RetroDeckRETRODECK

NVIDIA's 1 Petaflop RTX Spark Laptops: What the Number Means for Local AI

NVIDIA's RTX Spark laptops, on sale since October 2026 from $2,599, pair a Grace CPU and Blackwell GPU with up to 128GB of unified memory. The 1 petaflop figure is 4-bit FP4 and likely assumes sparsity. Capacity lets 120B-class models fit; about 300GB/s of bandwidth still limits how fast they generate.

What NVIDIA launched

The laptops in the viral clips are RTX Spark machines, built on NVIDIA's RTX Spark N1X superchip: an Arm-based Grace CPU and a Blackwell RTX GPU sharing one pool of LPDDR5X memory. ServeTheHome reported the launch on October 17, 2026, and NVIDIA's blog said on October 28 that the laptops were available from $2,599, with 128GB configurations on shelves. The six laptop makers are ASUS (ProArt P16), Dell (XPS 16), HP (OmniBook Ultra 16), Lenovo, Microsoft (Surface Laptop Ultra) and MSI (Prestige N16 Flip AI+). ServeTheHome said desktops were due the following month, with Microsoft's Surface RTX Spark Dev Box listed at $6,000.

NVIDIA's product page lists two configurations. The top one pairs a 20-core Grace CPU with a 6,144-core Blackwell GPU and up to 128GB of memory. The second pairs an 18-core CPU with a 5,120-core GPU and up to 64GB. Both are rated at 45 to 80 watts and run Windows 11. ServeTheHome describes the CPU as ten Cortex-X925 and ten Cortex-A725 cores, and the chip as the same or similar silicon to the GB10 inside DGX Spark, with faster LPDDR5X-9400 memory and a 55 TOPS NPU.

What "1 petaflop" actually measures

A petaflop is 10 to the 15th floating-point operations per second. NVIDIA's figure is "up to 1 petaflop of FP4 AI performance", and each qualifier matters.

FP4 is a 4-bit number format. It is the lowest precision Blackwell tensor cores accelerate, and it is only reached when a model's weights and activations are actually in an FP4 format such as NVFP4. The same chip delivers a fraction of that number at FP16 or in general-purpose code.

Sparsity is the second qualifier. NVIDIA's RTX Spark page does not footnote whether the figure assumes 2:4 structured sparsity. For DGX Spark, whose GB10 chip ServeTheHome describes as the same or similar silicon, NVIDIA partner PNY's spec page labels the identical 1 petaflop as theoretical FP4 performance using the sparsity feature. If RTX Spark follows the same convention, dense FP4 is about half. Treat that as likely, not confirmed.

"Up to" is the third. The figure belongs to the top configuration at its highest power limit, and a thin laptop at 45 watts will not sustain it.

Why the compute number barely affects chat speed

Running a language model has two phases. Prefill reads your whole prompt at once, which is compute-bound, so the petaflop helps there: long prompts, large documents and agent contexts get processed faster. Decode generates the answer one token at a time, and each token requires reading the model's active weights from memory. That phase is limited by memory bandwidth, not FLOPS.

NVIDIA has not published RTX Spark's bandwidth on its product page. ServeTheHome lists 300GB/s for N1X against 273GB/s for GB10, and Notebookcheck also cites up to 300GB/s. The arithmetic gives a ceiling: a dense 27B model at 4-bit is about 18GB of weights, so 300GB/s caps generation near 16 tokens per second before any overhead. A mixture-of-experts model reads only its active experts per token, which is why models like gpt-oss 120B, with about 5 billion active parameters, are the natural fit for this kind of machine. Real throughput lands below these ceilings.

What 128GB of unified memory changes for local models

Capacity is the real news. A discrete laptop GPU tops out at a fraction of this memory, so models above roughly 30 billion parameters have meant offloading to system RAM and a steep slowdown. With 128GB shared by CPU and GPU, a 120B-class mixture-of-experts model at 4-bit fits entirely in memory the GPU can address. NVIDIA's page claims models up to 120B parameters with a 1M token context. That second half depends on the model: long context costs KV cache memory, and only architectures with sparse or linear attention keep a million tokens affordable.

Capacity also lets several models stay loaded: a chat model, an embedding model, a speech model and an agent's working set at once, without swapping. The comparison between capacity and bandwidth is covered in more depth in the guides on AMD's 192GB Ryzen AI Halo against NVIDIA and on large unified memory PCs.

What it does not change

Software is the open question. These are Windows on Arm machines. ServeTheHome notes they rely on Microsoft's Prism emulator for x86 apps, that AVX2 support was added only recently, and that some anti-cheat systems still do not work. Check that your inference stack, whether Ollama, LM Studio, llama.cpp or a CUDA toolchain, has a native Arm64 Windows build before buying.

Model quality does not change either. A 120B open-weight model on a laptop is a large step for private and offline work, but the hardest long-horizon tasks still favour hosted frontier models. And every token a model reads, local or hosted, costs time: on a bandwidth-bound machine a shorter prompt is a faster answer.

That is where retrieval earns its place. RDK indexes folders and Obsidian vaults as encrypted private chunks that stay on your node, and agents query them through MCP or the CLI, getting a few hundred tokens of context instead of whole documents. On repeated or reference-heavy work that cuts token spend 80 to 90 percent, and on a local machine it shortens prefill on every turn.

Frequently asked questions

Is the RTX Spark laptop really 1 petaflop?
Only at FP4, the 4-bit format, in the top configuration, and probably with structured sparsity. NVIDIA's RTX Spark page does not footnote sparsity, but partner spec pages label the identical DGX Spark figure as sparse. Dense FP4 would then be about half, and higher precisions much lower. For general computing it is not a petaflop machine.
Can an RTX Spark laptop run a 120B model?
With 128GB, yes, at 4-bit. gpt-oss 120B is about 65GB in the Ollama library, leaving room for context and other apps. The 64GB configuration cannot hold it. Speed depends on memory bandwidth, around 300GB/s by third-party reports, so mixture-of-experts models with few active parameters run far better than dense models of similar size.
How is RTX Spark different from DGX Spark?
ServeTheHome describes them as the same or similar silicon. DGX Spark is a small Linux box built on GB10 with a ConnectX-7 network card at 140 watts. RTX Spark N1X targets Windows 11 laptops at 45 to 80 watts, with faster LPDDR5X-9400 memory, about 300GB/s against 273GB/s, and an NPU for low-power tasks.
Does unified memory make local models as fast as a desktop GPU?
No. Unified memory gives capacity: larger models fit without offloading. A desktop graphics card with GDDR7 has several times the bandwidth, so a model that fits in its memory generates tokens faster. Unified memory wins when the model is too big for the card, and loses on speed when it is not.