Intel GPU D-Series Delivers 1.3 PFLOPS AI Inference With 480GB Memory in Air-Cooled Design

Intel GPU D-Series Launches With Xe3P Architecture for AI Inference

Intel officially unveiled its GPU D-Series AI inference accelerator at the Open Compute Project Global Summit, giving the chip formerly code-named Crescent Island its commercial name and disclosing detailed specifications for the first time1

1

2

2

. The Intel GPU D-Series is built on the company's new Xe3P architecture, delivering 1.3 PFLOPS of FP8 compute alongside 1.5 TB/s of memory bandwidth and up to 480GB of LPDDR5X memory, all within a 350W TDP1

1

2

2

. The chip uses a standard PCIe form factor with air-cooled design, allowing deployment in existing data center environments without specialized cooling infrastructure2

2

.

Targeting Large Language Models and Trillion Parameter Workloads

The GPU D-Series is designed specifically for large language models, long-context inference, and concurrent inference services1

1

2

2

. According to Intel, a single card can support models with up to 300 billion parameters, while an eight-GPU system can handle models exceeding one trillion parameters2

2

3

3

. The first product in the lineup, the Intel GPU D270, features 32 Xe cores and 256 matrix engines optimized for agentic AI and large-memory workloads3

3

. "Inference is becoming the defining AI workload, and enterprises need infrastructure that delivers the right performance at the right cost," said Anil Nanduri, vice president of AI Products and GTM at Intel's Data Center Group2

2

.

LPDDR5X Memory Strategy Balances Capacity and Cost

Intel's choice of LPDDR5X memory for the GPU D-Series represents a strategic departure from competitors like Nvidia and AMD, who rely on high-bandwidth memory (HBM) for their data-center accelerators4

4

. While LPDDR5X generally offers lower bandwidth than HBM, it provides high capacity with lower cost and power consumption3

3

. The 480GB capacity significantly exceeds AMD's Instinct MI450X at 432 GB of HBM4 and Nvidia's Vera Rubin at 288 GB of HBM44

4

. This memory configuration positions the GPU D-Series for inference workloads where model residency and capacity are as important as peak arithmetic throughput3

3

. The 480GB capacity could reduce the need to partition large models across multiple accelerators, simplifying deployment and reducing communication overhead3

3

.

Performance Positioning Against Nvidia and AMD Alternatives

The GPU D-Series targets a different performance and infrastructure profile than Nvidia's highest-end hardware2

2

. Each of the 72 Rubin GPUs in Nvidia's Vera Rubin NVL72 rack-scale system delivers 17.5 PFLOPS of FP8 throughput, but requires liquid cooling and draws substantially more power1

1

2

2

. Among PCIe AI accelerator options, the Intel GPU D270 compares favorably: AMD Radeon AI PRO R9700 reaches 766 TFLOPs of FP8 at 300W, while Nvidia RTX PRO 5500 Blackwell delivers 846 TFLOPs of FP8 at 600W4

4

. The AMD Instinct MI350P reaches 2.3 PFLOPS but with FP8 micro-scaling4

4

. Intel's bundling of LPDDR5X at 160 GB for reference designs and up to 480 GB for partner models positions it as a value-optimized product4

4

.

Xe3P Architecture and Software Ecosystem Foundation

The Xe3P architecture powering Crescent Island is the first chip built on Intel's massively parallel computing architecture, lacking most graphics-specific functional units that would characterize a traditional GPU1

1

. Each Xe3P core packs 8 Vector Engines with FP8+FP4 precision and 8 XMX engines with 16-deep systolic arrays4

4

. The GPU D270 comprises four compute slices with 8 cores each, totaling 32 Xe3P cores, 256 Vector Engines, and 256 XMX units4

4

. Intel is highlighting the "common Xe software foundation" it has built for AI across its Xe Cores in integrated GPUs and discrete GPU products1

1

. The company announced Day 0 support for Meta's Muse Glimmer model and native upstream integration with AI frameworks including vLLM, SGLang, llm-d, and Nvidia Dynamo4

4

5

5

.

Market Implications and Deployment Timeline

Customer sampling of GPU D-Series products is currently underway, though Intel did not announce a broad launch date or pricing2

2

3

3

. Intel is working with a broad ecosystem of OEMs, ODMs, neoclouds, independent software vendors, appliance providers, and AI-native companies for commercialization1

1

5

5

. The company is also contributing to the Open Adapter Card workgroup to define a standardized sub-rack accelerator card form factor for improved interoperability5

5

. This represents Intel's latest attempt to gain traction in AI hardware infrastructure after the failure of the Data Center GPU Max family (Ponte Vecchio) and its mothballed follow-ons1

1

. In a market hungry for inference compute cycles, a product with proven architectural design targeted specifically at tokens-as-a-service providers in an accessible thermal and electrical envelope might provide Intel the toehold it needs as AI models advance and agentic workflows assume greater importance1

1

.