Intel D270 GPU: 480 GB LPDDR5X for AI Inference
Original: Intel D270 GPU AI Accelerator–Up to 480 GB LPDDR5X
Why This Matters
Massive LPDDR5X capacity offers a cost-efficient HBM alternative for inference at scale.
Intel launched the D270 GPU, an inference-optimized data center accelerator with up to 480 GB LPDDR5X memory, 350 W air-cooled TDP, 32 Xe3P cores, and 256 XMX matrix engines targeting agentic AI workloads.
Intel's D270 is the latest entry in its GPU D-series lineup, pitched squarely at inference rather than training. The card packs up to 480 GB of LPDDR5X memory per card — a notably large pool that Intel says allows entire models and long contexts to stay resident, reducing the need for high-bandwidth memory (HBM). The thermal envelope sits at 350 W air-cooled, using standard PCIe infrastructure.
Intel frames the design around 'agentic AI': workloads where agents run across thousands of steps, 100K-token contexts, and dozens of concurrent sessions. Die area that would normally go to graphics and ray-tracing pipelines has been reallocated to Xe3P cores and matrix compute — 32 Xe3P cores and 256 XMX matrix engines in total.
The pitch to operators is economic: dense, air-cooled, standard-rack deployment keeps tokens-per-watt and tokens-per-rack metrics favorable without requiring specialized power or cooling buildout. Intel explicitly targets customers who need to scale inference capacity within existing data center constraints, not new HBM-heavy builds.