Intel has begun placing its new D-Series data center GPUs into customers' hands for testing, marking the company's latest attempt to win a meaningful slice of the AI hardware market. The pitch is not about building the biggest or fastest training system available. Instead, Intel is aiming squarely at AI inference—the stage of the workflow in which already-trained models generate answers to user queries. Inference is the phase end users directly experience, and as more models move from development into daily service, running them efficiently has become an increasingly important focus for chipmakers.

What Intel is shipping

The D-Series was previously known under the codename Crescent Island and is built on Intel's Xe3P architecture. The flagship model is the D270, which Intel is positioning for AI inference and agentic workloads—in this context, AI systems that carry out multi-step actions rather than answering a single prompt and stopping.

On paper, the D270 delivers peak compute of approximately 1.3 PFLOPs at FP8 precision, a lower-precision number format that trades some numerical detail for substantially faster math. The card carries a thermal design power of 350W.

Under the hood, the chip packs 32 Xe3P cores and 256 XMX engines. The XMX engines are Intel's dedicated matrix-math units, which handle the heavy multiplications that neural networks run on. A unified 32 MB L2 cache ties the design together.

The memory bet

The most consequential design choice is memory. Instead of HBM—the high-bandwidth memory used in many competing accelerators—the D270 relies on LPDDR5X. The two memory types occupy different corners of the market: HBM stacks memory dies directly beside the processor to deliver top-end bandwidth at a premium price, while LPDDR5X is the same memory class used in smartphones and laptops, produced at far larger volumes and lower cost.

Intel's reference configuration ships with 160 GB of LPDDR5X, and partners can expand that to 480 GB. Peak memory bandwidth is estimated at around 1.5 TB/s. The two levers cut in different directions: capacity determines how much of a model and its working data a card can hold at once, while bandwidth governs how quickly that data moves during generation.

Intel says it is targeting long-context and agentic AI applications with a value-oriented approach. The company's argument centers on high memory capacity relative to power draw and cost, rather than chasing peak bandwidth numbers.

Built to slot into existing racks

The D270 uses an air-cooled design and a PCIe Gen5 form factor. PCIe cards plug into standard servers, meaning data center operators can add them without rebuilding their infrastructure around specialized cooling or proprietary boards. According to Intel, Crescent Island emphasizes scalability and integration over sheer training performance.

Timeline and Xeon connection

Customer sampling for the D270 is targeted for the second half of 2026. Intel planned previews at Computex and Hot Chips 2026 ahead of broader availability.

The D-Series also forms part of a larger strategic play: Intel wants to pair its GPUs with its Xeon processors to build more complete AI inference systems. The pairing reflects how inference stacks are typically assembled, with server CPUs handling data movement and orchestration while GPUs run the model math. Intel's OEM partners include ASUS, Dell, and HPE.

Competitive context

Intel is framing the D270 as a cost-effective option for AI workloads that competes with offerings from NVIDIA and AMD. The risks are real. LPDDR5X offers capacity, but its bandwidth sits well below what HBM-based designs can deliver, and some workloads are bandwidth-hungry.

Customer feedback from the sampling phase in the second half of 2026 will be closely watched, as early reactions should show whether enterprises view the D270 as a genuine alternative or a niche option. Between now and then, the milestones to track are the Computex and Hot Chips previews and the first OEM systems from ASUS, Dell, and HPE—along with whatever performance data emerges once sampled cards are in customers' hands.

Source: CryptoBriefing