Intel hasn’t given up on the accelerator market and announces a new data center GPU. Crescent Island is based on its own Xe architecture and stands out for its large memory capacity.
Intel unveils Crescent Island, a new AI accelerator tailored for the data center market. Crescent Island aims to provide customers with a solution for the growing demand for inference GPUs. It’s noteworthy that Intel still feels compelled to play a significant role in this space. The CPU specialist has been increasingly focusing on its core activity: building CPUs. Moreover, accelerator specialist Nvidia has just invested in Intel.
Nevertheless, the company believes there’s a role for its internally developed Xe architecture. Crescent Island is built on Xe3P and promises high performance per watt. Xe3P is a minor upgrade from Intel Xe3. Intel also uses Xe for the GPU components in its CPUs and as the foundation for its Arc GPUs for workstations.
Slow but Plentiful
Intel targets inference and therefore integrates 160 GB of LPDDR5X memory. Both the volume and type stand out: Crescent Island has a vast amount of memory, but not the most performant kind. Intel opts for Low Power DDR5X memory instead of HBM2 or GDDR6.
Intel differentiates itself from the competition in this way. Nvidia and AMD typically choose faster but consequently more expensive memory. Crescent Island promises to combine top specifications like the 160 GB memory with a lower price tag, though we don’t have details about this. For workloads where memory volume is more important than speed, Intel seems to be positioning an interesting component in the market.
For now, Intel appears to be keeping its promise to maintain its commitment to the GPU segment. Whether this ambition is sustainable in the long term, especially now that Nvidia is an investing partner, remains to be seen.