Intel Unveils D270 GPU: 1.3 PFLOPs FP8 Compute with 480GB LPDDR5X at 350W

Key points

• Intel’s D270 GPU delivers 1.3 PFLOPs of FP8 compute with a 350W TDP, targeting air-cooled data centers.
• The chip uses the Xe3P architecture with 32 cores, 256 Vector Engines, and 256 XMX units.
• Memory configuration includes up to 480GB of LPDDR5X, offering 1.5 TB/s bandwidth, avoiding HBM supply constraints.
• The D-Series supports models up to 300 billion parameters on a single card and over one trillion in an eight-GPU setup.
• Customer sampling for the D270 is expected in the second half of 2026.

Background

Intel previously teased the Xe3 architecture during its Panther Lake and Xe3 deep dives. The new Xe3P architecture represents a scalable upgrade from client iGPUs to data center AI GPUs, building on Intel’s existing multigenerational Xe install base to improve efficiency and software compatibility.

How outlets are covering it

Wccftech emphasizes the competitive advantage of LPDDR5X over HBM, noting that it lowers power, cost, and supply constraints compared to NVIDIA’s Vera Rubin and AMD’s MI450X, which rely on HBM4. Intel’s official release highlights the strategic importance of open infrastructure for agentic AI, stressing that Xeon CPUs and open software ecosystems are crucial for orchestrating complex workflows. Tom’s Hardware focuses on the technical specifications, confirming the 1.3 PFLOPs FP8 peak and 1.5 TB/s memory bandwidth. While Wccftech notes the trade-off in peak bandwidth for memory-bound tasks, Intel frames the high capacity as essential for long-context inference and large language models.

Why it matters

As AI inference becomes the dominant workload, enterprises face rising costs and power constraints. Intel’s D-Series offers a cost-effective, power-efficient alternative to high-end HBM-based GPUs, enabling scalable AI deployment without specialized cooling infrastructure. This positions Intel to compete in the value-oriented segment of the AI market, addressing the growing demand for flexible, economical infrastructure for agentic AI workloads.

What to watch

Intel plans to begin customer sampling for the D270 GPU in the second half of 2026. The company is collaborating with OEMs, ODMs, and software providers to develop validated, deployment-ready solutions. Future iterations of the D-Series will leverage optimizations from Intel’s existing Arc Pro B-series lineup, ensuring early access to performance enhancements.