Challenging NVIDIA? Qualcomm Takes on AI Computing Power Chips Again

Qualcomm, a giant in smartphone processors, has officially joined the ranks challenging NVIDIA. Qualcomm recently announced that it will collaborate with Humain, a newly established AI company in Saudi Arabia, to jointly develop AI computing chips for AI data centers.

According to the memorandum of understanding, Qualcomm and Humain plan to develop and build cutting-edge AI data centers in Saudi Arabia. Based on Qualcomm's edge and data center solutions, they will provide local and international customers with efficient and scalable hybrid cloud-to-edge AI inference solutions.

The two companies will also collaborate with the Ministry of Communication and Information Technology (MCIT) of Saudi Arabia to establish a Qualcomm semiconductor technology design center in the country.

It is worth noting that Humain announced partnerships with rival chipmakers NVIDIA and AMD separately this week to use their chips in its AI data centers.

In fact, this is not Qualcomm's first attempt to enter the AI computing chip market. As early as 2019, Qualcomm launched its first data center chip, AI 100, aiming to leverage its low-power and high-efficiency technologies accumulated in the mobile chip sector to penetrate the AI inference computing market for data centers.

At that time, Qualcomm's primary customer for AI 100 was social media giant Meta. According to informed sources, although Meta found the performance of the AI 100 chip to be quite outstanding during testing, with particularly notable computing power per watt of power consumption, which was crucial for Meta's data center operating costs supporting billions of users,

however, Meta's concerns about the maturity of the supporting software provided by Qualcomm, particularly its ability to consistently deliver the chip's maximum performance in long-term tasks, ultimately led to the collapse of the collaboration.

The key to NVIDIA's significant lead in the AI chip sector lies in its robust software support. Its CUDA platform and a series of optimization tools provide developers with a comprehensive and user-friendly ecosystem, making the training and deployment of AI models more efficient and flexible. In contrast, Qualcomm's AI chip software platform, which is still in its early stages of development, clearly has a long way to go.

However, Meta's rejection does not mean the complete failure of Qualcomm's AI 100 chip. Foxconn Industrial Internet (FII) was the first public customer of AI 100, using it in servers for analyzing traffic and security surveillance video streams, demonstrating the potential of Qualcomm's chips in real-world applications.

The core of AI inference computing lies in using trained models to make real-time decisions, such as in recommendation systems, speech recognition, or image analysis. Inference tasks not only require high computational efficiency but also must operate stably under strict latency and energy consumption constraints. This scenario aligns perfectly with Qualcomm's strengths in the field of low-power mobile chips.

According to Gartner's report, the global AI chip market is expected to reach $119.4 billion by 2027. This not only signifies a vast market opportunity but also implies that it will be difficult to maintain a leading position for an extended period. Despite NVIDIA's strength, it must continually address challenges from both traditional giants and emerging players.