Intel Arc Pro Driver Update Boosts System RAM Allocation to 93% for AI Models

Intel has released a driver update for its Arc Pro graphics line that significantly expands the amount of system memory available to integrated GPUs (iGPUs). The update allows select systems to allocate up to 93% of total system RAM to be used as video memory (VRAM), a move designed to facilitate the local execution of larger Artificial Intelligence (AI) large language models (LLMs).

The update, identified as driver version 32.0.101.8517 (Q1.26.R2), targets professional graphics solutions and workstation users. By increasing the allocatable memory from previous limits—which some reports indicate were around 87%—to 93%, Intel is attempting to lower the hardware barrier for users who wish to run complex AI workloads without investing in high-cost discrete GPUs with large onboard VRAM capacities.

Technical Scope and Hardware Compatibility

The new driver is optimized for a wide range of Intel’s professional graphics hardware. This includes discrete video cards from the Arc Pro B70, B65, B60, and B50 series, as well as models within the Arc Pro A family.

Crucially, the update provides support for integrated graphics processors, specifically the Arc Pro B390 and B370. These iGPUs are integrated into Intel Core Ultra Series 3 processors, which are developed under the code name Panther Lake.

In typical integrated graphics architectures, the GPU shares system memory (RAM) because it lacks its own dedicated high-speed memory chips. Traditionally, the operating system and BIOS limit how much of that system RAM the GPU can claim. By raising this ceiling to 93%, Intel enables the iGPU to access a much larger pool of memory, which is a primary requirement for loading the massive parameter sets found in modern LLMs.

Impact on Local AI Inference

The primary objective of this memory expansion is to improve local AI inference. When running an AI model locally, the entire model—or a significant portion of it—must reside in the GPU’s memory to ensure acceptable performance. If a model is too large for the available VRAM, the system must either offload the work to the much slower CPU or fail to load the model entirely.

By allowing an iGPU to utilize nearly the entire capacity of the system’s RAM, users with high-capacity memory configurations (such as 64GB or 128GB of RAM) can effectively treat their PC memory as a massive VRAM pool. This allows for the execution of larger models that would otherwise require enterprise-grade discrete GPUs.

This development is part of a broader shift toward AI PCs, where the goal is to move AI processing from the cloud to the edge. Local inference offers several advantages, including increased data privacy, reduced latency, and the elimination of recurring subscription fees associated with cloud-based AI services.

Industry Context

Intel’s approach targets a specific gap in the market: the space between entry-level laptops and high-end AI workstations. While discrete GPUs from NVIDIA are the industry standard for AI due to their dedicated VRAM and CUDA cores, they remain expensive and power-hungry.

By leveraging the unified memory architecture of the Core Ultra processors, Intel is positioning the Arc Pro line as a viable alternative for developers and creators who need to experiment with LLMs but cannot justify the cost of professional-grade discrete hardware.

The release of version 32.0.101.8517 marks a transition from simple maintenance updates to functional enhancements that change how system resources are partitioned. For users of the Panther Lake architecture, this update transforms the iGPU from a basic display driver into a more capable tool for computational AI workloads.