After all, Wang Yinan and Chen Zihao risked their lives to get seventy-four special CPUs using the latest technology of the toothpaste factory, of which sixty-four were working hard in the cabinet.

The supercomputing power of the Gödel system finally came in handy this time.

The specifications of the video source chosen by Wang Yinan are standard 1080P, that is, 1920 x 1080 resolution, and the bit rate of H265 is about 50% of that of H264, that is, 5000Kbps, which is about 0.7M bytes per second.

As a comparison, ordinary home broadband, if you use ADSL line, is basically 1Mbps or 2Mbps, about 140K or 280K bytes per second, which is far from the 800K bytes per second required by the H265 standard, so on ADSL broadband, there is no way to watch 1080P H265 video online, let alone H264 with higher bandwidth requirements.

Therefore, ADSL users can only download video files and watch high-definition videos offline, of course, after fiber to the home, access 10Mbps or 20Mbps bandwidth, you can support the demand for online viewing of 1080P videos.

If it is a three-dimensional video, then the left eye and the right eye are separate video streams, so the bandwidth demand is doubled, if it is a panoramic video, then although the range of human eyes is relatively small, but you can't wait for the eyes to turn over, and then transmit the part to be seen, that kind of experience is too bad, so the panoramic video needs to transmit video data from all angles at the same time, so it is at least three times more bandwidth overhead.

Three-dimensional plus panorama, which is commonly referred to as VR, brings at least six times the bandwidth requirements of ordinary HD, which is also the fundamental reason why VR live broadcast is currently hot, but it is basically not experienced, because not to mention the family, the office network of general companies does not have such a high bandwidth.

The length of the whole video is about fifteen minutes, and the neural network of Gödel's system only took less than twenty seconds to complete the compression process, "No", Wang Yinan didn't believe his eyes, "Such a fast speed, it can't be a bug",

This situation is not uncommon, if there is no convergence in the neural network training process, it is very likely to cause an invalid data processing process, that is, there is nothing in the compressed data, just some random noise.

Wang Yinan glanced at the size of the final generated data, 150M, that is, the neural network compression algorithm of the Gödel system, the compression efficiency is nearly five times that of H265, and the speed of compression and decompression is not known for the time being, because after all, it is a supercomputer, and the actual speed depends on the execution result after compiling into x86 or ARM instructions.

"No", Wang Yinan was stupid by this size, 150M size, can store almost 1000 seconds of 1080P video content, that is to say, the bit rate is 150K bytes per second, which is very close to the home ADSL if 1M broadband, the maximum data transmission rate that can be achieved.

In other words, if you use the neural network video compression algorithm of the Gödel system, with a little optimization, or a little reduction in quality, you can achieve 1080P full HD video live broadcast on the ordinary most Lo broadband.

In other words, Internet celebrity users, when using 4G for live broadcast, can now only play 480P live content, and using Wang Yinan's neural network compression algorithm, you can achieve full HD mobile phone live broadcast, of course, this also puts forward higher requirements for the appearance of Internet celebrities, after all, a small pimple on the face will be clearly seen by the audience in the live broadcast.

"Calm down, calm down", Wang Yinan kept reminding himself that he still needed to see if he could restore the video, this kind of joke has not happened before, there was once a company that claimed that no matter for any data, it can be compressed to one hundredth of the original size, and it also came up with a program that can indeed compress any file to one percent of the original size.

At the beginning, people were shocked, but if you think about it carefully, it's not right, this is against the basic laws of science, and then later, everyone found that this program can be compressed, but it can't be restored, and guess what, the company said, the description of my program is written like this, it can be compressed to one percent of the original size, but there is no guarantee that it can be restored!

If the Gödel system gave Wang Yinan such a hand, it would be crying, but if that's the case, I guess Wang Yinan can't care about crying, and quickly smash the Gödel system.

Wang Yinan inputs a compressed 150M video file into an inverse neural network, the format of which is completely customizable, and it is estimated that each video is different, and only the Gödel system or equivalent LLVM code can recognize it.

"It's going to be troublesome in the future.,There's no format and specification for video compression at all.,Neural network chooses the most efficient way.,Compress as much as you want.,Anyway, as long as you recognize it.,Maybe only the legendary four-dimensional space model,Can explain the compression mechanism inside the neural network.。

Wang Yinan was thinking about it for a while, the beep beep sounded again, looked at the time, and the decompression time was less than ten seconds, which is also normal, any compression algorithm will theoretically take more time than the decompression algorithm.

