Hashrate Exchange Rate

Lin Yuan summarized all problems into two points: computation and collection.

Regarding the computation problem, Lin Yuan had always wanted to know the actual performance of this computing power system. He wondered how much real-world computing power 1P of this system's computing power actually equated to.

Previously, he didn't have professional computing power equipment to test, but now, due to the testing task, wasn't Good Group's computing power equipment readily available?

Good Group's Zijin Branch had a test computing platform with a mix of H100 and A100 GPUs. Lin Yuan successfully applied for the right to use one A100. Although the A100 was significantly inferior to the H100, it was, after all, a genuine computing power card, easily outperforming various RTX and HD graphics cards used in computers on the market.

Lin Yuan first ran a general test targeting TF32 data. The A100's computing performance for TF32 generally stabilized above 35T/S. His own computing power system only needed approximately 10T/S of computing power to achieve the same level of performance.

Based on this, Lin Yuan finally calibrated the current computing performance of the computing power system: it was approximately three times better than real-world computing power. In other words, 1T unit of computing power from the computing power system was equivalent to 3T in the real world.

That is to say: the exchange rate between the system's computing power and the real world's was 1 to 3.

Lin Yuan suddenly felt a sense of "I am a developed country, and the entire real world is the third world," because his computing power was valuable.

But even so, Lin Yuan had little confidence in using the computing power system to run algorithm optimization. Because the computing power system's exchange rate was only 3 times, his accumulated 1000P computing power was simply not enough.

However, even out of curiosity, he still had to try it.

Lin Yuan then prepared to load a collected set of data into the computing power system. This data was real delivery data from a large group of delivery riderss over a period of time.

Lin Yuan quickly encountered a problem: how the hell was he supposed to input the data? This data was too large, a full 10GB, and it would take forever to input it manually by eye.

"System, I told you, put a WiFi in my brain, then you can communicate with electronic devices in the future."

However, the system did not respond to him.

It wasn't until Lin Yuan was about to give up that the system proactively popped up a prompt.

"Data loaded. Start training?"

'When was it loaded?'

Lin Yuan was astonished to see the exact same data on the system panel as on the server in front of him. The system read over 10GB of data in just a few moments, and Lin Yuan hadn't noticed the entire process.

"So you're not just limited to inputting data by eye."

Lin Yuan suddenly realized: could it be that the system itself was connected to the real world?

Without time to think further, Lin Yuan, with this curiosity, began the AI model training for the system.

He set the computing power to 50T/S, a safe power value that wouldn't cause him dizziness.

To Lin Yuan's surprise, the system completed a data training session in just one minute. 50T/S * 60S = 3000T, which was only 3P of computing power consumption.

Lin Yuan knew clearly that an A100 running at full load for the same data training would take ten minutes.

Given that the A100's full-load computing power was 35T/S, the computing power generated by an A100 running at full load for ten minutes was: 35T/S * 600 = 21000T = 21P.

In other words, to complete the same workload, the computing power system only used 3P of computing power, while the A100 used 21P. Since the A100's computing power value was a real-world computing power value, even if an H100 computing card were used, its work efficiency would improve, but the total computing power consumed would not change.

This led to the conclusion: for training this 10GB dataset, 1P of the computing power system's computing power was equivalent to 7P in the real world. The computing power exchange rate then became 1 to 7.

This didn't match the previous 1 to 3.

Although this system was mysterious, so far Lin Yuan had not found any signs of it breaking natural laws.

AI model training was essentially data computation. Since it was data computation, there shouldn't be a situation where it computes slowly one moment and quickly the next, because the data type Lin Yuan used hadn't changed; it was all TF32.

'Could the computing power system be anthropomorphic? Sometimes working fast, sometimes slow?'

With this question in mind, Lin Yuan used the computing power system to compute the same data at different times.

He tried it while eating, while on the toilet, before going to bed, and even while watching videos, but the final computing power consumption was constant.

'Damn it, is your exchange rate still fluctuating?'

Lin Yuan then switched back to using the computing power system to calculate the original TF32 test data, but he oddly obtained the initial 1 to 3 computing power exchange rate.

He then had to switch to another delivery data set, also around 10GB, and something strange happened: the computing power exchange rate became 1 to 6.

Subsequently, Lin Yuan continuously changed data samples, and finally he discovered: as the selected data sample changed, the computing power exchange rate actually varied.

Damn it, it was normal for the computing power consumed to change with different data samples. But what the hell was a changing computing power exchange rate? A changing computing power exchange rate meant that the computing power system's efficiency varied when processing different data samples.

This was like a computer running different programs, where different CPU usage was easy to understand. But for this computer to have its maximum CPU frequency change when running different programs? That was anti-science.

"Hey, System, tell me, are you a high-dimensional being?"

"The rules of the three-dimensional world can't explain you anymore."

"Ah ah ah~~~"

Finally, Lin Yuan had to delve into studying the AI model training process of the computing power system. Because when a black box exhibits problems, but you can't find the source of the problem, you can only dive into that black box to look.

Although, this black box was very complex.

Lin Yuan, with a nervous heart, asked the system to display the detailed training process.

"This..."

Lin Yuan's eyes widened. In the AR image floating in the void before him, the structural diagram representing the model was actually changing.

AI model training was essentially using a computing power card to throw collected data into a preset model for computation.

A model could be roughly thought of as a formula. (Anyway, if you want, you can even see the entire universe as a formula)

So, AI model training, simplified to the extreme, was: y = f(x).

x represented sampled data.

f represented the model.

y was the calculation result.

In reality, AI model training involved changing the parameters of f during the training process, for example: f = 2x+1 would become: f = 3x+1 as it ran. But it would absolutely not turn into f = 3x+1 / x+1 while running.

Unless the model was manually changed and training rerun.

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

In other words, the formula f was constantly changing with training. If it was changing, then it would be strange if the computing power exchange rate remained fixed.

This was like a computer that could change its own CPU structure while running.

This... an incredible thought suddenly popped into Lin Yuan's mind.