Microsoft Maia 300 Pushes HBM4 Demand as Intel Bets on LPDDR5X for Inference

Microsoft and Intel both stepped up their challenge to NVIDIA with new in-house AI accelerators unveiled on Monday. Microsoft's Maia 300 uses a multi-chiplet architecture built on TSMC's 2nm class process for core computing functions, tapping the foundry's N2P node for compute-intensive chiplets while shifting memory controllers and I/O functions to more cost-effective 3nm dies optimised for bandwidth and connectivity. Microsoft said the architecture delivers 36% more compute FLOPs, 2.5 times the HBM bandwidth and 14% higher scale-up bandwidth than its predecessor.

The memory content is the headline for suppliers. Maia 300 moves from six HBM3E stacks in Maia 200 to eight HBM4 stacks, increasing memory capacity from 216GB to 288GB. Combined with a doubling of accelerator density to 96 chips per rack, total HBM stacks per rack reportedly jump 2.7-fold from 288 to 768. Microsoft plans to roll out potentially more than one million Maia 300 accelerators next year, with each rack consuming around 167kW, and claims the platform delivers roughly 40% more computing output per megawatt than leading GPU-based systems. SK hynix is seen as the leading HBM4 contender having been sole HBM3E supplier for Maia 200, while Samsung could also secure orders with its 36GB 12-high HBM4 now in mass production.

Intel took the opposite route at the OCP Global Summit, formally branding its Crescent Island inference accelerator as the Intel GPU D-Series. The first card, the D270, is based on the enhanced Xe3P architecture with 32 Xe cores and 256 XMX matrix engines running at roughly 2.5GHz, delivering 1.3 PFLOPS of peak FP8 matrix throughput and 10.2 TFLOPS of FP64, with 32MB of L2 cache and a PCIe Gen5 x16 interface. Memory bandwidth is 1.5TB per second, delivered by a 20-chip, 80-channel LPDDR5X bus clocked at roughly 9,375MT/s, and the card is rated at 350W in an air-cooled PCIe CEM form factor.

Intel's own reference card ships with 160GB of LPDDR5X, while ODM partners will be able to offer versions with up to 480GB using 24GB LPDDR5X chips, a far larger capacity than any other PCIe form-factor accelerator announced so far. By comparison, AMD's similar PCIe CEM MI350P offers a theoretical peak of 2.3 PFLOPS of dense FP8, roughly 40% more, but in a 600W configuration; Intel is instead leaning on performance per watt and prefill efficiency. Intel confirmed it is sampling cards to customers but has not announced a launch date, so AI server vendors looking for a PCIe inference alternative are likely to wait until 2027.