Intel Panther Lake iGPU Can Use Up to 93% of System RAM
Intel has introduced a new Shared GPU Memory Override capability for its Panther Lake platform, significantly increasing the amount of system memory that can be assigned to compatible integrated GPUs.
With the latest Arc Pro graphics software and driver, supported Arc Pro B370 and B390 iGPUs can reportedly allocate up to 93% of installed system RAM. On a 64GB system, that translates to approximately 59.5GB of shared memory available to the integrated GPU.
The feature is primarily aimed at professional applications, local AI workloads, visualization, and other memory-intensive use cases where applications may impose minimum VRAM requirements. It does not turn system RAM into dedicated GPU memory, nor does it eliminate the performance limitations of an integrated graphics architecture.
⚙️ Panther Lake Raises the Shared Memory Ceiling #
Intel calls the new capability Shared GPU Memory Override. It is supported on the integrated Arc Pro B370 and B390 graphics processors associated with the Intel Core Ultra Series 3 / Panther Lake platform.
The feature is enabled through recent Intel graphics software and driver releases.
| Parameter | Configuration |
|---|---|
| Feature | Shared GPU Memory Override |
| Supported GPUs | Arc Pro B370 / B390 |
| Platform | Core Ultra Series 3 / Panther Lake |
| Default Allocation | Up to 57%, depending on system memory |
| Maximum Allocation | Up to 93% |
| 64GB System Example | Approximately 59.5GB |
| Graphics Software | Intel Graphics Software 25.26.1602.2 or newer |
| Driver | 32.0.101.6974 or newer |
The available allocation scales with the amount of installed system memory. Systems equipped with larger RAM capacities therefore have substantially more memory that can be exposed to GPU workloads.
Why Intel Is Raising the Limit #
The change appears designed primarily to improve flexibility rather than raw graphics performance.
Professional applications and AI software sometimes check the amount of GPU memory available before loading large projects or models. A relatively small default iGPU allocation can therefore become a software limitation even when the system has substantial unused RAM.
Increasing the reported and available GPU memory can help address these capacity-related restrictions.
🧠 Shared Memory Is Not Dedicated VRAM #
The most important technical distinction is that this memory remains system RAM.
An integrated GPU does not suddenly gain a dedicated pool of GDDR or HBM simply because more system memory has been allocated to it. CPU and GPU resources continue to access the same underlying DDR or LPDDR memory subsystem.
That means the change primarily increases capacity, not memory bandwidth.
| Memory Type | Typical Role | Key Characteristic |
|---|---|---|
| Dedicated GDDR VRAM | Discrete GPUs | High GPU-oriented bandwidth |
| HBM | High-end accelerators | Extremely high bandwidth |
| Shared DDR/LPDDR | Integrated GPUs | Common CPU/GPU memory pool |
This distinction is critical when evaluating the practical implications of the 93% figure.
A system with 64GB of RAM may be able to expose nearly 60GB to the iGPU, but the GPU still operates within the bandwidth and latency characteristics of the platform’s memory subsystem.
Capacity and Bandwidth Are Different Constraints #
Increasing available memory capacity can allow larger workloads to load, but it does not increase the rate at which the GPU can move data.
For memory-bandwidth-sensitive workloads, the underlying memory interface remains the limiting factor.
Consequently, the feature should be viewed as a capacity and compatibility enhancement, rather than a substitute for a discrete GPU.
🤖 Local AI Workloads Could Benefit #
One of the more interesting applications is local AI inference.
Large models can require substantial memory capacity, and integrated GPUs have historically been constrained by the relatively small amount of memory exposed to graphics workloads.
A system with 64GB or 128GB of RAM can potentially make a much larger portion of its memory available to GPU workloads through the new override.
This could make Panther Lake systems more flexible for:
- Local AI model execution
- AI development and experimentation
- GPU-accelerated professional applications
- Large texture datasets
- Visualization workloads
- Compute testing
- Media-processing applications
However, model size alone does not determine inference performance. Memory bandwidth, compute throughput, software optimization, quantization, and workload characteristics remain equally important.
