Intel D-series AI accelerator hits 1.3 PFLOPS FP8, beating Nvidia B300 memory capacity

Intel has announced a new graphics processing unit designed specifically for artificial intelligence inference workloads, featuring an unusually large memory configuration. The device, part of the D-series, was revealed at the Open Compute Project Global Summit. According to TechRadar, the accelerator breaks previous records for data centre GPU memory capacity. It is engineered to handle large language models and enterprise services that require substantial data accessibility during processing.

Root cause

The design choice stems from Intel’s focus on inference rather than training or gaming. TechRadar reports that Intel selected LPDDR5X memory, a technology typically found in laptops, instead of the GDDR7 standard used in modern discrete graphics cards. This decision prioritises memory capacity and power efficiency. The configuration allows the hardware to keep significant model weights accessible without the high power draw associated with traditional high-bandwidth memory solutions.

Attack path

This product does not involve a security vulnerability or attack vector. The source material describes a hardware specification and market positioning strategy. Intel aims to provide infrastructure that delivers performance at a lower cost for enterprises. The focus is on enabling practical deployment of high-capacity inference systems. There are no reported exploits, bugs, or weaknesses in the D-series architecture mentioned in the available reporting.

Affected versions

The D-series, previously known by the codename Crescent Island, is the specific product line discussed. TechRadar notes that pricing and availability details remain undisclosed. The hardware is intended for data centre environments. It is designed to integrate into existing infrastructure rather than replace consumer-grade components. No specific version numbers or firmware releases were provided in the announcement.

Mitigation

As this is a product announcement rather than a security incident, no mitigation steps are required. There are no patches or updates to apply. Security operations teams do not need to modify configurations or isolate systems. The information serves as market intelligence regarding competitor hardware capabilities. Monitoring for future software support or driver updates may be relevant for long-term planning.

Why it matters

Teams evaluating AI infrastructure must understand the trade-offs between memory capacity and bandwidth. High capacity allows larger models to run on fewer devices, reducing hardware costs. However, inference speed depends on how quickly data moves, not just how much is stored. Organisations should assess whether their workloads benefit from holding large models in memory or require faster throughput. This distinction guides procurement decisions for generative AI services.

Background: Generative AI

Generative AI creates new content, code, or media by learning statistical patterns from vast datasets. Traditional AI predicts outcomes or classifies inputs based on defined rules or supervised learning. The former introduces risks of hallucination and data leakage; the latter offers predictable, auditable decision-making paths for security operations.

Read the full guide: Generative AI vs Traditional AI: Operational Differences and Risk Profiles

What to do and how to stay safe: Intel

• Review current AI inference workloads to determine if model size exceeds existing GPU memory limits.
• Compare the total cost of ownership between high-capacity LPDDR5X solutions and high-bandwidth alternatives.
• Monitor vendor announcements for software support and driver compatibility with the D-series.
• Assess power consumption budgets, as 350W units impact data centre cooling and energy requirements.

Step-by-step guide: Stop Sensitive Data Leaks in AI Chatbots: Priority Controls

General security guidance from the Vector Bulletin newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

How much memory does the Intel D-series GPU have?

It features 480GB of LPDDR5X memory, which is the highest capacity for a data centre GPU according to the report.

Does this GPU support large language models?

Yes, a single card can support models with up to 300 billion parameters, while an eight-GPU system handles over one trillion.

How does its performance compare to Nvidia?

It offers 1.3 petaflops of FP8 performance and 1.5TB/s bandwidth, surpassing the B300 in memory capacity but not necessarily speed.