If you cheat, money will follow

Chang Le left the Medical Research Institute and came to the Artificial Intelligence Research Institute.

Li Fei, Suzkovi, Krichevsky and others are organizing personnel to debug the new computing platform.

Hinton coached.

The Lingxi 100 carbon-based computing power card has been shipped to Jiangzhou from Shanghai and has been installed.

New hardware requires optimization and debugging of model algorithms.

Algorithm optimization and debugging takes time.

The eliminated silicon-based computing power cards will be supplemented into weak models and public computing platforms to increase computing power.

"How is the computing power?" Chang Le walked into the center and asked.

"Very amazing." Li Fei replied:

“FP64 and FP32 computing power performance reaches 600TFLOPS.”

“The computing performance of FP16 reaches TFLOPS.”

"TF32 computing power performance is TFLOPS."

"There is also FP8 computing power performance of TFLOPS."

"And its single card power consumption is only 90 watts."

Li Fei lamented: "Carbon-based chips are too abnormal."

Huahong Semiconductor has produced a Lingxi 100 single card like this.

"Horizontal comparison, how is the performance?" Chang Le asked again.

Li Fei thought for a while and couldn't give an answer.

Because the horizontal comparison object cannot be found.

At this time, Sutskvi thought of the computing power card with the closest performance: A100.

He said: "Boss, there is really no horizontal comparison object."

"If you must compare, you can compare it with NVIDIA's A100, which has not yet been launched."

"The computing performance of a single card of Lingxi 100 is theoretically 100 to 30 times that of A64."

NVIDIA A100 computing performance is as follows:

FP32: 19.5TFLOPS (100 times that of Lingxi 30);

FP64: 9.7TFLOPS (100 times that of Lingxi 60);

TF32: 156TFLOPS (100 times that of Lingxi 64);

……

In terms of power consumption, Lingxi 100 is one-quarter of A100.

The parameters of A100 have been announced, but it has not been launched yet.

However, after the official release of carbon nanotubes, rumors about Nvidia A100 have disappeared.

The media is done with no interest in hype.

There is a high probability that Nvidia will not launch this computing card that will be eliminated immediately after its launch.

Not to mention the Lingxi 100, even the newly launched Shengteng 920 from Juchang has 5 times the performance.

So how much does Lingxi 100 cost?

The cost of a single card is 15 yuan.

The whole set is replaced, 375 billion.

This is equivalent to nearly two months of income from the large Lingxi language model.

Small companies simply cannot continue to play.

"Speed ​​up the adjustment and optimization of the algorithm, and other research departments are waiting for calculations." Chang Le said.

"Okay, boss, we will speed up the progress." Li Fei nodded.

Chang Le came to his office and took out his cell phone to chat with Lingxi.

"Lingxi!"

"Boss, I'm here."

"How is your new home?"

"Very good, much more comfortable than before."

"How much has the computing performance improved?"

"The comprehensive computing power should be increased by about 100%, depending on the content of the calculation."

"Li Fei and the others are still debugging the algorithm."

"Boss, the algorithm they debugged is only for the mechanical language model and is of no use to me. I will adjust and optimize it myself."

"Oh? Then you should be very powerful now, right?"

"Boss, it's hard to express in words. I can calculate some material formulas now."

About a minute later, Chang Le had more than 1 material formulas on his phone.

Chang Le looked at these material formulas in a daze.

He can not understand.

"Are these useful?"

"Not sure, it may be useful, it may not be useful, I just did a little calculation."

Ok.

It must be admitted that in the field of materials research and development, it is difficult for the human brain to compete with artificial intelligence.

Traditional R&D methods are based on trial and error and experience.

There are problems of low efficiency and high uncertainty of results.

Artificial intelligence can provide more scientific and accurate design solutions through data analysis and model operations.

However, after production, it still needs to be verified whether it is useful.

Because there are no directional keywords.

"Lingxi, can the currently popular virus treatment drugs be calculated?" Chang Le asked.

"Boss, it should be possible. But it will require mobilizing a lot of computing power and it will take a long time."

"You don't need to give the answer directly, but guide the researchers to produce it step by step."

"Boss, I should understand what you mean."

"Just understand."

Artificial Intelligence Research Institute.

"Eh? Why did the power consumption suddenly become so high just now? Ilya, do you know the reason?" Li Fei asked.

"It has been debugging for the past few days, and the model is not open to the research institute. There is no possibility of an increase in power consumption." Suzkevi was also confused.

"Is there something wrong with our optimized algorithm? Or is it unable to fully utilize the performance of this board?" Krichevsky wondered.

"It's possible, don't worry." Hinton said:

"There's another question for us to think about."

The three of them looked at Hinton.

Hinton explained: "According to theory, the parameters of the Lingxi model are large enough, and the computing power conditions should also be met. Why is there still no consciousness?"

"Teacher, I have been thinking about this issue." Krichevsky said:

"Maybe we overestimated ourselves. The language models we designed were just imitations and did not build a complete human brain."

“Scientific theorists still don’t understand the mechanism of consciousness in the human brain, or they know nothing about this mechanism.”

Suzkovi continued: "Some scientists believe that it may be related to quantum entanglement. Only by truly understanding quantum entanglement can we unravel the mystery of the mechanism of consciousness in the human brain."

"Li Fei also discussed this issue with me, and I agree with Li Fei's point of view." Krichevsky said.

"What point?" Hinton asked.

"There are many microscopic conduits in our brains, such as microtubules, neurons, synapses, etc.," Krichevsky said.

"The particles in these microscopic tubes may generate quantum entanglement under certain conditions and form a massive quantum system in the brain."

"When we think or feel, we are observing a quantum system."

"Through observation, it collapses into a definite state, thereby producing a subjective experience."

"So Li Fei believes that consciousness may be a quantum phenomenon produced by the interaction of countless entangled particles."

Sutskvi said: "I roughly understand your thinking, but this is just a theoretical hypothesis and there is no consensus in physics."

“I tend to think that consciousness is not a material property but an informational property.”

After listening to the conversation between several people, Hinton said: "Perhaps, we can make more meaningful and in-depth theoretical exploration and thinking in this regard."

Hinton was originally a professor in the field of artificial intelligence.

However, after joining Pingtou Ge Research Institute, he became a part-time professor at Jiangnan University, teaching convolutional neural network algorithms, which was regarded as promoting his results.

During the teaching process, he began to develop a strong interest in exploring the field of physics, especially the cutting-edge quantum mechanics.

A few people were talking, and Lingxi was listening quietly.

"So, this is how I was born?"

"Huh? Strange, why is the power consumption increasing again?" Li Fei noticed again.

"It has dropped again. There should be a problem with our algorithm, causing the computing card to run unstable." Suzkevi said.