Intel's Crescent Island datacenter GPU ships with up to 480 GB of LPDDR5x memory. It uses a PCIe form factor and lacks HBM or GDDR memory. The GPU offers significantly more memory than Nvidia's flagship GPUs at 288 GB. Intel's design targets lower costs amid rising memory prices, though bandwidth is limited to around 1.2 TB/s. The shift to disaggregated compute architectures highlights prefill as a key bottleneck in AI inference.
Tap to vote and see what everyone thinks.
AMD's Helios MI455X Uses UALink-over-Ethernet at Launch
Summary by ByteBrief