Intel driver update enables 93% system memory allocation to iGPUs for larger AI models

Intel Pushes Memory Allocation Boundaries for AI Workloads

Intel has released driver version 32.0.101.8517 for Arc Pro GPUs, introducing a significant capability that allows users to dedicate up to 93% system memory allocation to the integrated GPU 1

1

2

2

. This represents a notable increase from the previous 87% limit that Intel established last year with its "Shared GPU Memory Override" feature for Core Ultra Series 2 processors 1

1

. The driver release specifically targets Arc Pro GPUs including the Arc Pro B390 and Arc Pro B370, while also supporting discrete Arc Pro A and B-series cards from the Battlemage and Alchemist lineups 1

1

2

2

.

Enabling Larger AI Model Support on Accessible Hardware

The expanded system memory to iGPU allocation directly addresses one of the primary bottlenecks in running AI models locally: VRAM capacity. Traditional memory partitioning typically limits a GPU to 50% of system RAM, creating significant constraints for LLM inference tasks 1

1

. With this Intel driver update, a system equipped with 32GB of RAM can now allocate 30GB to the GPU, providing sufficient memory to run models like Qwen 2.5 32B at 4-bit quantization with a comfortable context window 1

1

2

2

. Workstations with 64GB of RAM gain even more capability, able to handle heavyweight Large Language Models like Llama 3 70B while maintaining enough headroom for the KV cache and system stability 1

1

.

Competitive Landscape and Performance Considerations

Intel's approach positions the company aggressively against AMD in the AI inference space. While AMD's Ryzen AI chips currently allow up to 87% memory allocation, AMD's Variable Graphics Memory (VGM) technology in high-end configurations like Strix Halo can allocate 96GB from a 128GB pool to the iGPU 1

1

2

2

. On AI MAX+ platforms, users can allocate a massive 112GB of memory to the GPU while running 128GB of system memory 2

2

. However, memory capacity alone doesn't determine performance. Intel's Core Ultra Series 3 (Panther Lake) chips feature fast LPDDR5X-9600 memory delivering bandwidth around 150 GB/s, while AMD's Strix Halo achieves 256 GB/s through its 256-bit memory bus 1

1

. Apple Silicon maintains an advantage with the M5 Max offering 614 GB/s memory bandwidth, though Intel and AMD are competing on flexibility and affordability through technologies like LPCAMM2 1

1

. Apple's Unified Memory Architecture eliminates traditional partitioning entirely, allowing the entire memory pool to be natively accessible to both CPU and GPU simultaneously 1

1

.