It seems to be a new trend for AI chip companies to directly provide cloud services.
Market leader Nvidia announced at GTC 2023 that it will provide AI cloud services based on Oracle’s cloud infrastructure.
American AI chip unicorn SambaNova is also doing the same thing, providing AI computing power in the form of cloud services.
Graphcore, a British AI chip unicorn company, has also recently adjusted its strategy. Its CEO Nigel Toon said that it will work closely with cloud vendors in the UK and the US to provide computing power in the form of cloud products instead of selling chip products separately.
However, Graphcore China did not follow up the British headquarters and did not sell chip products separately.product strategy.
“The strategies of China and the headquarters will be different in the future. We will continue to be consistent in business and strategy, and continue to provide AI computing power in the form of selling hardware and installing products ourselves.” Lu Tao, president of Graphcore and general manager of Greater China, said,
“In China, we still firmly follow the strategy of being integrated by cloud vendors.Graphcore China still has such a business model. “
Why not directly provide cloud services in China?
The core of the different strategies between Graphcore headquarters and the China region lies in whether to sell cloud services or directly sell chips. What makes people curious is why the China region has made a different choice?
Lu Tao explained,China’s industrial format is very different from that of the United States and Europe.For example, almost all Internet companies in the United States are built on the three clouds of AWS, Azure and Google Cloud. There are many excellent cloud vendors in China, but there are also many large Internet companies that choose to build their own data centers. Some companies were originally big customers of cloud vendors, but they also did a lot of self-construction as their size increased.
“The self-built behavior of these large-scale users, especially the top users, makes us find that they prefer to be provided with services in the form of hardware.The cloud is also very important, but we hope that Graphcore can become a part of the products of Chinese cloud manufacturers, rather than being a cloud of Graphcore’s own brand. “Lu Tao explained the reason why China insisted on the original business model.
In terms of customer groups, Graphcore China will focus on large Internet companies.
This mainly has two meanings. First, large Internet companies use GPU products the most deeply, thoroughly and widely, and have strict requirements on the technical parameters, usability, technical support services, commercial support services, and product stability of non-GPU products. high.
“Only by working with demanding customers can we really polish our products. The process is very painful, but we can learn a lot from it.” Lu Tao has a deep understanding, “We have deployed a comparison for a large domestic customer before. For the large IPU cluster, they put forward more than 300 software feature requirements to us. If these requirements were not raised by customers, we would not be able to imagine many requirements.The higher and more demanding the customer’s requirements, the faster the progress.”
The significance of the second aspect is from a commercial point of view. Taking China as a single overall market, China’s Internet industry accounts for 60% of the entire AI application market, and the top 10-20 companies in this industry account for more than 80% of the market.
“From a commercial point of view, if your product performs satisfactorily, the final commercial return will be relatively reasonable.“Lu Tao said.
According to Leifeng.com, Graphcore is cooperating with domestic Kingsoft Cloud and Tencent Cloud, and the details of cooperation with another large domestic cloud vendor are under planning.
How to seize the opportunities brought by generative AI?
Today, the large models of major Internet companies have also brought strong demand for AI chips with large computing power. Only by quickly implementing support for large models can it be possible to seize market opportunities.
Currently, the large models supported by Graphcore include GPT-2 XL version, GPT-J open source version, Dolly 2.0 (a model with more than 10 billion parameters that has just been open sourced), and ChatGLM-6B (a relatively popular open source model in China).
“For companies with large models, one of our advantages is that we can achieve 3.5 milliseconds per token, Hundreds of words can come out in a second. “Lu Tao used a more vivid way to reflect this speed,”ChatGPT is presented verbatim and we can display screen by screen.This is an important value point for those who make large models, which can bring a different experience.
After realizing support for multiple large models, it also means that Graphcore can quickly support other large models.
Lu Tao said, “It took about two weeks from the release of Dolly 2.0 to when we supported this model.The key to supporting large models is the size of the model and how to split the model between multiple IPUs.Various underlying capabilities are required, such as Transformer-related operators, model parallel APIs, operator parallel APIs, and Tensor parallel APIs. These underlying technologies are relatively mature. “
certainly,If you want to attract customers who are familiar with GPUs to use Graphcore’s IPU, it is best to achieve zero code migration.Compatible with CUDA.
When Graphcore launched the C600 product that focuses on reasoning and training last year, it released the software toolkit PopRT at the same time. This toolkit can automatically convert the model trained by GPU in TensorFlow. It is relatively seamless without CUDA compatibility. Run the GPU software on the IPU in a timely manner.
“We just released version 1.0 last month. At present, some customers’ feedback is not bad, because it took a long time to do some model migration in the past. Now that we have tools, the customer’s own model migration is still relatively fast.” Lu Tao said, “Of course, the advantage of our chip IPU is that there are many cores, and the storage and bandwidth are very large.”
For the current numerous AI start-up chip companies with large computing power, Nvidia does not leave too many opportunities for competitors, and differentiated competition can increase the possibility of success.
This differentiation includes adopting more suitable market strategies for different markets. Therefore, Graphcore adopts different strategies in the UK and Chinese markets, which is a manifestation of differentiation and flexibility. This flexibility will help Graphcore achieve success in China. success.Leifeng.com(Public number: Leifeng.com)
Leifeng.com original articles, unauthorized reprinting is prohibited. See Reprint Instructions for details.