Image: Alpha Photo/Flickr
Meta has begun testing its first chip for training artificial intelligence systems. This will allow the company to develop its own semiconductors and reduce dependence on external suppliers such as Nvidia, two sources told Reuters.
- Meta has already released a small batch of chips for testing, and if the tests are successful, the company plans to increase production.
- The new chip is a specialized accelerator, meaning it is designed exclusively for AI tasks. This allows it to be more power efficient compared to integrated graphics processing units (GPUs), which are commonly used for AI computing.
- The chip is being manufactured in collaboration with TSMC of Taiwan.
- The test deployment began after the first stage of chip development, known as tape-out, a critical stage in semiconductor manufacturing when the initial design is sent to the fabrication plant.
- This process costs tens of millions of dollars and takes three to six months, with no guarantee of success. If it fails, Meta must identify the errors and repeat the tape-out process.
Meta has previously attempted to develop its own custom chip, but abandoned it at the small-scale testing stage, similar to the one currently underway for the new chip.
Instead, the company was betting on buying billions of dollars worth of Nvidia's GPUs in 2022.
Meta uses these chips to train its artificial intelligence models, which are used in recommendation systems, advertising, and other services.
This year, the value of GPUs has been called into question. In January, Chinese startup DeepSeek unveiled new low-cost models that work more efficiently with less processing power than most other models on the market.
Following DeepSeek's launch, shares of AI-related companies fell sharply, with Nvidia shares losing up to 20% of their value. They have since recovered most of their losses as investors continue to believe in Nvidia's dominance.
