NVIDIA RTX Spark: The Arm superchip threatening Apple M5 and reinventing the Windows PC at Computex 2026
🔎 A PC that thinks for itself, finally
NVIDIA just did to Apple what Apple did to Intel: change the rules of the game with a single chip. Unveiled on May 31, 2026, at Computex, the RTX Spark is not just a simple processor. It's an ARM superchip that fuses the Grace CPU, Blackwell RTX GPU, and 128 GB of unified memory into a single package for Windows PCs.
Why now? Because Windows on Arm was sorely lacking a high-end flagship. Qualcomm offered one, but it wasn't enough for heavy AI workloads. Apple M5 was cruising on the closed macOS ecosystem with no rival on the PC side. NVIDIA saw the void and dove in with the architectural brute force that is its signature.
Microsoft's Signal Laptop Ultra, the first device to integrate it, targets a 110W TDP according to Tom's Hardware's hands-on sessions. That means: workstation power in a 15-inch laptop. The era of the agentic AI PC has just begun, and it no longer speaks x86.
The essentials
- The RTX Spark combines a 20-core ARM Grace CPU and a Blackwell RTX GPU with 6,144 CUDA cores, connected by NVLink-C2C
- Up to 128 GB of unified memory shared between CPU and GPU — a game-changer for local AI
- First device: Microsoft's Surface Laptop Ultra 15" with a 110W thermal budget
- Primarily targeting workstation laptops and mini desktops, not consumer PCs initially
- Direct competition: Apple M5, Qualcomm Snapdragon X Elite, and indirectly Intel/AMD in the creative segment
- Promises to run 120-billion parameter models locally
Recommended tools
| Tool | Main usage | Price (June 2026, check on site.com) | Ideal for |
|---|---|---|---|
| Hostinger | Web hosting to deploy AI apps | Starting from 2,99 €/month | Devs hosting local interfaces |
| NVIDIA TensorRT | Local inference optimization | Free (NVIDIA SDK) | Maximizing RTX Spark perfs |
| Ollama | Running LLMs locally | Open source | Testing 70B+ models locally |
RTX Spark Architecture: Grace CPU + Blackwell GPU Under One Roof
The RTX Spark is not a traditional SoC. It's a superchip, a term NVIDIA has mastered since its Grace Hopper architecture for servers, adapted here for the PC world.
The CPU is a 20-core ARM NVIDIA Grace, derived from the same family as the datacenter chips. It is not a Cortex-X modified by a partner: it's pure NVIDIA, designed with the help of a Taiwanese firm according to CNBC. The GPU is a Blackwell RTX featuring 6,144 CUDA cores and 5th-generation Tensor Cores supporting the FP4 format, according to the official NVIDIA press release.
The link between the two? NVLink-C2C, the same interconnect found in high-performance computing systems. This means the CPU and GPU share memory without going through the traditional PCIe bus. No bottlenecks, no unnecessary copying of data between the two units.
This architecture is strongly reminiscent of what NVIDIA had outlined around Vera Rubin and N1X ARM, except that here everything is condensed into a single chip for PCs. The philosophy is identical: unify the CPU and GPU at the silicon level rather than stacking them on a motherboard.
Unified memory: the real silent killer
128 GB of unified memory. This figure alone justifies the existence of the RTX Spark. On a traditional PC, even a high-end one, your GPU is limited to 16 or 24 GB of VRAM. You want to run a heavy model? The system swaps to the SSD and performance collapses.
With the RTX Spark, the CPU and GPU draw from the same 128 GB pool. A 120-billion parameter model in FP4 takes up about 60 to 70 GB. It fits entirely in memory, without aggressive compression, without destructive quantization. The GPU accesses it at full bandwidth via NVLink-C2C.
This is exactly the advantage Apple has claimed with its unified memory since the M1 chip. Except that here, you also get 6,144 CUDA cores and 5th-generation Tensor Cores. The marriage of the Apple approach (unified memory) and the NVIDIA approach (raw GPU power), in short.
128 GB of unified memory: what it actually changes for an AI dev
Unified memory is the difference between a model that runs and a model that crashes. Let's take a real-world case.
A model like Nemotron 3 Ultra 550B, NVIDIA's flagship open-source model, theoretically requires datacenter resources. But with 128 GB of unified memory and the FP4 format supported by 5th-generation Tensor Cores, models in the 70B to 120B range become perfectly executable locally.
For an AI developer, the implications are immediate:
- Rapid iteration: no more need to push every test to a cloud GPU at $2/hour. You iterate locally, you push to prod when it's clean.
- Sensitive data: proprietary code, medical data, legal documents — everything stays on the machine. No possible leaks.
- Zero network latency: local inference eliminates network latency. For agentic applications that make dozens of LLM calls per minute, this is a massive gain.
Tom's Hardware points out that the platform promises to transform Windows into an "agentic AI OS". With 128 GB of unified memory and a Blackwell GPU, a PC can literally run an agent based on GPT-5.4 or Claude Opus 4.7 locally, with long context and integrated tools, without calling a remote API.
