Intel says its Falcon Shores GPUs will be ready to support the next generation of LLMs despite a previously announced delay.

Intel wants to play a bigger role in training AI models. The Intel Xeon Scalable CPUs are an important part of the servers on which Large Language Models (LLMs) are built, but in fact they mainly serve a supporting role. The hard work is on the GPU accelerators, and Nvidia is the master there.

Intel therefore wants to throw a new weapon into the fight against the next generation of LLMs: Falcon Shores. The Falcon Shores GPUs should go to customers in 2025, the manufacturer confirms. This timing isn’t particularly spectacular, as Intel was previously forced to delay Falcon Shores.

Worth the wait?

Exactly what we are waiting for is also not so clear. Intel will share some details at the ISC conference in Hamburg. The manufacturer plans to equip the chips with a whopping 288 GB of HBM3 memory. For comparison: The currently strongest accelerator is the Nvidia H100 Hopper with “only” 80 GB RAM. In addition, Intel states that the chip will support 8-bit floating point workloads, which is relevant for efficient training of new models.

However, when Falcon Shores launches, the landscape of HPC accelerators will look very different. Nvidia does not stand still and wants to continue to impress the market with the super chip Grace Hopper, which combines ARM CPUs with GPU power. AMD has a similar concept ready with the Instinct MI300. A Falcon Shores chip with a CPU component was planned but has since been scrapped.