Intel Lunar Lake’s 10GB Memory Tax: Independent Audit of Xe2 VRAM Regression & Bug #1435

Status: CASE #1435 (Closed as Duplicate / Locked)

Internal Report ID: 22022038499

Audit Date: May 11, 2026

📑 Executive Summary

This document serves as a public record of a critical architectural flaw discovered in the Intel Core Ultra 7 258V (Lunar Lake) processors featuring the Arc 140V graphics suite. The audit focuses on a severe memory allocation regression during Large Language Model (LLM) inference using the Vulkan 2.13 API.

Despite providing high-fidelity logs, the official reporting channel was suppressed under the guise of "moderation," necessitating this external, non-deletable record.

💻 System Specifications

• Processor: Intel Core Ultra 7 258V
• Integrated Graphics: Intel Arc Graphics 140V (Xe2 Architecture)
• Driver Version: 32.0.101.8724 (WHQL)
• Testing Environment: LM Studio / Vulkan 2.13 Backend

🔍 Technical Analysis: The "9.7GB Overhead"

During inference testing with the Qwen 3.6 35B A3B (Q3_KS) model, a massive discrepancy in memory management was identified:

1. Expected Allocation: ~20.0 GB (including standard system overhead).
2. Actual Physical Allocation (Lunar Lake): 29.7 GB.
3. The Regression: A sustained, unexplained overhead of 9.7 GB (~33% loss of effective VRAM capacity).

Consequences:

The overhead triggers an immediate TDR (Timeout Detection and Recovery) event as the system exceeds shared memory limits. This leads to a $0x380$ in bios error, resulting in a driver crash and total system instability.

📉 Post-Crash Performance Decay

Following the initial TDR event, the Xe2 architecture exhibits "zombie state" behavior:

• Performance Loss: A ~40% drop in tokens-per-second (TPS) compared to baseline, even after reloading smaller models.
• Output Corruption: Text generation becomes incoherent ("gibberish output"), indicating that the driver fails to clear or re-initialize memory buffers correctly after a Vulkan crash.

🚫 Audit Trail: Corporate Information Suppression

The handling of Issue #1435 on the Intel GPU Community Issue Tracker (IGCIT) raises significant concerns regarding transparency:

• False Equivalency: The mobile-specific Lunar Lake issue was summarily closed as a "Duplicate" of #1330 (a general desktop driver issue), ignoring the unique memory-sidecar architecture of the 258V series.
• Conversation Lockdown: The thread was marked as "Too Heated" and locked, preventing the submission of 16 redundant mirrors containing additional diagnostic logs.
• Tone Policing: Technical critiques regarding the regression were characterized as "threatening behavior" by moderators (Karen-Intel), effectively ending the technical dialogue in favor of PR damage control.

🛡️ Evidence Preservation

To prevent the erasure of these findings, all technical logs, memory dumps, and correspondence have been mirrored across decentralized archives:

• Wayback Machine: Verified snapshots as of 11.05.2026.
• Archive.ph: High-fidelity mirrors of the original GitHub Issue #1435.
• Decentralized Storage: Log files including $0x380$ signatures have been distributed to prevent single-point-of-failure deletion.

Conclusion: The Intel Lunar Lake Xe2 architecture currently suffers from a documented 9.7GB VRAM regression that renders Large LLM inference unstable. Consumers and developers should be aware that Internal Bug #1435 remains unresolved despite being closed by the manufacturer.

Sources:

https://github.com/IGCIT/Intel-GPU-Community-Issue-Tracker-IGCIT/issues/1435

Mirror: https://archive.ph/Q5F7m

INDEPENDENT CONSUMER AND AUDITOR