Xueba's computing system

Chapter 22 Hash Rate
Lin Yuan summarized all the problems into two points: calculation and collection.

Regarding the computing problem, Lin Yuan has always wanted to know the actual performance of this computing system. How much computing power does 1P of this system roughly equal to in the real world?

In the past, he did not have professional computing equipment for testing, but now because of the testing task, the computing equipment of the good group is ready-made.

The Zijin branch of Haotuan has a test computing platform that mixes H100 and A100 GPUs. Lin Yuan successfully applied for the right to use an A100. Although A100 is much worse than H100, it is a serious computing card after all, and it can definitely beat all kinds of RTX and HD computer graphics cards on the market.

林远先跑了次针对TF32数据的通用测试，A100对TF32的计算性能基本稳定在35T/S以上。自己的算力系统要到达同级别表现只需要将近10T/S的算力功率。

Based on this, Lin Yuan finally calibrated the current computing power performance of the computing power system: about three times better than the computing power in the real world. In other words, with the same 1T computing power, the computing power system is equivalent to 3T in the real world.

In other words: the computing power exchange rate between the system and the real world is 1 to 3.

Lin Yuan suddenly had a feeling that "I am a developed country, and the entire real world is the third world", because my computing power is valuable.

But even so, Lin Yuan had no confidence in using the computing power system to run algorithm optimization. Because the computing power exchange rate of the computing power system was only 3 times, his 1000P accumulated computing power was not enough.

However, even if it is just out of curiosity, you still have to give it a try.

Lin Yuan immediately prepared to load the collected data into the computing system. This data was the real delivery data of a large group of riders over a period of time.

Lin Yuan soon ran into a problem: How the hell was he going to input the data? The data was too large, a full 10G, and it would take forever to input it with his own eyes.

[System, I told you, build WiFi in my brain, and you will be able to communicate with electronic devices in the future. ]

However, the system ignored him.

It was not until Lin Yuan was about to give up that the system popped up a prompt.

[The data has been loaded. Do you want to start training? ]

'When did it finish loading?'

Lin Yuan was surprised to see the exact same data on the server in front of him on the system panel. The more than 10G of data was read by the system in just a few moments, and Lin Yuan was unaware of the entire process.

[It turns out that you can input data with more than just your eyes. ]

Lin Yuan suddenly realized: Could it be that the system itself is connected to the real world?
Without having time to think too much, Lin Yuan started systematic AI model training out of curiosity.

He set the computing power to 50T/S, which is a safe value that will not cause him dizziness.

To Lin Yuan's surprise, the system only took one minute to complete a data training. 50T/S*60S=3000T, which means only 3P of computing power consumption.

Lin Yuan knew clearly that it would take ten minutes for A100 to run the same data training at full capacity.

已知，A100的满负荷算力为：35T/S，则A100满负荷运行十分钟产生的算力是：35T/S*600=21000T=21P。

In other words, to complete the same amount of work, the computing system only used 3P computing power, while the A100 used 21P. Since the computing power value of the A100 is the computing power value in the real world, even if the H100 computing power card is used, its work efficiency will be improved, but the total computing power consumed will not change.

进而得出：训练这份10G数据集，算力系统1P的算力相当于现实世界的7P。那算力汇率就变成1比7了。

This doesn't match the previous 1:3.

Although this system is magical, Lin Yuan has not found any signs that it has broken the laws of nature so far.

AI model training is actually data calculation. Since it is data calculation, there is no such thing as slow calculation at one time and fast calculation at another, because the data type used by Lin Yuan has not changed, it is all TF32.

"Is the computing system anthropomorphic? Sometimes it works fast, sometimes it works slowly?" With this question in mind, Lin Yuan used the computing system to calculate the same data at different times.

He tried it while eating, squatting, before going to bed, and even while watching a movie, but the final computing power consumption was constant.

'Fuck, your exchange rate is still fluctuating?'

Lin Yuan then used the computing power system to calculate the original TF32 test data, but he happened to obtain the original computing power exchange rate of 1 to 3.

He had to change another takeaway delivery data of about 10G. A strange thing happened. The computing power exchange rate became 1 to 6.

Then Lin Yuan kept changing the data samples, and finally he found that the computing power exchange rate actually changed with the selected data samples.

Damn, it is normal for the computing power consumed by different data samples to change. But what is the change in computing power exchange rate? The change in computing power exchange rate means that the efficiency of the computing power system when processing different data samples is different.

It's like the same computer. When running different programs, the CPU usage is different, which is easy to understand. But when this computer runs different programs, the highest CPU frequency actually changes. This is unscientific.

[Hey, system, tell me, are you a higher-dimensional creature? ]

[The rules of the three-dimensional world cannot explain you. 】

【Ah ah ah~~~】

Finally, Lin Yuan had to study the AI model training process of the computing system in depth. Because when a black box shows a problem, but you can't find the problem, then you have no choice but to drill into the black box to see.

Although, this black box is very complicated.

With a nervous mood, Lin Yuan asked the system to show the detailed training process.

"this..."

Lin Yuan widened his eyes. In the AR image that appeared in the void before his eyes, the structural graphics representing the model were actually changing.

AI model training is essentially using computing power cards to throw the collected data into a preset model for calculation.

A model can be roughly thought of as a formula. (If you want, you can even think of the entire universe as a formula.)
Therefore, the ultimate simplification of AI model training is: y=f(x).

x represents the sample data.

f stands for model.

y is the calculation result.

现实中的AI模型训练是在训练过程中改变f的参数，比如：f=2x+1，跑着跑着就会变成：f=3x+1。但绝对不会跑着跑着变成f=3x+1/x+1。

Unless you manually change the model and re-run the training.

However, the computing system changed the model structure of f during training.

In other words, the formula f has been changing with training. If it is changing, then there is no way the computing power exchange rate can be fixed.

This is like a computer that can change its CPU structure while running.

This... An incredible idea suddenly popped up in Lin Yuan's mind.

(End of this chapter)