all, there is no point in the number of channels if it is impossible to watch them.
Below is a graph of quality comparison by PSNR metric: Intel AVC with original stream (blue line) / Nvidia AVC with original stream (red line).
Based on the graph, we see that the quality of the received streams is close in terms of the PSNR value. Let’s compare using the VMAF metric.
In the following graph, let’s look at comparing Intel HEVC to original stream (blue line) / Nvidia HEVC to original stream (red line).
As you can see from the graph, our comparison was not entirely correct, since the maximum number of encoded Nvidia channels = 14, and their quality is almost 2 dB higher than 13 channels on Intel.
Therefore, we made a few more changes, and with the highest possible quality on Nvidia and Intel in GAcc mode (GPU Accelerated – when encoding occurs not only on the graphics, but the CPU is also connected), we got the following result. Intel HEVC GAcc with original stream (blue line) / Nvidia HEVC with original stream (red line):
The encoding quality is almost the same, but the performance of both systems has dropped significantly. Now Nvidia encoded just 4 FHD HEVC channels, while Intel only encoded 2.
Let’s recalculate the price per channel based on the new data:
1500/2 = 750 for one transcoded HEVC channel on Intel;
3000/4 = 750 per HEVC transcoded channel on Nvidia.
And, in fact, we got the same ratio for the price per channel as in the situation with AVC.