Models practically executable
With 128 GB and FP4, here is what becomes realistic:
| Model size | Recommended format | Estimated memory usage | Expected inference speed |
|---|---|---|---|
| 7B-13B | FP16 | 14-26 GB | Ultra-fast, real-time |
| 30B-35B | FP8 | 30-35 GB | Very smooth, interactive |
| 70B | FP4 | ~35 GB | Good, productive use |
| 120B | FP4 | ~60-70 GB | Acceptable, agentic tasks |
| 550B (extreme quantization) | FP4 + pruning | ~110-128 GB | Slow but functional |
Top-tier agentic models like GPT-5.5 (score 98.2) or Gemini 3 Pro Deep Think (95.4) remain out of reach locally — their actual size far exceeds 128 GB even in FP4. But models in the 70B-120B range, which include weights like DeepSeek V4 Pro (Max) scoring 88 in agentic or Kimi K2.6 in self-host at 88.1, become perfectly viable.
Windows on Arm: Microsoft's magic wand
The RTX Spark wouldn't exist without Microsoft's desire to turn Windows on Arm into a serious platform. Until now, WoA was an intermittent promise: x86 emulation worked, but with an overhead that negated the appeal for power users.
The RTX Spark changes the game because it targets an audience that doesn't need to emulate x86: creatives and AI devs. These users already work in natively ARM-compatible ecosystems — Python, PyTorch, TensorFlow, CUDA runtimes — or can compile natively without friction.
Ars Technica notes that the first RTX Spark devices will be "workstation laptops and mini desktops". This is no coincidence. Microsoft and NVIDIA are targeting the pro segment first, where users are willing to pay for power and adapt to a still-young WoA ecosystem. The general public will come later, when prices drop and software compatibility is proven.
The Surface Laptop Ultra 15" is the flagship of this strategy. With a 110W TDP, Microsoft has clearly chosen raw performance over battery life. It's a "portable workstation" positioning, not an "ultrabook". A statement of intent: Windows on Arm is no longer a compromise; it's a choice for power.
The slow death of x86 in creative workflows?
Not so fast. x86 remains dominant for traditional PC gaming, legacy enterprise software, and the entire installed base. But in the creative/pro niche — video editing, 3D, local AI — the argument for x86 is crumbling.
Apple proved this with the Apple Silicon transition. NVIDIA is now replicating this dynamic on the Windows side, with an advantage: CUDA cores. Creatives using Blender, DaVinci Resolve, or CUDA-optimized Adobe tools will find in the RTX Spark a familiar environment, but with the energy efficiency and unified memory of ARM.
RTX Spark vs Apple M5 vs Qualcomm : the honest comparison
TechRadar titles it bluntly: "Watch out, Apple — Nvidia just unveiled its RTX Spark Arm superchip". Is the hyperbole justified? Let's look at the facts.
| Specification | NVIDIA RTX Spark | Apple M5 (estimated) | Qualcomm Snapdragon X Elite |
|---|---|---|---|
| CPU Architecture | ARM Grace 20 cores | ARM Apple (est. 16-20 cores) | ARM Oryon 12 cores |
| GPU | Blackwell RTX 6144 CUDA cores | Apple GPU (est. 40 cores) | Adreno 680 |
| Max Unified Memory | 128 GB | Up to 128 GB (M5 Ultra) | 64 GB (X Elite) |
| Native AI Format | FP4 (Tensor Cores Gen 5) | FP16/INT8 | INT8 |
| CPU-GPU Connectivity | NVLink-C2C | Native unified architecture | Traditional bus |
| Target OS | Windows on Arm | macOS | Windows on Arm |
| Availability | Q3 2026 (estimated) | Early 2026 | Already available |
| Target TDP | 110W (Surface Ultra) | 30-120W (depending on model) | 45-80W |
The RTX Spark does not beat the Apple M5 across the board. The Apple M5 Ultra in a Mac Studio will likely remain more efficient in watts per TFLOP for pure creative workloads. But for local AI, the RTX Spark has a decisive advantage: Tensor Cores in FP4 and the CUDA ecosystem.
AI developers don't choose a chip for its raw specs. They choose an ecosystem. And CUDA remains the most mature AI ecosystem in the world. Tom's Guide sums it up well: the RTX Spark is "the superchip that will change laptops forever", precisely because it brings CUDA to a unified ARM architecture.
Against Qualcomm, the match is more expeditive. The Snapdragon X Elite is an excellent chip for daily productivity, but its 64 GB max memory and lack of a high-performance GPU place it in a different category. The RTX Spark is a superchip; the X Elite is a highly successful mobile SoC. Not in the same league for AI.
DLSS 4.5 and gaming on RTX Spark
The NVIDIA Computex 2026 announcement is not limited to RTX Spark. DLSS 4.5 is announced simultaneously, and the two are linked.
DLSS 4.5 brings new frame reconstruction and pixel generation improvements, directly leveraging the 5th-generation Tensor Cores of the Blackwell RTX. On an RTX Spark, this means that native resolution gaming is no longer necessary — the GPU can render at 1080p and reconstruct to 4K with indistinguishable quality, freeing up resources for other tasks.
