Intel Reveals D270 at OCP 2026: 1.3 PFLOPS FP8, 480GB LPDDR5X, No HBM
Intel used OCP 2026 to officially name and spec its long-previewed Crescent Island AI accelerator. The product is called the Intel GPU D-Series, and the first SKU is the D270, built specifically for AI inference with no 3D graphics capability.
Specs
| Metric | D270 |
|---|---|
| Peak FP8 | 1.3 PFLOPS |
| Memory bandwidth | 1.5 TB/s |
| Max memory | 480 GB LPDDR5X |
| Reference card | 160 GB LPDDR5X |
| TDP | 350W |
| Form factor | PCIe 5.0 x16, air-cooled |
| Cores | 32 Xe3P, 256 XMX engines |
| L2 cache | 32MB unified |
| 3D graphics | None |
Memory Architecture: LPDDR5X, Not HBM
The D270 is unusual in its memory choice. Most competing AI accelerators use high-bandwidth memory (HBM) for density and bandwidth. Intel chose LPDDR5X, arguing the trade-off favors inference: LPDDR5X enables higher capacity at lower cost per gigabyte, which matters when you need to keep large model weights resident in memory for long-context and concurrent workloads.
Intel’s reference card ships with 160GB. Manufacturing partners can configure cards up to 480GB, giving significant headroom for very large models without multi-card sharding.
Scale
A single D270 can hold models of roughly 300 billion parameters. An eight-card server can support models exceeding one trillion parameters, putting it in range of the largest deployed frontier models. Intel frames the target workloads as large language models, long-context inference, and high-concurrency serving — not training.
Why No Graphics Hardware
Intel removed conventional 3D-graphics and ray-tracing units from the die, reclaiming silicon area for the 256 XMX matrix-multiplication engines and the 32MB unified L2 cache. The D270 is a GPU in architecture only; it will not drive a display. This is the same trade-off AMD made with the MI series, choosing compute density over graphics versatility.
Status
Customer sampling is underway with server manufacturers, cloud providers, and software vendors. Intel has not announced pricing or a general availability date. The air-cooled, standard PCIe form factor means deployment into existing server infrastructure without liquid-cooling retrofits.
Intel’s previous D-Series preview at Computex in June 2026 established the Xe3P architecture and the air-cooled form factor; OCP is where the company disclosed final compute and bandwidth figures.