Intel's mysterious new datacenter GPU is what Nvidia's Rubin CPX nearly was

Intel's Crescent Island datacenter GPU uses LPDDR5x memory with up to 480 GB, contrasting with industry-standard HBM/GDDR. It aims to address AI workloads similar to Nvidia's shelved Rubin CPX, which prioritized cost-effective prefill acceleration. The shift to disaggregated compute architectures separates prefill and decode phases, reducing reliance on high-bandwidth memory.