The west wind comes from the Wafei Temple
量子位 | 公众号 QbitAI
Nvidia, which has a market value of $2.3 trillion and is killing all sides in the AI acceleration market, the whole industry can't bear it.
Google, Intel, Qualcomm and others jointly established the "anti-CUDA alliance" - the UXL Foundation, to start with software that complements Nvidia's chips and break its monopoly.
More than 4 million developers around the world rely on NVIDIA's CUDA software platform to build AI and other applications.
The computer code accumulated in the CUDA ecosystem for nearly 20 years is almost as important as Nvidia hardware.
In the latest news, Vinesh Sukumar, head of artificial intelligence and machine learning at Qualcomm, said in an interview with Reuters:
We're actually showing developers how to migrate away from the NVIDIA platform.
In addition to the UXL Foundation, other giants such as AMD have made similar moves.
According to an insider, Microsoft is trying to combine AMD's MI300 chip with its software ecosystem ROCm, and has successfully run GPT-4 on the AMD platform:
As for Nvidia, Ian Buck, its vice president and chief director of accelerated computing, said in a statement:
The world is accelerating, and new ideas for accelerated computing from across the ecosystem will help push the boundaries of AI and accelerated computing applications.
Break the CUDA ecological moat
UXL基金会全称Unified Acceleration Foundation，创始成员包括Arm Holdings plc、富士通、谷歌、Imagination Technologies、英特尔公司、高通公司和三星。
Reuters recently broke the news that since Intel developed oneAPI, UXL has planned to build an open-source project to create a set of software and tools that can drive a variety of AI accelerator chips, so that computer code can run on any machine, regardless of the chip and hardware used.
According to an executive, the UXL Technical Steering Committee is preparing to finalize the technical specifications in the first half of this year, and engineers plan to refine the technical details to a "mature" state by the end of the year.
In addition to initially participating companies such as Google, Intel, Qualcomm, etc., UXL will also attract cloud computing companies such as Amazon and Microsoft Azure, as well as more chipmakers. UXL also accepts third-party technical contributions, such as from outsiders who are passionate about open source technologies.
Ultimately, their goal is to support NVIDIA's hardware and code.
It's worth mentioning that UXL is just one part of the anti-CUDA army.
According to Reuters, PitchBook data shows that venture capitalists and corporate funds have invested $4 billion in 93 different related projects. Startups in the name of undermining Nvidia's leadership received more than $2 billion in investment in 2023, compared with $580 million the year before.
NVIDIA's long-term layout around CUDA
But does this kind of "encirclement and suppression" really work?
Jay Goldberg, CEO of D2D Advisory, believes that CUDA's features aren't really important yet:
The important thing is that people have been building code around CUDA for 15 years.
After CUDA was first introduced in 2006, NVIDIA has carried out a long-term layout around CUDA, using the CUDA software ecosystem to help GPU hardware, create a software and hardware ecosystem, and form industry barriers.
In the early days, Nvidia adopted a free promotion strategy and did not seek short-term returns.
For example, giving free hardware to startups, including OpenAI.
Last month, Lao Ma, who participated in the founding of OpenAI, also posted a photo of Huang presenting OpenAI with the world's first DGX-1 supercomputer in 2016.
Nvidia scientist Jim Fan also opened the microphone and said that he witnessed that moment with his own eyes:
In addition, NVIDIA has also launched the "Startup Acceleration Program" to provide startups with various support such as software and hardware discounts, and when these companies grow, the CUDA code programmed for GPUs will become the industry standard.
As CUDA has expanded and startups have grown, CUDA has become deeply embedded in many AI and high-performance computing projects.
Now NVIDIA's GPU is stepping on the new era, holding the lifeblood of OpenAI, Google and other AIGC developers. CUDA, the moat of the software ecosystem, has also become more solid, and at the same time, the policy is also being tightened.
Some time ago, an engineer discovered that when installing CUDA 11.6, a new clause was added to the user license agreement:
You may not reverse engineer, decompile, or disassemble any output generated using SDK elements for purposes that are intended for non-NVIDIA platforms.
In fact, in previous versions, there was a provision prohibiting CUDA from running through the translation layer on other hardware platforms.
Coming back to the trend of UXL this time, netizens are also waiting to see the lively state, "Which is the standard?":
What do you think about this?
Reference Links:
[1]https://www.reuters.com/technology/behind-plot-break-nvidias-grip-ai-by-targeting-software-2024-03-25/
[2]https://x.com/denizerkan_nyc/status/1772461965954154830?s=20
— END —
QbitAI · Headline number signed
Follow us and be the first to know about cutting-edge technology trends