By Walter Schulze
Publication Date: 2026-10-13 00:00:00
Intel finally gave its inference chip a real name and a memory number big enough to make hyperscalers look twice: the GPU D270 can carry up to 480GB, triple what Nvidia’s most common inference cards typically ship.
Intel used the OCP Global Summit on October 12 to drop the codename and introduce the real product: the GPU D270, the first entry in a new Intel GPU D-Series built specifically for AI inference. For more than a year the chip was known only as Crescent Island, a placeholder that outlived its usefulness the moment Intel decided to actually ship the thing. Now it has specs, and the specs tell you exactly what Intel is betting on.
The headline number is memory. Intel’s reference design ships with 160GB of LPDDR5X, but the architecture lets outside manufacturers build cards carrying up to 480GB. That’s not a typo or a rounding error, it’s the whole strategy. Where Nvidia and AMD chase raw FLOPS with GDDR7 or HBM, Intel went the other way: cheaper, more abundant LPDDR5X, stacked high. The tradeoff shows up in bandwidth, 1.5TB per second, which trails the HBM-equipped cards from Nvidia and AMD but comes at a fraction of the cost and none of the current HBM supply crunch.
On compute, the D270 delivers 1.3 peak petaflops of FP8 throughput across 32 Xe3 cores and 256 XMX engines, inside a 350-watt envelope. AMD’s Instinct MI350P, by comparison, hits 2.3 petaflops of FP8, nearly double, but it draws up to 600 watts to get there. Intel isn’t trying to out-muscle…
