[News] AI Chip Memory Strategies Diverge: Microsoft’s Maia 300 Fuels HBM4 Demand as Intel Bets on LPDDR5X

Microsoft and Intel unveiled their latest in-house AI accelerators on Monday, stepping up their challenge to NVIDIA with distinct approaches to AI inference performance and efficiency. According to Reuters, Microsoft’s Maia 300 features a multi-chiplet architecture leveraging TSMC’s 2nm process for core computing functions. It also packs eight HBM4 stacks, up from six HBM3E stacks in Maia 200, Seoul Economic Daily notes.

Intel, meanwhile, is taking a different route. Its newly branded D-Series GPU, formerly known as Crescent Island, opts for LPDDR5X over the increasingly costly and supply-constrained HBM favored by NVIDIA and AMD, Wccftech reports, adding that built on the Xe3P architecture, the accelerator supports up to 480GB of LPDDR5X memory.

Microsoft’s Maia 300 Eyes Million-Chip Rollout, Fueling HBM4 Demand

Notably,