Updated October 25, 2026: This report separates NVIDIA’s published RTX Spark specifications and model-capacity claims from performance that still needs independent, configuration-specific measurement.
RTX Spark is a new class of Windows laptop and compact desktop built around an NVIDIA Blackwell RTX GPU, an NVIDIA Grace CPU and unified LPDDR5X memory. NVIDIA says systems start at $2,599 and are available from partners including ASUS, Dell, HP, Lenovo, Microsoft and MSI. The headline is capacity: supported configurations offer up to 128GB of unified memory rather than the much smaller dedicated-memory pools common in conventional consumer GPUs.
What NVIDIA’s specifications establish
| Published item | What it supports | What it does not prove |
|---|---|---|
| Up to 128GB LPDDR5X unified memory | A larger shared CPU/GPU capacity ceiling on eligible configurations | How much memory a particular workload can allocate, or its speed |
| Up to a 6,144-core Blackwell RTX GPU and 20-core Grace CPU | The platform and highest listed configuration | Cross-system application performance from core counts alone |
| 45–80W listed superchip TDP for laptop configurations | The published processor power range | Complete-system draw, sustained clocks, acoustics or battery life |
| Windows 11 and the CUDA/TensorRT/NVFP4 stack | The intended operating system and NVIDIA software path | Compatibility with every local-AI frontend, model format or kernel |
| NVIDIA claim: models up to 120B parameters with up to 1M context | The vendor’s advertised upper-end use case | A universal fit or speed result across precisions, models and applications |
Capacity is the interesting part
A shared 128GB pool can make workloads possible that cannot fit wholly in 24GB or 32GB of dedicated VRAM. That does not make the full 128GB an application budget. Windows, active applications, model weights, KV cache, runtime workspaces, visual inputs and concurrent requests all use the same pool. A smaller-memory configuration changes the conclusion again.
Parameter count is not a memory specification. Precision, quantization, architecture, projector or adapter files, context length, cache type and runtime implementation determine the complete workload. NVIDIA’s “up to 120B” statement should therefore be read as a vendor platform claim, not a promise that every 120B checkpoint and one-million-token session will fit or respond at an acceptable rate.
Why this article no longer estimates tokens per second
The earlier version derived broad token-per-second ranges from memory bandwidth. That was too weak for a buying decision. Bandwidth can constrain autoregressive inference, but observed speed also depends on the exact model, quantization, prompt and generated lengths, cache, batching, kernels, thermals, power mode, software version and whether the workload remains in the intended memory path. A bandwidth estimate is not a substitute for a measured application result.
HighVRAM will publish a speed comparison only when the source reports the exact RTX Spark configuration and enough workload detail to interpret the result. Until then, the defensible conclusion is about platform and capacity, not a tokens-per-second ranking.
What to verify before buying
- Exact system: confirm the GPU/CPU tier, installed memory, storage, cooling design and whether memory is upgradeable.
- Software path: verify the intended model format, frontend, CUDA/TensorRT support and Arm compatibility where relevant.
- Complete workload: record model files, context, cache type, batch or concurrency, modality and background applications.
- Measured output: capture load success, time to first token, generated tokens per second, peak shared-memory use, power mode, temperatures and output correctness.
- Alternative system: compare delivered price and workflow support with a discrete-GPU desktop, a larger-memory workstation or another unified-memory computer.
Bottom line
RTX Spark’s strongest documented distinction is a Windows-and-CUDA platform with up to 128GB of unified memory in portable or compact systems. That is potentially valuable for developers and creators whose complete workloads exceed ordinary consumer-GPU capacity. It is not evidence that every large model runs quickly, that 128GB is fully available to the model, or that the $2,599 starting configuration is the configuration needed for the advertised upper-end workloads.
Use the local AI hardware recommender to estimate capacity, compare platform trade-offs in the hardware guide, and build a reproducible benchmark record for an exact RTX Spark configuration.
Primary sources
- NVIDIA Newsroom: RTX Spark PCs available today — availability, starting price, vendor claims and platform summary.
- NVIDIA RTX Spark product page — advertised use cases, memory tiers and software stack.
- NVIDIA RTX Spark specifications — listed CPU/GPU, memory, operating-system and power configurations.
- NVIDIA Developer: local AI hardware overview — NVIDIA’s platform positioning and supported model-size claim.