According to the latest report, Qualcomm is the world’s largest supplier of smartphone processors, and the chips are very perfect in terms of computing power and energy efficiency.
In 2019, Qualcomm announced that based on its technology and experience in the field of smartphone chips, it will enter the fast-growing market of data center artificial intelligence chips.
Moreover, Qualcomm had approached Facebook’s parent company Meta Platforms, hoping that Meta would become a benchmark customer for Qualcomm’s first data center AI chip, AI 100.
Join tip3x on Telegram
After Qualcomm released the chip in the fall of 2020, Meta tested the chip against a range of other options, including chips the company had been using before, as well as a dedicated chip Meta developed in-house for processing AI computing.
According to people familiar with the matter, Qualcomm’s chip performed well in the test, with the best performance per unit of energy consumption. For a company like Meta, since its data centers serve billions of users, improvements in energy efficiency can lead to huge optimizations in operating costs.
However, by the spring of 2021, Meta said it would refuse to use Qualcomm’s chips, people familiar with the matter said. The specific reason is that Meta doubts that the supporting software of Qualcomm chips is not mature enough to exert the best performance of the chips in specific computing tasks in the future. After evaluating various options, Meta decided to stick with existing chips, a person familiar with the matter said.
The incident has never been reported by the media before, but it also shows that software has become one of the core factors in winning customers for AI chips. Sales of AI chips are expected to reach $13.5 billion this year and grow to $41.3 billion by 2026, according to IDC analyst Shane Rau. “At least for the next 15 to 20 years, the market demand for AI chips is almost limitless,” he said.
Qualcomm chips are at the heart of billions of smartphones around the world, and they also underpin AI features such as smartphone camera optimization, but the AI 100 is the company’s first attempt at competing with Nvidia.
In the field of data center AI chips, Nvidia currently has an overwhelming advantage. The company’s dominance comes not just from the chips, but also from the accompanying software. Nvidia’s software is the current gold standard in the AI industry.
“It’s not just Qualcomm, everyone is in an arms race with Nvidia CEO Jen-Hsun Huang,” said Peter Barrett, general partner at venture capital firm Playground Global. His software efforts help maintain the company’s leadership.” The playground has also invested in companies such as MosaicML, which help AI clients match their models to the right hardware.
To be sure, Meta’s rejection is likely just a temporary setback for Qualcomm in the field of AI chips. Just in September 2021, after Meta’s test, the AI 100 chip achieved multiple firsts in the MLPerf basic test. The MLPerf benchmark is a set of industry standards for measuring the performance of AI chips.
Furthermore, industry watchers expect Qualcomm’s chips to perform well in tests again this spring. Qualcomm has announced the first customer for AI 100: Foxconn Industrial Internet. The company is using the chip in a server that analyzes video from security and traffic cameras.
Meanwhile, Qualcomm continues to woo other potential customers such as Microsoft. A Microsoft spokesman declined to comment on this dynamic. Qualcomm plans to use the AI 100 chip for inference computing, which uses AI models trained on massive data to make real-time decisions. In the context of Meta, this usually means deciding what content to show the user in milliseconds based on a recommendation model.
To achieve better performance, the trained model must also be optimized for the hardware on which the model is run. If the optimization is poor, the model is likely to use only a fraction of the hardware’s available performance, resulting in wasted power. However, the optimization of the model can consume a lot of developer time.
Often, software that can optimize code written in various languages and automatically match the underlying hardware is more likely to be favored by developers. Nvidia’s software excels here.
The CEO of Ceremorphic, a startup that develops AI processing, said that if the chip is given directly to developers, without optimization software, it is like giving the user a 100-speed car. A bike in the right position, and then counts on him to figure out how to ride on unknown terrain and in which gear.
“You can’t give developers 100 gears, you have to make the configuration look like three gears,” he said. “At the moment, most chip companies don’t do that.”
Engineers capable of writing the software that goes with the chip are scarce. That’s a challenge for a big company like Qualcomm, and dozens of other startups targeting the same market.
The development of such software requires developers to have specialized experience in compilers. The compiler translates the code written by the developer into the machine language used by the chip.
“This type of talent is sought after and very scarce,” said Shahin Farshichi, a partner at Lux Capital, which has invested in AI chip startups Mythic and Flex Logix.
