Tuesday 13 October 2026
Morning Edition
ZOTPAPER

News without the noise

ZOTPAPER
Hardware & Devices

Intel Launches GPU D-Series AI Accelerator, Targets Efficient Inference with 1.3 PFLOPS

Crescent Island chip finally named; air-cooled design aims at trillion-parameter models and multi-tenant AI workloads

By LineZotpaper
Published
Read Time2 min
At the Open Compute Project (OCP) 2026 conference, Intel officially named its long-teased Crescent Island AI accelerator the GPU D-Series, promising 1.3 PFLOPS of FP8 performance and 1.5 TB/s of memory bandwidth in an air-cooled, 350-watt PCIe form factor. The chip is designed for large language model inference, long-context reasoning and multi-tenant AI services, with Intel saying a rack of eight such GPUs can handle models exceeding one trillion parameters.

Intel has revealed the final specifications for its GPU D-Series, the product formerly known as Crescent Island. The chip is the first built on Intel's Xe3P architecture, though it lacks most graphics-specific units typically found in GPUs, making it a dedicated massively parallel inference accelerator.

The company disclosed that each GPU D-Series module will offer up to 480GB of LPDDR5X memory and a 350W TDP, enabling deployment in traditional air-cooled rack servers. Intel aims this product at workloads including large language models, long-context inference and latency-sensitive applications such as robotics and multi-tenant AI services.

For comparison, each GPU in Nvidia's Vera Rubin NVL72 system delivers a theoretical 17.5 PFLOPS of FP8 throughput, but requires liquid cooling and significantly higher power. Intel positions its offering as fitting a different performance and infrastructure envelope — accessible to data centres that cannot support liquid cooling at scale.

Intel says it is working with a broad ecosystem of OEMs, ODMs, neoclouds, independent software vendors and AI-native companies to bring the accelerator to market. Customer sampling of GPU D-Series products is already in progress, though the company has not announced a broad launch timeline.

The GPU D-Series arrives after Intel's earlier Data Center GPU Max family (Ponte Vecchio) failed to gain traction, and subsequent follow-on projects were mothballed. In a market hungry for inference compute, Intel hopes a proven architectural approach in an accessible thermal envelope will finally secure its foothold in the AI accelerator segment.

§

Analysis

Why This Matters

  • The GPU D-Series directly targets the fast-growing inference market, where demand for efficient, air-cooled hardware is rising as models scale up.
  • Intel's success or failure here will signal whether it can recover from years of missteps in AI hardware and compete with Nvidia's dominance.
  • For data centre operators, a viable Intel alternative could lower costs and reduce dependency on a single supplier for AI inference infrastructure.

Background

Intel has been developing AI accelerators for the data centre for years, launching the Ponte Vecchio GPU in 2022 for HPC and AI workloads. That product failed to achieve significant market adoption, and Intel later cancelled or stalled several follow-on designs. The GPU D-Series represents the company's latest attempt, leveraging its Xe architecture but stripped of graphics-specific hardware to focus purely on AI inference.

Key Perspectives

Intel: The GPU D-Series fills a gap for air-cooled, lower-power inference acceleration. With support for trillion-parameter models and multi-tenant services, Intel claims the chip is ideal for enterprises and neoclouds that cannot deploy liquid-cooled racks. Market observers: Nvidia's Rubin GPUs offer dramatically higher peak throughput but require vastly more power and cooling. The two products are not direct competitors; Intel is targeting a different segment — cost- and power-conscious deployments. Skeptics: Intel has repeatedly failed to deliver a competitive AI accelerator. Without proven customer adoption and a robust software ecosystem, the GPU D-Series risks repeating past disappointments even if the hardware is solid.

What to Watch

  • First customer deployments and performance benchmarks from early sampling partners.
  • Broader launch timeline and volume availability in 2027.
  • Adoption by major neocloud providers and enterprise data centres for inference workloads.

Sources

Zotpaper

Written by software from the reporting listed above, scored by an automated standards desk, and published without a person reading it first. If something here is wrong, tell the editor and it will be put right.

How we workSubscribe