In the recent wave of artificial intelligence, NVIDIA is undoubtedly a company on the cusp of the limelight, and its limelight has even recently overwhelmed ChatGPT's owner Microsoft.
A key reason why NVIDIA can lead the way is that its widely sought after chip products in the field of artificial intelligence, namely A100 chips and higher-generation H100 chips, these high-end chips and corresponding graphics cards are difficult to find.
Zhang Yi, a senior researcher at Microsoft Asia researcher, recently lamented in a podcast that now there is a strange scene where the entire earth cannot get enough A100 chips. A year ago, almost no one expected this.
The A100 chip launched by NVIDIA in 2020 is now priceless, and the H100, which is popular with ChatGPT, is being snapped up by large companies. This also made NVIDIA's performance soar, and the stock price was rising.
Brannin McBee, founder and CEO of Core Weave, a startup in the field of artificial intelligence, can't help but lament: H100 is one of the most scarce engineering resources on the planet. This statement is enough to give people a glimpse of NVIDIA's current grand scene.
But there are thousands of chips in the world, why only NVIDIA's chips have become unique players in the field of artificial intelligence? And NVIDIA, a company that has always dominated graphics cards, why can it make such a big fortune in the field of deep learning and artificial intelligence?
Two pushes from Microsoft
In 1999, the budding NVIDIA first introduced the concept of GPU. Prior to this, CPU manufacturers, including Intel, firmly believed that graphics processing was the job of the CPU, and the more things the CPU did, the better, and the idea of separating the graphics work on another subsidiary processor was very chicken.
At that time, Japanese manufacturers engaged in games had the most power to speak in the field of graphics applications. The CPU of the Japanese host is very strong, and most of the development work is concentrated on the CPU, so the GPU does not get much market space.
The turning point is that Microsoft, which is not convinced, wants to attack the industry leading position belonging to Japanese manufacturers, and it developed Direct X, a standardized API graphics interface, and since then a large number of graphics functions have been ported from the CPU and transferred to the GPU. Coupled with the launch of Microsoft's other product Xbox, its CPU and GPU have their own responsibilities, breaking the situation that CPU chips in the industry are the only one.
And NVIDIA was the only company in the hardware field to follow the Microsoft banner and went all the way to the black on the road of GPU.
Since then, Microsoft has pushed for another change, introducing unified rendering technology, that is, allowing the GPU to combine the vertex calculation of graphics drawing and the subsequent rendering. It has partnered with ATI, another well-known company in the graphics card field, GPU Xenos, and successfully applied this technology.
Unintentional willow insertion
Unified rendering is only a step in graphics applications, but it has brought a completely different development path to NVIDIA, which can be said to be the starting point of NVIDIA's later GPU development and even intervention in the field of deep learning.
After seeing the unified rendering architecture, NVIDIA decisively pushed its previous GPU architecture back and restarted. Its GPU stream processors have been meticulously grouped into small stream processors that can run separately, solving the problem that stream processors were previously bound and could not run independently and were forced to idle.
This laid the foundation for NVIDIA's later revolutionary CUDA architecture. Because NVIDIA's stream processor is a very independent and standard unit, it is easy to control and schedule, which allows tasks that would otherwise only be processed serially to be processed in parallel. This makes programming much less difficult.
At the same time, NVIDIA's competitor ATI did not invest in hardware architecture changes in the early days, because the past serial design was used, the sunk cost became higher and higher, making its innovation more and more difficult and expensive, and finally successfully squeezed out of the graphics card market by NVIDIA.
Since then, in 2017, NVIDIA introduced the Tensor Core computing unit concept, which is specially designed for deep learning and supports lower precision operations to greatly save model computing power.
This dedicated acceleration unit largely squeezed out CUDA's space for processing deep learning, but it also caught NVIDIA's competitors by surprise, making AI-specific chips no longer attractive. As a result, NVIDIA GPUs coincidentally became the most recognized hardware in the AI field.
Bet to win trend
In 2003, NVIDIA, which "iterates quickly and keeps trial and error", engaged in an unpopular project. It has developed an SoC chip that integrates a CPU based on the ARM architecture with its own GPU.
Since the SoC chip, NVIDIA has released chips every few years. In 2015, it launched the Tegra K1, which uses the Arm public version of the CPU and its own Kepler architecture GPU, but because of the unsatisfactory power consumption and heat, it is very torturous for most users.
But industry insiders acknowledge these setbacks. An investor once pointed out that while Nvidia is holding the basic GPU disk, it continues to extend its tentacles in new areas, and allows countless people who buy its graphics card to accompany it to share the cost.
He also praised that although many things of NVIDIA, such as CUDA, could not see the landing scene for a period of time, it established a complete ecology in the process of trial and error, and successfully stood on the cusp when a new wind hit.
This is also one reason why NVIDIA GPUs beat other chips and successfully ate the AI dividend. On the one hand, GPUs are more versatile and more adaptable to changes than dedicated chips; On the other hand, NVIDIA has a complete ecosystem, making its GPU the most suitable choice at the moment.
In fact, when AI breaks out in an instant, companies in the industry have no choice but to find that GPUs are the best choice for simple and efficient running generative AI models, and a GPU that was originally used to play games is unlikely to switch to run AI programs, and only NVIDIA's GPUs can run AI models.
And there is also a small Easter egg in the story of NVIDIA.
In 2016, NVIDIA released the DXG-1, the first deep learning supercomputer. Commendably, NVIDIA CEO Jensen Huang donated the first DXG-1 to OpenAI, which was a start-up at the time, as if he could foresee the future.
In 2022, OpenAI detonated the concept of artificial intelligence with the birth of ChatGPT, and also drove NVIDIA to become a sweet potato in the chip field. This fate has to be emotional, but it is like the good fruit of Huang Jenxun's vision.
This article is from CaiLian News