Overview
The Intel GPU D270 is the first product in Intel data centre inference GPU line after it was formally named at the 2026 OCP Global Summit. The family previously appeared under the codename Crescent Island and is built on the Xe3P architecture. It targets LLM serving, long-context workloads and concurrent inference, with Intel stating that a single card handles a 300B parameter model and an eight-card system handles a 1T parameter model.
On specifications, the D270 carries 32 Xe cores and 256 XMX engines with 1.5TB/s of FP8 compute, paired with up to 480GB of LPDDR5X memory at 1.5TB/s of bandwidth. Board power is 350W, the card uses a standard PCIe AIC form factor with PCIe Gen5 support, and it allows air cooling, which means it drops into conventional server chassis.
Beyond multi-tenant AI serving, Intel lists robotics and other latency-sensitive workloads as targets. Alongside the card, Intel and its partners showed new rack designs based on Xeon 6+ processors, with some configurations scaling past 46,000 CPU cores at up to 133.45kW.
Key Features
- Xe3P architecture: Data centre inference GPU, previously known by the codename Crescent Island
- 32 Xe cores and 256 XMX engines: FP8 compute rated at 1.5TB/s
- Up to 480GB LPDDR5X: Memory bandwidth also rated at 1.5TB/s
- 350W board power: Allows air cooling, no liquid cooling retrofit required
- Standard PCIe AIC form factor: PCIe Gen5 support, fits conventional server chassis
- 300B model on one card: Intel states one card serves a 300B parameter model and an eight-card system reaches 1T
- Built for long context and concurrent inference: Designed for LLM serving, long-context workloads and concurrent inference
- Robotics and latency-sensitive work: Targets include multi-tenant AI serving, robotics and other latency-sensitive workloads
Use Cases
- Multi-tenant inference serving in the data centre
- Deploying long-context LLMs for inference
- Robotics and other latency-sensitive inference loads
- Sites without liquid cooling that need air-cooled deployment
- Teams already standardised on PCIe Gen5 infrastructure planning expansion
Pros
- Up to 480GB of memory on a single card, well above typical accelerators
- Air cooling at 350W, so no cooling retrofit needed
- Standard PCIe AIC form factor with PCIe Gen5, fits existing server chassis
- A single card covers a 300B parameter model, eight cards reach 1T
- Memory bandwidth and FP8 compute both at 1.5TB/s, a balanced pairing for long-context work
Pricing
Intel has not published pricing or general availability for the D270. Pricing, availability and channel details follow Intel official announcements.
Summary
The Intel GPU D270 is the first card in Intel newly named data centre inference GPU line, moved from the codename Crescent Island to the Intel GPU D-series at the 2026 OCP Global Summit and built on Xe3P. The specification is clear enough: 32 Xe cores and 256 XMX engines with 1.5TB/s of FP8 compute, paired with up to 480GB of LPDDR5X at the same 1.5TB/s of bandwidth. The two numbers that matter here are memory capacity and cooling, since 480GB on one card at 350W with air cooling, in a standard PCIe AIC form factor with PCIe Gen5, means no special work on the data centre. Intel positions it for LLM serving, long-context and concurrent inference, stating one card covers a 300B parameter model and eight cards reach 1T, with robotics and latency-sensitive loads in scope. Pricing and availability are still unpublished, so if you plan capacity around it, Intel official announcements are what to track.
Version History
- Crescent Island line named Intel GPU D-series at OCP 2026, first product is the D270 (2026-10-13): At the 2026 OCP Global Summit, Intel announced that its data centre inference GPU based on the Xe3P architecture, previously known by the codename Crescent Island, is now named the Intel GPU D-series, with the Intel GPU D270 as its first product. The D270 carries 32 Xe cores and 256 XMX engines with 1.5TB/s of FP8 compute, paired with up to 480GB of LPDDR5X memory at 1.5TB/s of bandwidth. Board power is 350W, the card uses a standard PCIe AIC form factor with PCIe Gen5 support and allows air cooling. The series targets LLM serving, long-context workloads and concurrent inference, with Intel stating one card handles a 300B parameter model and an eight-card system handles a 1T parameter model, aimed at multi-tenant AI serving, robotics and other latency-sensitive workloads. Intel and its partners also showed new rack designs based on Xeon 6+ processors, with some configurations scaling past 46,000 CPU cores at up to 133.45kW.