🖥️ Professional Software Compatibility Is a Key Target #
The feature may be particularly useful when applications perform strict VRAM-capacity checks.
Some professional workloads require a minimum amount of GPU memory before allowing a project or feature to run. An integrated GPU could therefore encounter a capacity check even when the underlying system has substantial unused RAM.
By increasing the amount of memory exposed to the GPU, Intel can potentially remove some of these artificial capacity barriers.
This is especially relevant for thin-and-light notebooks and compact workstations that rely exclusively on integrated graphics.
The change effectively gives system designers another way to exploit large RAM configurations without adding a discrete GPU.
📈 More Memory Does Not Mean Discrete-GPU Performance #
The headline 93% allocation figure should not be interpreted as meaning Panther Lake’s integrated GPU can compete directly with a discrete graphics card equipped with a comparable amount of VRAM.
A discrete GPU typically has dedicated high-bandwidth memory optimized for graphics and compute workloads.
An integrated GPU instead shares the system memory subsystem with the CPU and other platform components.
For demanding workloads such as high-refresh-rate gaming, large-scale GPU rendering, or compute-intensive AI inference, memory bandwidth and GPU compute resources can become much more important than raw capacity.
The new feature therefore addresses a different problem: making more of the system’s existing memory available to workloads that need it.
🔬 Real-World Performance Still Needs Testing #
The practical impact of the feature will depend heavily on software and platform implementation.
Several areas require independent testing before the 93% allocation limit can be evaluated properly.
Driver and BIOS Behavior #
OEM BIOS configurations, Windows memory management, and Intel’s graphics drivers will determine how consistently the feature behaves across different systems.
Laptop manufacturers may also implement their own memory-allocation policies.
Dynamic Memory Reallocation #
A key question is how smoothly memory is reclaimed when CPU workloads require additional RAM.
If GPU memory allocations can be dynamically adjusted without significant latency or system stalls, systems with large RAM capacities could make better use of otherwise idle memory.
Independent testing is needed to determine whether aggressive allocations introduce measurable stuttering or responsiveness issues.
Application-Level Performance #
Benchmarks should focus on actual workloads rather than memory capacity alone.
Useful test scenarios include:
- Local AI inference
- Blender and other GPU renderers
- Video encoding and export
- Professional visualization
- Gaming
- Multitasking
- Large project-file workloads
These tests will reveal whether the additional memory produces meaningful performance improvements or primarily helps applications pass VRAM-capacity checks.
🧩 A Practical Upgrade for AI PCs #
Intel’s Shared GPU Memory Override is best understood as a resource-management feature for integrated graphics, not a fundamental change to GPU architecture.
The ability to allocate up to 93% of system RAM gives Panther Lake systems substantially more flexibility when running memory-intensive software, particularly on configurations equipped with 64GB or more of RAM.
Its biggest potential benefit is therefore not higher peak GPU performance, but the ability to prevent memory-capacity limitations from unnecessarily blocking workloads.
For professional users and AI developers operating on compact systems without discrete GPUs, that distinction could be meaningful.
🔭 Panther Lake Expands the Role of Integrated Graphics #
The move toward much larger shared-memory allocations reflects a broader change in how integrated GPUs are being positioned.
As AI PCs increasingly use the iGPU as a general-purpose accelerator rather than merely a display engine, memory capacity becomes an increasingly important system-level resource.
Intel’s 93% ceiling does not solve the bandwidth and compute limitations inherent to integrated graphics, but it provides a more flexible memory model for workloads that are constrained primarily by capacity.
Ultimately, the feature’s value will be determined by application compatibility, driver maturity, memory-management behavior, and real-world benchmarks.
The headline number is impressive, but the more significant development is Intel treating system memory as a larger shared resource for AI and professional GPU workloads on Panther Lake.