Intel announced its plans to delve into the development of chips to power artificial intelligence applications. After years of delays and a move that widened the gap with NVIDIA, the manufacturer is determined to return. Intel will launch Falcon Shores, an AI chip that it will look to compete head-on once it lands in 2025.
During its participation in the ISC conference of high performance computers, Intel presented a GPU with up to 288 GB of memory. Falcon Shores is designed to accelerate deep learning training and supports 8-bit floating point computing. The next-generation GPU will be part of Max, a family of high-performance computer (HPC) chips.
In terms of performance, Falcon Shores boasts 288 GB of HBM3 memory with a bandwidth of up to 9.8 TB/s. Intel’s new AI chip supports data types ranging from FP64 to BFLOAT16 or FP8 and supports Compute Express Link (CXL), a standard that enables resource sharing for higher performance. According to Intel, Falcon Shores is based on a modular architecture tiled that allows you to implement multiple discrete GPU and CPU configurations.
Intel confirmed that its new AI chip will be available from 2025. Falcon Shores will rise from the ashes of Rialto Bridge, a GPU with which he planned to close the gap with NVIDIA. In a presentation made in May, the company confirmed that Rialto would be the successor to Ponte Vecchio and would land in mid-2023, followed by Falcon Shores in 2024.
To the surprise of some, Intel canceled the plans a few weeks ago and announced that we would not see an AI chip before 2025.
With its new AI chip, Intel is going for NVIDIA, although it may be a little late
In an interview with ReutersJeff McVeigh, interim director of the Accelerated Computing and Graphics Group (AXG), said that the company is taking its time to develop new chipss. Falcon Shores is a key player in plans to combine a CPU and GPU in the future, however, this chip will be released as a GPU.
“While we aspired to have the best CPU and GPU on the market, it was hard to say that one vendor was ever going to have the best combination of them,” McVeigh said. Intel mentioned during its presentation that mixing CPU and GPU cores in Falcon Shores was premature. The explosion of generative AI is changing the requirementsso the optimal combination of these nuclei is constantly changing.
After canceling Rialto Bridge to focus on its next-generation AI chip, Intel said the goal was to maximize return on investment. This means that Ponte Vecchio will be the only GPU for high-performance computers until 2025, a decision that would further distance them from the competition. When Falcon Shores hits the market, it will have to deal with Grace Hopper or Instinct MI300, solutions from NVIDIA and AMD, respectively.
At the moment, NVIDIA leads the production of chips for the artificial intelligence industry. The A100 has become the workhorse for supercomputers used to train Long Language Models (LLMs). Companies like Stability AI employ up to 5,400 A100s for Stable Diffusion, according to a State of AI report.