Nvidia's RTX Spark is a superchip that fuses a 20-core Arm CPU and a Blackwell GPU on one package with 128GB of unified memory, and it exists for one reason: to run AI agents on the PC itself rather than shipping every request to a data center. Unveiled at Computex 2026, it delivers up to 1 petaflop of AI compute and is Nvidia's clearest statement yet that the next Windows machine is not a faster laptop, it is a local AI computer. Spark laptops and compact desktops arrive this fall from Asus, Dell, HP, Lenovo, Microsoft Surface and MSI.

  • RTX Spark is one superchip: a 20-core Nvidia Grace-class Arm CPU joined to a Blackwell RTX GPU with 6,144 CUDA cores over the NVLink-C2C interconnect.
  • It carries 128GB of unified memory at up to 300 GB/s, so CPU and GPU share one pool instead of copying data across a PCIe bus.
  • The GPU uses fifth-generation Tensor Cores with FP4 precision to hit roughly 1 petaflop of AI compute for on-device agents.
  • Hardware ships this fall from six major OEMs, with Acer and Gigabyte to follow, positioning Spark as a platform rather than a single product.
RTX Spark unified memory versus a discrete GPU PC A traditional PC splits CPU RAM and GPU VRAM across a PCIe bus. RTX Spark puts a 20-core CPU and Blackwell GPU on one package sharing 128GB of unified memory. TRADITIONAL PC CPU GPU system RAM VRAM PCIe RTX SPARK 20-core CPU Blackwell GPU 128GB unified memory NVLink-C2C, up to 300 GB/s, one shared pool One memory pool means large models load without a CPU-to-GPU copy, which is what lets a laptop hold an agent-sized model in the first place. genztech.blog
Fig 1 A normal PC splits RAM and VRAM across PCIe. Spark's unified memory lets the CPU and Blackwell GPU share one 128GB pool, so a large local model does not have to be shuttled between chips.

What is a superchip, and why fuse the CPU and GPU?

A superchip is Nvidia's term for putting a CPU and GPU on one package linked by a fast chip-to-chip interconnect instead of a slow motherboard bus. In RTX Spark, a 20-core Arm CPU sits next to a Blackwell RTX GPU with 6,144 CUDA cores and fifth-generation Tensor Cores, connected over NVLink-C2C. The payoff is the shared 128GB memory pool. On a conventional PC, a model living in GPU VRAM has to be copied from system RAM first, and VRAM caps how big a model you can even hold. Unified memory erases that wall, which is the difference between a laptop that can run an agent-scale model locally and one that cannot.

RelatedOpenAI's First Chip, Jalapeno, Targets Cheaper Inference

Why does Nvidia want your agent to run locally?

Because the cloud has real costs the marketing rarely mentions: latency on every request, a recurring bill, and your data leaving the machine. An agent that watches your screen, drafts your email and edits your files is exactly the workload where those costs bite hardest. Running it on-device makes it instant, private and free at the margin. Nvidia describes Spark as moving the PC "from tool to teammate," which is marketing, but the technical claim underneath is sound: 1 petaflop of FP4 compute and 128GB of memory is enough to keep a capable model resident and responsive without a network round trip.

SpecRTX SparkTypical AI laptop 2025Cloud inference
CPU20-core Arm (Grace class)x86 mobileServer CPU
Memory for models128GB unified8 to 24GB VRAMEffectively unlimited
AI compute~1 petaflop (FP4)Fraction of thatVast, shared
LatencyLocal, no round tripLocal, limited sizeNetwork dependent
Running costOne-time hardwareOne-time hardwareRecurring per token

Where does Spark fit next to Vera Rubin?

Spark is not a data-center part. Nvidia's Vera Rubin architecture, now in production, is the successor to Blackwell for the huge racks that train and serve frontier models in the cloud. Spark is the other end of the same strategy: push inference to the edge, onto the desk, so Nvidia sells silicon on both sides of the AI workload. The two are complementary. Vera Rubin handles the training and the giant models; Spark handles the local agent that talks to them and does the everyday work offline. Owning both ends is the point.

The catch: an Arm Windows bet

Spark runs an Arm CPU, and Windows on Arm still carries a compatibility tax. Native Arm apps fly; x86 software runs through translation with a performance and reliability cost that varies by app. Microsoft has spent years narrowing that gap, and putting Surface on the OEM list signals confidence, but anyone buying Spark for legacy desktop software rather than AI workloads should test their exact apps first. The unified-memory advantage is real. The ecosystem maturity is the variable.

RelatedHBM4 Enters Production for NVIDIA's Vera Rubin

What to watch · fall 2026
  • Real local model sizes. 128GB unified memory is the headline. Watch which model sizes actually run well, and how fast.
  • Arm app compatibility. The make-or-break for mainstream buyers is whether their existing Windows software just works.
  • Price versus a discrete GPU. Spark competes with cheaper laptops plus a cloud subscription. The value case rests on price.
  • Battery under AI load. Sustained agent workloads are power hungry. Real-world runtime is the spec to verify.

Our take

RTX Spark is the most coherent hardware argument for the local-AI PC anyone has made, precisely because it fixes the real bottleneck instead of just adding cores. Memory, not raw compute, is what kept agent-sized models off laptops, and 128GB of unified memory is the honest answer to that. The strategy is clean too: Vera Rubin owns the cloud, Spark owns the desk, and Nvidia collects either way. The open questions are all execution, not vision. Arm compatibility has to be seamless, the price has to undercut a laptop-plus-subscription, and the battery has to survive a real agent workload. If those land, Spark is the template every PC maker copies. If they slip, it is a brilliant chip waiting for the software to catch up. Either way, the direction is set: the interesting silicon war of 2026 is being fought on the desktop, not just in the cloud.

Primary sources

Original analysis by GenZTech. Figures current as of July 2026. Source: nvidianews.nvidia.com