Intel's D270 fits a 300B model on one card with 480GB LPDDR5X
Intel's inference card bets memory over peak FLOPS: 1.3 PFLOPS FP8, 1.5TB/s, eight-card rigs past a trillion params — sampling now, broad availability in 2027.
Copy markdown
A 300B model on a single card
Intel's new D270 — the Xe3P "Crescent Island" accelerator — packs up to 480GB of LPDDR5X at 1.5TB/s, enough to hold a ~300B-parameter model in memory on one card. Today you'd stripe a model that size across a multi-GPU HBM box; this collapses it to a single slot.
The bet: cheap capacity over peak FLOPS
At 1.3 PFLOPS FP8 and 350W, the D270 sits below Nvidia's Blackwell Ultra and AMD's MI455X on raw speed. Intel's wager is that swapping pricey HBM for denser, cheaper LPDDR5X wins on cost per token for inference — the number your API bills ultimately ride on.
Eight cards clear a trillion params
An eight-GPU D270 system holds models past 1T parameters, aimed at agentic inference where long context and KV caches devour memory. If ODMs ship it near the implied price, self-hosting a frontier-size model gets materially cheaper per token.
The catch: sampling now, GA in 2027
Only select customers are sampling today; broad availability is targeted for 2027, so there's nothing to plan a 2026 deploy around yet. Note also that Intel's own board tops out at 160GB — the headline 480GB configs come from ODM partners.