59.5GB for the iGPU! Intel's New Driver Pushes Shared Memory Cap to 93%
Intel has released a new driver (version 32.0.101.8804) for Arc Pro workstation GPUs that raises the shared memory allocation limit for Panther Lake's integrated graphics from 87% to 93%. On a 64GB system, this means up to 59.5GB can be allocated to the iGPU as video memory — a massive pool that opens the door for serious local AI inference on integrated graphics.
💡 What You Will Learn
Intel has released a new driver (version 32.0.101.8804) for Arc Pro workstation GPUs that raises the shared memory allocation limit for Panther Lake's integrated graphics from 87% to 93%. On a 64GB sy
Intel's latest driver update (32.0.101.8804) targets the Panther Lake processor's built-in Arc Pro B390 iGPU, pushing the shared memory ceiling from 87% to 93%. On a 64GB machine, that's roughly 3.8GB of additional headroom — the difference between "can't fit" and "barely fits" for many quantized open-source LLMs in the 30-35GB range.
Panther Lake's iGPU is no slouch: - Architecture: Xe3 - Cores: 12 Xe3 cores, 1536 ALUs - AI Units: 96 Tensor cores - Max Frequency: 2500MHz - TDP: 80W
It supports LPDDR5X-9600 (up to 96GB) or DDR5-7200 (up to 128GB) dual-channel memory. The combination of high-frequency memory and a higher allocation cap means the traditional iGPU bottleneck — competing with the CPU for memory — is being addressed from both ends.
Two caveats: 1. 59.5GB of shared memory ≠ 59.5GB of dedicated VRAM. The iGPU shares system memory bandwidth with the CPU, so running a 50GB model means the CPU is essentially starved. 2. The 93% cap primarily benefits workstation users with 64GB, 96GB, or 128GB configurations. Casual users with 16GB or 32GB laptops won't see much difference.
Bottom line: This isn't just a driver update — it's Intel removing the last barrier to making iGPUs viable for local AI workloads. A Panther Lake laptop with 96GB of RAM can effectively function as an 89GB "AI workstation." Bandwidth remains the bottleneck, but capacity is no longer the wall it used to be.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.
