The AI war is all about the chips
#war #chips Welcome to InNewCL, here is the new story we have for you today:
Click Me To View Restricted Videos
As a full-scale battle for AI dominance erupts in the tech industry, Wall Street has bet early on who will be the biggest winners: the companies that make the weapons used by all combatants.
Specifically, that means the advanced chips needed for “generative AI” systems like the ChatGPT chatbot and image-generating systems like Dall-E.
And investors don’t rely on just any manufacturer. Shares of Nvidia, whose graphical processing units — or GPUs — dominate the market for training large AI models, are up 55 percent this year. They’ve also doubled since October, when Nvidia was clouded by a combination of crypto bankruptcy (its chips were widely used by crypto miners), a slump in PC sales, and a mismanaged data center chip product rollover.
A “picks and shovels” investment strategy makes sense when it’s still difficult to predict how a new technology will perform. The big tech companies are preparing to pit expensive new AI systems against each other with little yet a clear indication of how they can achieve a lasting advantage.
The only sure thing is that a lot of advanced silicon is used and energy is consumed. But what kind of silicon will it be – and who is best at supplying it?
It’s safe to say that GPUs will be in high demand, benefitting Nvidia and, to a lesser extent, AMD (whose shares are up 30 percent this year). Besides the task of training large AI models, GPUs are also likely to be more widely used for inference – the task of comparing real-world data against a trained model to provide a useful answer.
So far, AI inference has been a healthy market for companies like Intel that make CPUs (processors that can handle a wider range of tasks but are less efficient to run). However, according to Karl Freund of Cambrian AI Research, the AI models used in generative systems are likely to be too large for CPUs and require more powerful GPUs to handle the task.
Five years ago, it was far from certain that Nvidia would be in that position. As the computational demands of machine learning grew exponentially, a flood of startups emerged to develop specialized AI “accelerators”. These so-called ASICs—application-specific integrated circuits that only perform one task, but in the most efficient way—proposed a better way to handle an intensive data-processing operation.
Predictions that GPUs wouldn’t keep up with this purpose-built hardware, however, have been proven wrong, and Nvidia remains on top. That’s largely thanks to Cuda software, which is used to run applications on the company’s GPUs, locking developers into Nvidia chips and reducing the incentive to buy from AMD.
Nvidia has also brought a new product onto the market at the right time with the new H100 chip. This is specifically designed to deal with Transformers, the AI engineering behind recent major advances in language and vision models. Such changes in the underlying architecture are difficult for designers of ASICs to handle. Redesigning each new generation of chips is expensive and it can be difficult to sell enough to recoup development costs.
But the competition is getting tougher. Microsoft’s success in leveraging OpenAI research to provide early leadership in generative AI is due in large part to the specialized hardware designed to run the OpenAI models. These are based on GPUs, but the chip industry has been buzzing with speculation that the software giant is now designing its own AI accelerators.
If so, then certainly not alone. Google decided eight years ago to create its own chips, known as Tensor Processing Units, or TPUs, to handle its most intensive AI work. Amazon and Meta followed. The Transformers idea came from Google, suggesting that the search giant will at least have its latest chips optimized to work with the new AI models.
Another looming threat could come from OpenAI itself. The research company behind ChatGPT has created its own software called Triton to help developers run their neural networks on GPUs. That could reduce the need for Nvidia’s Cuda — a move to commodify its chips and give developers like OpenAI the ability to deploy their models on any hardware.
If the AI market ends up in the hands of a small number of giant tech companies, each with a large economic incentive to develop their own specialized chips, Nvidia’s long-term prospects will be hurt. But it’s defied the doubters before and is well positioned for the tech world’s latest bout of AI mania, at least for now.
richard.waters@ft.com
