NVIDIA RTX Spark: What 128 GB of Unified Memory Means for Local AI
NVIDIA used Computex 2026 to do something it had never done before: ship a system-on-a-chip for ordinary Windows laptops and small desktops. The RTX Spark pairs a 20-core Arm CPU with a Blackwell RTX GPU and — the part that matters most if you run models locally — up to 128 GB of unified memory, with native CUDA support. If you have ever fought the 24 GB VRAM ceiling on a laptop, that is the headline.
This is a preview, not a review. RTX Spark systems are slated to ship in fall 2026, and NVIDIA has not published pricing — the product page only offers a "Notify Me" signup. Here is what we actually know, and what it means for local inference. And to be clear up front: this is neatly packaged, but it is not a new idea — RTX Spark is NVIDIA's consumer entry into a category AMD's Strix Halo and NVIDIA's own DGX Spark already defined: a big unified-memory APU built for local AI.
RTX Spark is not DGX Spark. They share a name and NVIDIA's Grace-Blackwell lineage, but they are different products. DGX Spark is the $3,999 GB10 desktop "AI supercomputer" for developers that shipped in late 2025 and holds models up to ~200B parameters. RTX Spark is a consumer SoC for mainstream Windows laptops and compact desktops. When you read coverage online, make sure it is about the one you think it is.
The specs that matter
| Spec | RTX Spark (as announced) |
|---|---|
| CPU | Up to 20-core Arm |
| GPU | Blackwell RTX, up to 6,144 CUDA cores (~RTX 5070 laptop-class, per early coverage) |
| Memory | Up to 128 GB unified (LPDDR5X) |
| AI performance | Up to 1 petaflop (FP4) |
| Graphics | Full DLSS suite, ray tracing, Reflex, G-SYNC; AV1 + 4:2:2 encode/decode |
| Software | Native CUDA |
| Form factors | Laptops and compact desktops |
| Availability | Fall 2026 |
| Price | Not announced (early reports suggest roughly $1,800–$2,900 by configuration — treat as rumor) |
Launch laptops named by NVIDIA include the ASUS ProArt P16, Dell XPS 16, HP OmniBook X 14, Lenovo Yoga Pro 9n, Microsoft Surface Laptop Ultra, and MSI Prestige N16 Flip AI+.
Why unified memory is the real story for local LLMs
On a normal gaming laptop, the GPU's VRAM is a hard wall. A 24 GB RTX 4090 laptop can hold a 30B model at Q4 but chokes on a 70B; the moment weights spill into system RAM, throughput collapses. (The rough math is ~0.5 GB per billion parameters at Q4, plus headroom for the KV cache — see our guide to running a 27B model locally.)
RTX Spark borrows Apple Silicon's trick: one pool of memory shared by CPU and GPU. With up to 128 GB, the GPU can address far more than any discrete laptop card:
- A 70B model at Q4 (~40 GB with context) fits with room to spare.
- 100B+ MoE models with modest active-parameter counts become viable on a laptop.
- Long-context work, where the KV cache balloons, stops being a memory emergency.
And unlike Apple Silicon, RTX Spark runs native CUDA — so the tooling you already use (CUDA builds of llama.cpp, vLLM, PyTorch, and the wider CUDA ecosystem) works without the MLX/Metal detour. That combination — Mac-like memory capacity plus the CUDA stack — is what makes it genuinely interesting for local AI.
The catches: bandwidth and Windows on Arm
Two caveats keep this honest:
- Bandwidth is the ceiling — and we can already estimate it. Token generation is memory-bandwidth-bound, and unified LPDDR5X has far less bandwidth than the GDDR7 on a discrete RTX card or the HBM in the datacenter. We do not need to wait for RTX Spark to know roughly where it lands, because its siblings are already on desks: DGX Spark runs ~273 GB/s and AMD's Strix Halo ~256 GB/s (closer to 212 GB/s measured). NVIDIA has not put RTX Spark's figure on a spec sheet, but early reporting from its developer forums pegs memory bandwidth at ~300 GB/s — a touch above DGX Spark — which puts it squarely in the same band. (Ignore the 600 GB/s number you may see quoted: that is the die-to-die chip interconnect, not memory bandwidth.) The consequence is the important part: because token generation is bandwidth-bound, every machine in this class performs similarly — DGX Spark manages roughly 38 tokens/sec on a 120B model, only marginally ahead of Strix Halo. Expect RTX Spark in the same neighbourhood. 128 GB of LPDDR5X is generous on which models you can load and modest on how fast they run.
- It is Arm — so Windows on Arm applies. RTX Spark is an Arm SoC, which means the Windows-on-Arm software story matters. Arm-native builds of your AI toolchain will fly; anything still x86-only runs under emulation. Before you plan a local-AI workflow around it, confirm your stack — inference engine, drivers, Python wheels — has Arm-native Windows support.
How it stacks up
- vs a discrete RTX laptop (24–32 GB): RTX Spark trades peak GPU horsepower and memory bandwidth for far more addressable memory. Bigger models, possibly slower per token.
- vs AMD Strix Halo (Ryzen AI Max+ 395): The closest comparison, and the reason RTX Spark is evolutionary, not revolutionary. Strix Halo has shipped since 2025 with the same recipe — up to 128 GB unified LPDDR5X at ~256 GB/s, a large RDNA 3.5 iGPU, an XDNA 2 NPU — and runs 70B models at Q4 today. It is x86 (no Windows-on-Arm caveat) but runs on ROCm/Vulkan, not CUDA. The RTX Spark trade is essentially: same memory model and similar bandwidth, but you get NVIDIA's CUDA stack and RTX graphics in return for going Arm.
- vs Apple Silicon (M-series): A comparable unified-memory philosophy, but RTX Spark brings native CUDA — a big deal if your workflow is CUDA-first. (See our take on Apple Silicon for local models.)
- vs DGX Spark: Closer than the names suggest. Early community teardowns describe RTX Spark as essentially the same GB10-class silicon as DGX Spark on a near-identical board — just without the ConnectX-7 networking, at a lower TDP, and running Windows instead of DGX OS. DGX Spark is the pricier ($3,999) developer desktop with a 200B-parameter ceiling; RTX Spark is the mainstream consumer sibling in a laptop you would actually carry. Same engine, different trim.
Should you wait for it?
So what is RTX Spark, really? A neatly executed product — but not a new one. AMD's Strix Halo and NVIDIA's own DGX Spark already proved out the 128 GB unified-memory APU for local AI; RTX Spark brings that recipe to mainstream Windows laptops with NVIDIA's CUDA stack and RTX graphics attached. Because the whole class shares the same LPDDR5X bandwidth ceiling, do not expect it to out-run its siblings on tokens/sec — what you are really buying is the CUDA ecosystem and RTX features in this form factor, not a speed breakthrough.
If you keep hitting the VRAM wall on a laptop and your toolchain is CUDA-first, that is a genuinely useful package. If you simply want the most local-model-per-dollar today, a Strix Halo mini-PC already does the job on ROCm. The two things still worth waiting for are official pricing and real Arm-native software support — both of which arrive closer to the fall 2026 ship date. We will benchmark it against Strix Halo, Apple Silicon, and discrete RTX the moment units land.
Source: NVIDIA RTX Spark product page and NVIDIA's Computex 2026 announcement. Specifications are as announced; pricing and final performance are not yet confirmed.