Source: Wall Street News Author: Chang Jiashuai
Compared to Nvidia, building a competitive chip is already a big challenge, let alone building an entire software ecosystem and getting people to start using it.
In the AI chip market, Nvidia has a nearly monopoly position due to its strong GPU performance and CUDA software ecosystem, but supply shortages and high prices are driving its customers to seek alternatives.
Major customers and competitors join forces to participate in open source projects to replace CUDA
Nvidia's big customers, including OpenAI, Meta, Microsoft, Google, and Amazon, as well as Nvidia's competitors such as Intel, AMD, and Qualcomm, seem to have formed a delicate consensus to overthrow Cuda's hegemony. These companies are all participating in Triton, an open source language project initiated by OpenAI.
Tech giants like Meta, Microsoft, and Google that spent tens of billions of dollars to buy Nvidia chips hope that Triton will help them break Nvidia's monopoly on AI hardware; Intel, AMD, and Qualcomm also hope to use Triton to target Nvidia's customers.
Triton was first released by OpenAI in 2021 to enable code to run smoothly on all types of GPUs.
According to OpenAI's official explanation, Triton's goal is to provide an open source environment to write fast code with higher productivity than CUDA, while being more flexible than other existing DSLs.
Analysts believe that the reason why Nvidia has been able to become a “seller” in the AI industry over the past year is mainly due to its CUDA system, which has been developed for more than 20 years. This is an insurmountable barrier for competitors.
Nvidia CEO Hwang In-hoon once said that his company “not only produces chips, but also builds entire supercomputers, from chips to systems to interconnects... but most importantly software.”
He called CUDA the “operating system” for AI.
Challenging CUDA can be very difficult
Since CUDA was born in 2006, Nvidia has invested billions of dollars to develop hundreds of software tools and services to speed up and simplify running AI applications on its GPUs. In terms of the number of employees, Nvidia has twice as many software engineers as hardware engineers.
David Katz, partner at AI investment firm Radical Ventures, said:
I think people are underestimating what Nvidia actually built. They've built an efficient, easy-to-use, and actually workable software ecosystem around their products, making complex things simple.
Despite this, the high price of Nvidia products and long waiting queues to buy the most advanced devices prompted its major customers to seek alternatives or develop their own GPUs.
However, since most AI systems and applications run in Nvidia's CUDA ecosystem, rewriting code for other GPUs (such as AMD's MI300, Intel's Gaudi 3, or Amazon's Trainium) takes a lot of time and risk.
Gennady Peshmenko, CEO of AI startup CentML and associate professor of computer science at the University of Toronto, told the media:
To compete with Nvidia in this field, you need not only build competitive hardware, but also be easy enough to use. The performance of the Nvidia chip is indeed excellent, but in my opinion, its biggest advantage is in terms of software.
Compared to Nvidia, building a competitive chip is already a big challenge, let alone building an entire software ecosystem and getting people to start using it.
Although Triton may weaken Nvidia's market share, Citigroup analysts estimate that by 2030, Nvidia's share of the generative AI chip market will still be as high as 63%, which means it will remain dominant for many years to come.
edit/lambor