AI can democratize one of technology’s most important resources

Nvidia He is the undisputed king of artificial intelligence chips. But thanks to the AI ​​he helped build, the hero may soon face increased competition.

Modern AI is based on Nvidia’s designs, a dynamic that has propelled the company to a market value of more than $4 trillion. Each new generation of Nvidia chips allows companies to train more powerful AI models using hundreds or thousands of processors linked together inside massive data centers. One of the reasons for Nvidia’s success is that it provides software to help program each new generation of chips. This may not be a special skill soon.

A startup company called chip It is training AI models to do one of the most difficult and important jobs in AI: optimizing code so that it runs as efficiently as possible on a given silicon chip.

Emilio Andere, co-founder and CEO of Wafer, says the company implements reinforcement learning on open source models to teach them to write kernel code, or software that interacts directly with the hardware in the operating system. Weaver is also adding “utilities” to existing programming models such as Anthropic’s Claude and OpenAI’s GPT to enhance its ability to write code that runs directly on chips, Andere says.

Many prominent technology companies now have their own chips. Apple and others have used custom silicon for years to improve the performance and efficiency of software that runs on laptops, tablets and smartphones. At the other end of the scale, companies like Google and Amazon are minting their own silicon to improve the performance of their cloud computing platforms. Recently dead He said It will deploy 1 gigawatt of computing capacity with a new chip developed with Broadcom. Deploying custom silicon also involves writing a lot of code so that it runs smoothly and efficiently on the new processor.

Wafer works with companies including AMD and Amazon to help optimize software to run efficiently on its devices. The startup has so far raised $4 million in seed funding from Google’s Jeff Dean, OpenAI’s Wojciech Zaremba, and others.

Andere believes his company’s AI-led approach has the potential to challenge Nvidia’s dominance. A number of high-end chips now offer performance comparable to raw floating-point — a key industry standard for a chip’s ability to perform simple arithmetic operations — of Nvidia’s best silicon.

“The best AMD hardware, the best Trainium hardware (from Amazon), the best TPUs from Google, give you the same theoretical failures as Nvidia GPUs,” Andere told me recently. “We want to maximize intelligence per watt.”

Performance engineers with the skills needed to optimize code to run reliably and efficiently on these chips are expensive and in high demand, Andere says, while Nvidia’s software ecosystem makes it easy to write and maintain code for its chips. This makes it difficult for even the largest technology companies to do it alone.

When Anthropic partnered with Amazon to build its AI models on Trainium, for example, it had to rewrite its model code from scratch to make it work as efficiently as possible on hardware, Andere says.

Of course, Anthropic’s Claude is now one of many AI models that are now superhuman at writing code. So Ander thinks it may not be long before AI starts eating into Nvidia’s software advantage.

“The moat is in the programmability of the chip,” Andere says, referring to the libraries and software tools that make it easier to optimize code for Nvidia hardware. “I think it’s time to start rethinking whether this is really a strong moat.”

Besides making it easier to optimize code for different types of silicon, AI may soon make it easier to design the chips themselves. Recursive intelligenceA startup founded by two former Google engineers, Azalea Mirhosseini and Anna Goldie, is developing new ways to design computer chips using artificial intelligence. If the technology succeeds, more companies could branch out into chip design, creating custom silicon that runs their software more efficiently.