NVIDIA unveiled RTX Spark at Computex 2026 on May 31, an Arm-based superchip co-developed with MediaTek that combines a 20-core Grace CPU, a Blackwell RTX GPU with 6,144 CUDA cores (CUDA: Compute Unified Device Architecture, NVIDIA’s GPU programming platform), and up to 128GB of unified LPDDR5X memory on TSMC’s 3-nanometer process. The chip targets 1 petaflop of FP4 (floating-point 4-bit precision) AI performance and is designed to run 120-billion-parameter language models entirely on-device. Systems from ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI are set to ship this fall, with Acer and Gigabyte following shortly after.
The 20 cores use Arm’s Cortex-X925 performance cores and Cortex-A725 efficiency cores, a combination MediaTek engineered. Qualcomm’s Snapdragon X2 ships Oryon V2 cores, and Apple’s current MacBook Pro runs M4, both microarchitectures that benchmark observers place ahead of Cortex-X925 in single-threaded throughput. NVIDIA has not published CPU performance results for RTX Spark; at a Computex press Q&A, the company confirmed it is leaving those numbers to OEMs (original equipment manufacturers) and the press once devices ship.
A 70-Billion-Transistor Superchip
The chip, internally codenamed N1X, fuses two chiplets over NVIDIA’s NVLink-C2C (chip-to-chip) interconnect, linking the GPU and CPU dies with 600 GB/s of chip-to-chip bandwidth, roughly five times PCIe Gen 5 at substantially lower power, per NVIDIA’s official RTX Spark announcement. Both chiplets share the same LPDDR5X memory pool at 300 GB/s, so the GPU accesses the same data as the CPU without crossing a discrete memory bus. The full package is a 70-billion-transistor design fabricated on TSMC’s 3nm process.
NVIDIA was plain at the press Q&A about where this design came from. The full-spec chip is “based on the same system” as the GB10 superchip inside the DGX Spark desktop AI workstation, which sells for $3,999, with adjustments for laptop thermals and power budgets. That origin gives the GPU architecture a shipping history: the same Blackwell silicon has been running in DGX Spark units that developer customers have used since the workstation launched.
OEM thermal engineering will define the sustained real-world performance of any N1X system. NVIDIA confirmed a TDP (Thermal Design Power) of 45-80W for the flagship N1X configuration, with each manufacturer selecting its own cooling approach. A vapor-chamber design in a 14mm chassis will sustain different throughput than a thicker notebook with more copper and airflow running the same chip. StorageReview flagged this variation in its Computex coverage, noting that the OEM that builds the better thermal solution could open a real sustained performance gap over a thinner competitor on identical silicon.
The CPU Cores Behind the Petaflop
MediaTek co-designed the CPU chiplet using Arm’s Cortex-X925 performance cores and Cortex-A725 efficiency cores, 10 of each in the flagship N1X. Tom’s Hardware’s Computex roadmap coverage noted these are “client cores,” newer than the Neoverse V2 cores in NVIDIA’s server-class Grace CPUs but a generation behind the Oryon V2 microarchitecture in Qualcomm’s current Snapdragon X line and the M4 in Apple’s MacBook Pro. Per HotHardware, the N1X CPU cluster peaks at 4.1 GHz.
That comparison sharpens for workloads outside NVIDIA’s GPU sweet spot: browser-heavy multi-tab sessions, compiled build times, and office productivity software lean on single-threaded CPU performance. For a developer running inference on a 70-billion-parameter model through CUDA and TensorRT (NVIDIA’s inference optimization library), the Blackwell GPU and unified memory pool dominate and the CPU clock matters far less. MediaTek’s RTX Spark CPU chiplet design collaboration covered the CPU architecture, memory subsystem, power management, and connectivity, making MediaTek the exclusive CPU architecture owner for this first generation.
Also embedded in the CPU chiplet: an NPU (neural processing unit) based on NVIDIA’s Deep Learning Accelerator, designed for lighter inference tasks without engaging the full Blackwell GPU. The NPU also satisfies the dedicated AI acceleration requirement for Windows Copilot+ PC certification.
No independently verified CPU benchmark for RTX Spark existed at this article’s publication. No IPC (Instructions Per Clock) figures, SPECint scores, or Geekbench results have been shared publicly. The first numbers will arrive when OEMs hand retail hardware to reviewers.
Windows Does the Heavy Lifting
Microsoft rebuilt parts of Windows 11’s task scheduler specifically for RTX Spark’s heterogeneous layout, implementing what Microsoft calls workload profile scheduling (WPS), which manages real-time decisions about which tasks land on performance CPU cores, efficiency cores, the Blackwell GPU, or the NPU when they compete for bandwidth simultaneously. Poor scheduling on a chip running CUDA workloads alongside translated x86 apps and a background AI agent can produce stalls that look like mysterious slowdowns rather than obvious failures. Microsoft published the technical details in its Windows Experience Blog post on RTX Spark integration on May 31. Prism, Microsoft’s x86-to-Arm64 translation layer, was also tuned for the N1X microarchitecture. The blog states:
Prism has been tuned for the microarchitecture of RTX Spark and when combined with the raw power of the silicon, unlocks great performance for developers, creators and gaming workloads running under emulation.
Microsoft’s Windows Experience Blog, May 31, 2026. Prism converts legacy x86 instructions to Arm64 at runtime; support for AVX (Advanced Vector Extensions) and AVX2, added last year, lets apps like Ableton Live that previously couldn’t run under translation install and work on Arm hardware.