However, the proportion of this time is also beyond Wang Yinan's expectations, because according to the experience of H264 or H265, the computing power required for compression is at least ten times that of decompression, that is, under the same hardware, if it is an algorithm similar to H264 and H265, the decompression time should be about two seconds.

"It seems that the compression algorithm of the Gödel system is completely different from the existing system, but in this field, as in Go, the machine has far surpassed the human, at least, the best algorithm conceived by man is more than five times worse than the neural network."

"It's just that human beings will never be able to understand the algorithm of the Gödel system", Wang Yinan sighed,

"It doesn't matter if you don't understand, just be able to copy it, I'm really a genius, I actually thought of the trick of functional replication, but I don't know if Mige's dog lets people copy it."

Happy to happy, Wang Yinan still didn't forget the last step, he asked the Gödel system to play back the decompressed image sequence on the screen at a speed of 30 frames per second, Wang Yinan took a few paragraphs and watched it for a while, at least the naked eye can't see any difference with the original film source.

However, it is useless to feel subjectively, and it still needs data to speak, Wang Yinan ran an open-source PSNR evaluation program, that is, to compare the peak signal-to-noise ratio between the original image and the compressed/decompressed image, but something unexpected happened to Wang Yinan.

The difference in peak signal-to-noise ratio is very large, that is, the evaluation program believes that the difference between the two images before and after is very large, even greater than the difference between H265 and compressed to H264 after doubling the bit rate.

But this is completely opposite to Wang Yinan's subjective feelings,Wang Yinan has also seen the H264 video that has been reduced by half of the bitrate,The quality of the whole picture has dropped very obviously,What blurs、Glitches have appeared,But now the video stream,Wang Yinan doesn't feel any stiffness in the image。

What's the problem?,It should be said,Wang Yinan still has a strong confidence in his eyesight,You know,When I watched Ms. Bodo's films,Wang Yinan can even tell the difference between the bitrate 500K and the bitrate 550K.

It seems that the problem can only lie in the PSNR algorithm, because if there are some small changes in the structure of the image, or the overall displacement, etc., it will not affect the overall subjective feeling of the video, but PSNR looks like the error is very large.

Wang Yinan immediately consulted the relevant literature, and sure enough, there was a lot of complaints about the PSNR algorithm on the Internet, and on the open source machine vision website, Wang Yinan found the latest complex wavelet structure similarity index algorithm, which should be the best machine vision detection index in line with the subjective evaluation from the perspective of multiple literatures.

Wang Yinan downloaded the latest complex wavelet structure similarity index algorithm code, and re-evaluated the decompressed image sequence.

The results are also unexpected, and the overall difference index is very low, or even negligible, which is also consistent with Wang Yinan's subjective feelings.

Wang Yinan ran to the bathroom, rinsed his face with cold water, and calmed down his excitement a little, watching 1080P high-definition video on the 1M ADSL line, or, under the ordinary 10M, 20M household fiber broadband conditions, watching the three-dimensional plus panoramic 360-degree real-time live broadcast, what does it mean?

This is not only a breakthrough in video compression technology, but may even bring about a revolution in the entire video entertainment industry!

Now the only question remains, whether it is feasible to compile the neural network into the IR code (machine-independent middle-layer representation) of the LLVM, and what level of performance can be achieved with the X86 or ARM instructions generated by the IR code of the LLVM?

If you reach the level of Gödel's system, you don't have to think about it, if you achieve real-time decompression, that is, the speed of decompression can keep up with the speed of network transmission, it will be terrible.

If you accidentally achieve real-time compression, that is, the speed of compression can keep up with the speed of the actual picture, and it is an incredible square!

Wang Yinan slept well, then cleaned himself up, and went to Zhou Hui for a big meal, well, it is said that the buffet of Pangu Baxing is good, and you can also see the scenery of my imperial capital, although Wang Yinan is not very interested in the cuisine of the Sun Country, he still went to have a beautiful meal.

Next, Wang Yinan started the retreat again, after the performance of the algorithm was confirmed, the motivation was completely different, the neural network compilation system, this kind of big killer, it was exciting to think about.

Can the compiled code run?

Can a huge neural network be compressed into the memory of an ordinary PC or mobile phone?

Is the running code efficient enough, can it achieve real-time decompression, or real-time compression?

Or are all the ideas too optimistic and slow like a snail?