NVIDIA describes RTX Spark as one physical memory pool shared by the CPU and integrated GPU. Windows presents that pool in separate GPU-dedicated, shared, and CPU-only regions, so configurable GPU memory does not mean the system lacks unified memory.
What Paul Thurrott’s review criticizes
Paul Thurrott published a negative RTX Spark review on October 31, 2026, after reporting more than two weeks with the ASUS ProArt P16, HP OmniBook Ultra 16, and Lenovo Pro 9n. His criticisms include price, laptop heat, weight and noise, memory configuration, Windows 11 functionality, and local-AI usability. Those are the reviewer’s assessments.
Thurrott describes the GPU memory as manually configured rather than unified. The terms address different things: NVIDIA’s memory architecture guide describes the physical pool, while GPU allocation controls determine how Windows presents portions of it. NVIDIA says users can adjust the GPU allocation in NVIDIA App and recommends leaving at least 16GB for the CPU, Windows, and basic apps.
One physical pool, three Windows regions
Unified memory means the CPU and integrated GPU share physical memory. Windows exposes that pool as a GPU-dedicated carveout, shared system memory that both CPU and GPU can access, and CPU-only memory that cannot be assigned to the GPU.
That makes the remainder after a GPU allocation more than one undifferentiated block. It is still part of the installed system memory, not extra RAM, and the CPU-only portion is not GPU-accessible.
The 128GB system with a 96GB GPU allocation
ETA PRIME demonstrates a 128GB ASUS ProArt P16 with 96GB assigned to GPU memory. Subtracting the allocation leaves 32GB. Under NVIDIA’s nominal partition formula, that remainder divides into 16GB of shared memory and 16GB of CPU-only memory. The result is a calculation from the formula, not a live reading of the laptop’s memory counters; Windows and hardware reservations affect usable capacity.
| Memory region | Amount in this example | Role |
| Installed physical memory | 128GB | Total system memory in the demonstrated configuration |
| GPU-dedicated carveout | 96GB | Portion reserved as dedicated GPU memory |
| Remaining after the carveout | 32GB | 128GB minus 96GB |
| Shared system memory, nominal | 16GB | Accessible to both CPU and GPU |
| CPU-only memory, nominal | 16GB | Available to the CPU, not assignable to the GPU |
What the split means for local AI
NVIDIA says large language models and AI agents can use additional system memory when they exceed the GPU allocation. The software must support and manage that use, and shared-memory access can be slower than access to the GPU-dedicated region, as NVIDIA explains in its CUDA memory guidance and RTX Spark support instructions. NVIDIA also warns that reserving too much for GPU workloads can leave too little memory for the host and make the system unresponsive.
For a separate look at local-AI coding workloads on the Lenovo Yoga 9n, see NeoTeo’s report on the Yoga 9n’s coding tests.
