Qualcomm Repurposes Phone Chips to Challenge Nvidia in AI
ā¼ Summary
– Qualcomm is launching AI200 and AI250 chips in 2024 and 2027 to challenge Nvidia’s market dominance.
– The new chips are designed for deploying AI models rather than training them, marking a shift from Qualcomm’s traditional mobile and telecom focus.
– These processors are based on Hexagon neural processing units used in mobile devices and can function in racks of up to 72 chips as a single computer.
– The AI200 features 768GB of RAM optimized for AI inference, while the AI250 promises significantly improved power efficiency.
– Humain, backed by Saudi Arabia’s Public Investment Fund, plans to use both chips for AI datacenters in Saudi Arabia.
Qualcomm is strategically entering the artificial intelligence chip market by launching two new processors designed to compete directly with industry leader Nvidia. The company has announced the upcoming release of its AI200 chip, scheduled for next year, followed by the more advanced AI250 model in 2027. Both processors leverage Qualcommās established mobile neural processing architecture, marking a significant pivot for a firm traditionally known for supplying components to smartphones, laptops, tablets, and telecom infrastructure.
These new chips are engineered specifically for AI inference tasks, which involve running already-trained AI models, rather than the computationally intensive training phase. This focus allows Qualcomm to target a different segment of the AI workflow. The underlying technology is derived from the company’s Hexagon neural processing units, the same hardware that accelerates AI functionalities in its popular mobile and laptop processors.
A key feature of these AI processors is their scalability. They are designed to be deployed in large-scale computing racks, where up to 72 individual chips can be integrated to function as a unified system. This approach mirrors the deployment strategies used for high-performance GPUs from competitors like Nvidia and AMD.
The AI200 model is reported to include a substantial 768GB of RAM, with its performance finely tuned for efficient AI inference operations. Its successor, the AI250, is promised to deliver a generational leap in power efficiency, potentially resulting in significantly lower energy consumption for data centers. The commercial potential of these chips is already being realized. Humain, an AI firm operating under Saudi Arabia’s Public Investment Fund, has confirmed it will utilize both the AI200 and AI250 chips. This partnership is part of a broader initiative to develop and power new AI data centers throughout Saudi Arabia.
(Source: The Verge)
