Intel GPU D270 is an inference-optimized data center GPU built for agentic AI, long contexts, and high-concurrency serving at practical power and cost.
Hi everyone!
Intel's new D-series makes a pretty unusual choice for an AI GPU: LPDDR5X instead of HBM, with up to 480GB on a single card.
Intel's own card has 160GB. The 480GB configurations are for partner-built versions.
It's aimed at long contexts and agents sharing a GPU, where keeping model weights and KV caches in memory matters. The D270 is rated for 1.5TB/s bandwidth and 1.3 PFLOPS peak FP8 compute in a 350W, air-cooled PCIe design.
That's a lot of memory, though bandwidth is well below the fastest HBM accelerators. So how many tokens per dollar can it actually serve?
Replies