Applications shipping natively for Windows on Arm at the RTX Spark launch include:
- Creative tools: Blender, DaVinci Resolve, Cinema4D, Redshift, Topaz Photo, CapCut, Affinity apps by Canva
- Developer tools: GitHub Copilot, Claude Code, ComfyUI, Cursor
- Gaming titles confirmed: League of Legends, VALORANT (Riot Games), PUBG: Battlegrounds, Alan Wake 2, War Thunder
- Adobe Premiere Pro and Photoshop, running natively on Arm with additional RTX Spark-specific optimizations in development
That list doesn’t cover every enterprise edge case. Driver-dependent legacy software, anti-cheat systems beyond Epic’s Easy Anti-Cheat and BattlEye (both confirmed compatible), and line-of-business software with kernel-mode dependencies can still fail under translation. Microsoft hasn’t published a compatibility percentage, and the practical test comes when IT departments run the specific stacks that govern their purchasing decisions.
Eight Partners, No Price
NVIDIA confirmed six OEMs for launch: ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI. Acer and Gigabyte follow. Jensen Huang said to expect over 30 laptop designs and roughly 10 desktop configurations in the opening wave. Form factors span 14 to 16 inches, with the most aggressive designs targeting 14mm chassis profiles.
Two chip configurations are confirmed for the launch window, per MediaTek’s Computex disclosures:
| Configuration | CPU | GPU | Max Memory | TDP |
|---|---|---|---|---|
| RTX Spark (N1X) | 20 cores: 10x Cortex-X925 + 10x A725 | 6,144 CUDA cores | 128GB LPDDR5X | 45-80W |
| RTX Spark (N1) | 12 cores: 8x Cortex-X925 + 4x A725 | 2,560 CUDA cores | 64GB LPDDR5X | 18-45W |
An 18-core variant, described by HotHardware as binned silicon from the 20-core N1X die, is also expected but hasn’t been officially confirmed by NVIDIA or MediaTek as of publication.
Microsoft’s Surface Laptop Ultra is the most detailed device announced. It carries a 15-inch mini-LED PixelSense Ultra display peaking at 2,000 nits, a 3:2 aspect ratio, HDMI, USB-A, a full-size SD card slot, and the largest trackpad Microsoft has ever built. Under 4.5 pounds and less than 18mm thick, Microsoft describes it as the thinnest and most powerful Surface laptop yet. ASUS’s ProArt P16 goes further on display: an OLED panel up to 4K at 120Hz with NVIDIA G-SYNC and a 99.9 Wh battery.
NVIDIA has not announced prices. Morgan Stanley, in a note cited by TechTimes, suggested N1 configurations near $1,799 and N1X machines around $2,899, though NVIDIA hasn’t confirmed either figure. HotHardware flagged that ongoing DRAM and NAND shortages add cost pressure that could push street prices toward the upper bound of analyst estimates.
The Field Nvidia Just Entered
NVIDIA’s dominance in discrete graphics and data-center AI didn’t hand it a ready-made lead in the laptop SoC market. That category already had a chip pairing large unified memory with a powerful integrated GPU: AMD’s Ryzen AI Max+ 395, codenamed Strix Halo.
Our earlier coverage of RTX Spark against Strix Halo reports that AMD’s chip had already landed in 35 consumer products by the time Jensen Huang took the Computex stage, shipping since January 2025. At a press roundtable that day, AMD VP Rahul Tikoo told HP VP Jim Nottingham: “We have 35 products with Strix Halo in market. Welcome, Nvidia, to the modern compute journey.” The Ryzen AI Max+ 395 packs 16 Zen 5 CPU cores and a 40-compute-unit RDNA 3.5 GPU with a large unified memory pool, and its x86 architecture means every Windows application runs natively, without any translation layer.
NVIDIA’s pitch to OEMs and enterprise buyers rests on the CUDA software stack, which no competitor can match directly. CUDA has been the default platform for AI model development for over a decade; TensorRT handles inference optimization, and most major AI frameworks treat NVIDIA hardware as their reference implementation. AMD’s ROCm (Radeon Open Compute) has narrowed the gap significantly in recent years, but the developer mindshare CUDA carries into RTX Spark’s platform is a structural advantage the GPU core count alone doesn’t capture.
Current benchmark comparisons place Apple’s M4 ahead in single-threaded CPU throughput, and its unified memory architecture has had four years of software optimization on macOS. Qualcomm debuted its Snapdragon X2 Elite in a mini-PC at Computex; its Oryon V2 CPU architecture benchmarks ahead of the N1X’s MediaTek-designed cores in per-core throughput from existing Snapdragon X laptop reviews.
Will Three Generations Be Enough?
The easiest path to RTX Spark’s failure is the one every previous Windows on Arm chip faced: OEMs and developers treat it as a one-generation experiment, investment thins, software support lags, and the platform never reaches the scale needed for self-sustaining adoption. Jensen Huang addressed that risk at Computex by publicly committing to three hardware generations of the Spark platform.
After the Grace Blackwell N1X shipping this fall, a Vera Rubin generation follows in 2027 on LPDDR6 memory, and Rosa Feynman comes after that. The cadence mirrors NVIDIA’s data-center roadmap, which has shipped consistently for years. MediaTek is the exclusive CPU architecture partner for the first generation, embedded in the silicon design from the ground up rather than supplied as a reference design, with the partnership confirmed to carry through at least Vera Rubin.
Every OEM that tooled a thin chassis around a 45-80W laptop SoC needs more than one product cycle to recover that investment. The Surface Laptop Ultra, ASUS’s ProArt P16, and HP’s OmniBook all reflect specialized mechanical engineering around the N1X’s thermal envelope, and designing systems that thin at that power level demands supply-chain commitments that typically span two or three product cycles. ISVs like Adobe, which is rearchitecting Photoshop and Premiere for the platform, face the same calculus. Sustained performance benchmarks on retail hardware won’t exist until devices ship this fall.
