The 2024 edition of the Computex – a Taipei-based electronics show where many Asian companies are usually showcasing their latest motherboards, PC power supplies or laptops – turned into an AI festival last week with a flurry of non-Asian companies taking the stage to pitch their roadmaps, new chips and systems not only for AI training (performed at the data center level) but also for inference (the roll-out of trained AI models on devices). This emphasis on inference coincided with the reveal yesterday by Apple of its iOS Intelligence features, suggesting that AI inference is about to come to billions of devices in a near future
In Taiwan, Nvidia, AI’s poster child, announced a new server-class GPU architecture called Rubin which will extensively use HBM4 memory and is expected to be released in 2026. The company reiterated its plan to double the cadence of new chip release (one every year) to keep competition at bay, even if it remains to be seen whether customers will be able/willing to follow such a fast upgrade cycle…
Archrival AMD showcased its next generation server-class GPU Instinct, expected to be released at the end of the year to compete head-to-head with the dominant H200 solution from Nvidia. On the inference side, the company announced a new generation (Zen 5 architecture) of laptop/desktop CPUs, optimized to locally run GenAI models and to deliver top-notch performances for Microsoft’s AI assistant Copilot.
Intel also took part in the AI show by introducing the next iteration of its server-class CPU chip called Xeon and the upcoming Lunar Lake laptop processor family which, without surprise, are all optimized to efficiently process AI models. The company also revealed the Gaudi 3 chip, a solution that should help Intel somewhat catch up in the AI server race where the company has a close to zero market share. Many assemblers plan to offer Gaudi 3-based systems with an expected price 50% below its competitors’ offering for the same computing power.
Lastly, Qualcomm was quite vocal about its AI strategy in the smartphone segment and… in the PC one as well!
According to the company’s benchmarks, its Snapdragon X Series chip, based on the ARM microprocessor architecture and designed for the PC market, is leaving AMD/Intel duopoly’s latest solutions in the dust.
AI training is, for the moment, largely dominated by Nvidia (more than 85% market share), which will soon face more intense pressure from its historic competitors but also from its major customers (Google, Microsoft…) which are developing in-house solutions based on the ARM architecture. But as we said previously, any Nvidia market share decline should not be a major source of concern in the current environment given that this should be largely offset by a much stronger than expected market growth.
On the inference side, the race remains fully open!
The introduction of AI features in the Apple ecosystem will obviously give a major boost to the roll-out and usage of AI apps in the next couple of years and we can then expect most smartphones, tablets, PCs, cars or other connected devices to come equipped at some point with inference chips to process AI tasks on-device.
The market for these inference chips and AI-enabled devices is massive. If we take the example of the iPhone, only 5% of the installed base has the sufficient processing to handle AI, suggesting that 95% of the devices on the market need to be upgraded.
This refresh cycle will be powerful for most consumer electronics industries but should not really start before 2025/26 in our view. Indeed, Apple Intelligence globally failed to impress yesterday with an enhanced Siri (integrated with ChatGPT), the standout, but no killer app. And the tech will only be available to English-speaking users in the Fall.
One sure thing is that this highly fragmented market (between smartphones, PCs, cars…) opens the opportunity for many semiconductor players to penetrate segments where they had no or no significant positions.
For example, the computing cartel formed by Nvidia, Intel and AMD is inexistent in the smartphone industry, currently dominated by Qualcomm and Taiwanese Mediatek. A supposed collaboration between Nvidia and the Taiwanese chip designer could well soon change this situation (with another ARM-based chip for the PC market).
After many attempts that spanned over the last 15 years, it seems that ARM-based solutions will finally find their way into our personal computers, a trend that was set many years ago when Apple decided to standardize its full product line on ARM chips to get rid of Intel. ARM’s CEO, which also gave a keynote at the Computex, expects to grab 50% of the Windows PC market by 2029 (it currently stands at 10%).
Whatever happens in the market share game across all these segments is yet quite unknown, but one thing stands out: GenAI is in the early stages of boosting the global semiconductor industry which grew 15% year-on-year in April and is helping TSMC print revenue growth close to 30%.