The implication is fascinating: an RTX Spark laptop could game at full throttle while running an AI model in the background for dynamic NPCs, real-time voice translation, or gameplay assistance. FP4 allows Tensor power to be shared between DLSS and generative AI without one cannibalizing the other.
This is the vision of "agentic gaming" that NVIDIA was already outlining during the announcements around Vera Rubin and the era of AI agents. The RTX Spark is the first consumer hardware (if one can call 110W workstations "consumer") capable of making it a reality.
NVIDIA's domination strategy: $40 billion and a locked-down ecosystem
The RTX Spark is not an isolated product. It is part of a massive strategy that Jensen Huang is deploying across the entire company. As analyzed in our article on the $40 billion NVIDIA is investing in AI in 2026, every move by NVIDIA serves a coherent plan: to be present at every layer of the AI stack.
In the datacenter: Grace Hopper, Blackwell B200, Vera Rubin. In the cloud: GPU instances across all major providers. At the edge and on the PC: RTX Spark. In models: Nemotron, NIM. Each product reinforces the others.
The RTX Spark locks the Windows on Arm ecosystem into CUDA. A developer who optimizes their models for the RTX Spark uses TensorRT, CUDA, and NVIDIA libraries. If they want to deploy in prod, they move to NVIDIA GPUs in the datacenter. If they want to distribute, they use NIM. The cycle is virtuous — for NVIDIA.
It's exactly the same strategy as Apple with its chip + macOS + App Store, except that NVIDIA is playing it on the open ground of Windows and open source. More subtle, potentially more effective.
❌ Common mistakes
Mistake 1: Confusing RTX Spark and N1X
Tom's Guide points out right in its title: the RTX Spark "is not called N1X". The N1X rumor had been circulating for months before Computex, but the final commercial name is RTX Spark. N1X might refer to the platform or a server variant, but for PCs, it's RTX Spark. Do not mix up the two in an article or a technical presentation.
Mistake 2: Believing that RTX Spark replaces discrete GPUs
The RTX Spark is an integrated superchip. It does not replace an RTX 5090 for pure gaming or professional 3D rendering. The 6,144 CUDA cores are powerful, but a high-end discrete GPU will have more cores, more dedicated memory bandwidth, and a much higher TDP. The RTX Spark excels in local AI thanks to unified memory, not in brute-force rendering.
Mistake 3: Thinking that everything runs natively on Windows on Arm
Software compatibility is the Achilles' heel of WoA. If your AI toolchain is in Python/PyTorch with CUDA, you will probably be fine — NVIDIA has had months to optimize. But if you rely on tools compiled in x86, specific drivers, or niche software, check compatibility before buying. Windows' x86 emulation works, but it adds a non-negligible overhead.
Mistake 4: Underestimating the 110W TDP
110W is the thermal budget of the Surface Laptop Ultra. This means the machine will get hot, the fan will spin up under load, and battery life will not rival a 30W MacBook Air M5. It's an accepted compromise: power versus mobility. Don't expect 15 hours of battery life during heavy AI inference.
❓ Frequently Asked Questions
Can the RTX Spark really run a 120B model locally?
Yes, thanks to the 128 GB of unified memory and the FP4 format of the 5th-generation Tensor Cores. A 120B model in FP4 takes up about 60-70 GB, which fits in the shared memory pool. The inference speed will be moderate but usable for agentic tasks.
When will the first RTX Spark PCs be available?
Announcements took place on May 31, 2026, at Computex. The first devices, including the Surface Laptop Ultra, are expected in Q3 2026. Workstation laptops and mini desktops will arrive first, with the general public likely following in 2027.
Is the RTX Spark compatible with existing x86 software?
Windows on Arm includes x86 emulation that works for most applications. However, for optimal performance — especially in AI — native ARM binaries are required. NVIDIA and Microsoft are pushing developers to compile for native ARM via their toolchains.
How does the RTX Spark compare to the Nemotron 3 Ultra 550B?
These are two different things. Nemotron 3 Ultra 550B is a language model. The RTX Spark is the hardware designed to run this type of model. With 128 GB, the RTX Spark can run quantized versions of Nemotron, but the full model in normal precision will still require a server.
Should you wait for the RTX Spark to buy an AI PC?
If you have an urgent need for local AI with more than 24 GB of VRAM, the RTX Spark is the first laptop to offer this capability via unified memory. If you can wait, it will be wise to look at the first real-world benchmarks and the software maturity of Windows on Arm before taking the plunge.
✅ Conclusion
The RTX Spark is the breakthrough that Windows on Arm has been waiting for for three years: a chip designed by the undisputed leader in AI computing, with 128 GB of unified memory and FP4 Tensor Cores that make running 120B parameter local AI models practical in a laptop. Apple's M5 has a serious competitor in the unified memory arena, and this time, it speaks CUDA.
If you are developing AI agents and want to move out of the cloud without sacrificing power, the RTX Spark deserves your full attention upon its expected release in Q3 2026.