Chinese semiconductor thread II
- Thread starter vincent
- Start date
Not entirely sure, forget now, a hundred years ago, Henry Ford had his company source everything in house. Maybe they even owned the coal mines that went into steel productions for Ford cars.
Back in those days, it was kind of normal for Ford to do everything in house.
We have returned to those days in manufacturing. Want to be a manufacturer and compete in China? Huawei and BYD showed this is how it will be done. Vertical integration is a must. If you want to be the big time.
Also, with this vertical integration, and how AI will help lower cost across the entire organization, we can see where this is going. The vertical integration makes too much sense with AI integrating it along. That also means this is a barrier to entry, of how manufacturing is going to be done. Restricting competition.
This AI thingy, will produce winners and losers in manufacturing. Seems that would can vertically integrate more will win.
That put some places in a tough spot. For instance, the Germans cannot even get a handle on their energy costs. Thinking how to vertically integrate then use AI to reduce the cost, is not what they are thinking about. They have to survive first. Many companies will not make it.
This tech war started by the Americans, really is like Nazi Germany invading Russia. Right now, seems like China has reached and equilibrium, where everyone is focused, and moving forward.
No one can beat Huawei and BYD because of their operations being vertically integrated to such a degree. Mix in the AI, then we see the writing on the wall.
A lot of important stuff is posted in this thread. This is a great idea.
However, I would like to translate, because I feel many posters do not fully understand, therefore they do not appreciate it as much as they should.
Instead of laughing, some wonder what is this. So I would like translate a bit.
There was the CPU the central processing unit, the semiconductor chip for the computer, the brains that adds and subtracts, then stores the results into memory.
After the CPU, some people wanted to play games, so Jensen Huang delivered the GPU the graphics processing unit. This chip was not to replace the CPU chip, it was to help it, so we could play games, back in the day.
Then, as technology evolved, people realized that the GPU was great at training the LLM or the AI models.
That is where we are today.
Now this post by comrade tokenanaylst, shows a new design for a new chip, specifically designed to handle the tokens from Large Language Models.
If we had CPU, the GPU, then this maybe a LPU, whatever, you get the picture.
The GPU was to handle the graphics better so we can play games. Well, maybe not me, but I knew people who played a lot of computer games, they still do too.
What seems curious, as I really do not know much about this stuff even though I am trying to translate it, is that the LLM uses a lot of matrix calculations.
Look at that design of this chip. They design like matrix, to basically handle those LLM matrix calculations.
Whatever goes through the LLM calculation, what comes out is not created equal, so you have those weights. So with a more effecient way of calculating or processing those weights, the more effecient or better your model, aka your LLM, will be.
So this design of this chip, is to improve preformance of the LLM, and it could improve it radically. Maybe we can run even more advanced LLM on the simple cell phone. Like on the cell phone, with this chip. Eventually, that is what someone will want and drive it there.
That is why it is imperative that Jensen Huang and Nvidia has access to the China market. That way he knows what is going on. If someone is designing a chip like this, and you're not, then guess what? Now you're IBM or something.
Some easy reading. Blah blah blah.
Agree, but there are some issues, modern GPUs are general purpose use, that is why despite being made for graphics can also be used for training and running AI models. A LPU would in theory be WAY more efficient than a GPU but the question is: They will be flexible and general purpose enough to accommodate different LLM architectures? Or the entire AI landscape has to be accommodated to LPUs?A lot of important stuff is posted in this thread. This is a great idea.
However, I would like to translate, because I feel many posters do not fully understand, therefore they do not appreciate it as much as they should.
Instead of laughing, some wonder what is this. So I would like translate a bit.
There was the CPU the central processing unit, the semiconductor chip for the computer, the brains that adds and subtracts, then stores the results into memory.
After the CPU, some people wanted to play games, so Jensen Huang delivered the GPU the graphics processing unit. This chip was not to replace the CPU chip, it was to help it, so we could play games, back in the day.
Then, as technology evolved, people realized that the GPU was great at training the LLM or the AI models.
That is where we are today.
Now this post by comrade tokenanaylst, shows a new design for a new chip, specifically designed to handle the tokens from Large Language Models.
If we had CPU, the GPU, then this maybe a LPU, whatever, you get the picture.
The GPU was to handle the graphics better so we can play games. Well, maybe not me, but I knew people who played a lot of computer games, they still do too.
What seems curious, as I really do not know much about this stuff even though I am trying to translate it, is that the LLM uses a lot of matrix calculations.
Look at that design of this chip. They design like matrix, to basically handle those LLM matrix calculations.
Whatever goes through the LLM calculation, what comes out is not created equal, so you have those weights. So with a more effecient way of calculating or processing those weights, the more effecient or better your model, aka your LLM, will be.
So this design of this chip, is to improve preformance of the LLM, and it could improve it radically. Maybe we can run even more advanced LLM on the simple cell phone. Like on the cell phone, with this chip. Eventually, that is what someone will want and drive it there.
That is why it is imperative that Jensen Huang and Nvidia has access to the China market. That way he knows what is going on. If someone is designing a chip like this, and you're not, then guess what? Now you're IBM or something.
Some easy reading. Blah blah blah.
Even for newer DRAM which needs EUV, SK Hynix wanted to upgrade the Chinese fab as much as possible and move the wafers to Korea just for the EUV layer and then bring them back. This news is from last year, so they may already be doing that.
Decoupling is better for everybody it seems
Will this "refrain" Trump from "asking for share ownership" as well?