NVIDIA RTX Spark review: Windows on Arm gets serious about graphics and local AI

Windows on Arm has spent years trying to answer a fairly straightforward question: can a laptop move away from x86 without making familiar software feel like a compromise? Well, Qualcomm certainly has put in a lot of efforts into making this a reality with their Snapdragon X and X2 platform. NVIDIA’s RTX Spark adds another question. Can that same laptop also deliver the graphics capabilities, creative application acceleration and local AI capacity associated with an RTX machine?

✅ Thank you for completing the survey!

Looks like we have an answer, even though it might be an early one. The ASUS ProArt P16 H7607BA is part of a new wave of devices that’s poised to bring a sea change in the PC ecosystem. Its RTX Spark N1X combines a 20-core Arm CPU, Blackwell RTX graphics with 6,144 CUDA cores, and 64 GB of unified memory in the configuration assessed here. It runs Windows 11 Pro, supports NVIDIA’s CUDA ecosystem and brings the familiar RTX feature set into a different processor and memory architecture. If all this seems familiar, then yes, it’s pretty much everything that was present in one of NVIDIA’s DGX Spark put together for Windows. Just that the DGX Spark had a much higher TDP whereas the RTX Spark machines are mostly limited to 80 Watts.

The combination of an NVIDIA CPU and GPU isn’t just about slapping on another AI badge on a laptop. Conventional creator machines usually divide their resources between system RAM and a separate graphics memory pool. RTX Spark offers an alternative in which the CPU and GPU can access a large shared memory capacity. Similar to what Apple does with their unified memory approach for the Macbooks and the Mac Minis. For local models, complex scenes and certain creative projects, fitting the workload into memory can be as consequential as the speed of the processor executing it.

The results suggest that NVIDIA has built a credible high-performance Windows platform. They also show why its strongest argument is the combination of capabilities. RTX Spark does not lead every CPU benchmark, and a large memory pool does not automatically make it faster than a powerful discrete GPU. Its appeal lies in bringing useful CPU performance, RTX graphics and substantial local AI capacity together in a portable system.

From Computex to a broader Windows platform

ASUS, like many other OEMs, announced its RTX Spark ProArt systems at Computex on 1 June 2026, including the P16, P14 and a compact desktop. Subsequent announcements have expanded on the software and Windows integration behind those machines. NVIDIA positions RTX Spark around creators, gamers and personal AI agents. Each audience places different demands on a PC. Creators need reliable acceleration in applications and enough memory for their projects. Gamers need compatibility, frame rates and responsive input. Local AI users need supported runtimes, model capacity and acceptable response times.

RTX Spark attempts to address all three through a common foundation: an Arm CPU, a Blackwell RTX GPU, unified memory and NVIDIA’s software stack. The proposition extends beyond laptops to compact desktops, although the ProArt P16 remains the focus of today’s review here.

Its rivals consequently come from several directions. Qualcomm’s Snapdragon X2 Elite Extreme features a similar Windows on Arm system for comparison. Intel’s Core Ultra 9 290HX Plus and AMD’s Ryzen 9 9955HX represent powerful x86 laptop processors. AMD’s Ryzen AI Max+ 395 is particularly relevant because it also combines substantial integrated graphics capability with a large shared memory architecture. We don’t see the conventional Intel or AMD laptops with GeForce graphics getting dethroned anytime soon. However, it goes without saying that the two incumbent x86 companies have more to worry about with NVIDIA throwing its hat into the ring.

RTX Spark isn’t the same as the DGX Spark. Sharing parts of NVIDIA’s accelerated-computing ecosystem does not make them interchangeable products. The DGX Spark lives and beatles in the familiar Linux space where most development takes place, whereas with RTX Spark, the proposition is a Windows PC intended to run everyday applications, games and creative software alongside local AI.

The N1X configuration in this ProArt uses ten Arm Cortex-X925 performance cores and ten Cortex-A725 efficiency cores. NVIDIA uses the Grace CPU branding and developed the CPU, I/O and memory tile with MediaTek. The core arrangement is distinct from both Qualcomm’s Oryon implementation and the x86 designs from Intel and AMD. The performance cores handle demanding work, while the efficiency cores provide additional throughput and a lower-power option for suitable tasks. Actual behaviour depends on scheduling, power management and the application. Twenty cores do not translate directly into a predictable advantage over an 18-core Qualcomm processor or a 16-core AMD CPU with simultaneous multithreading.

NVIDIA specifies maximum performance-core clocks of up to 4.0–4.1 GHz, with efficiency cores reaching 2.8 GHz. On the graphics side, the larger N1X configuration contains 48 streaming multiprocessors, packing in 6,144 CUDA cores, with fourth-generation RT cores and fifth-generation Tensor cores. Like always, CUDA cores execute general graphics and compute work, RT cores accelerate ray-tracing operations, and Tensor cores handle supported AI calculations. 

The Tensor hardware matters to both local AI and gaming. It supports the lower-precision calculations used by compatible models, as well as neural rendering technologies. NVIDIA advertises up to 1 petaflop of FP4 AI performance. However, that 1 petaflop figure should not be read as directly comparable with an NPU’s INT8 TOPS rating or as a prediction of language-model output speed.

Unified memory changes which workloads fit

The CPU and GPU tiles communicate through NVLink-C2C. NVIDIA specifies 600 GB/s of bidirectional interconnect bandwidth, while the LPDDR5X memory system provides up to 300 GB/s of peak memory bandwidth. These are different links in the architecture. The larger figure is not the GPU’s memory bandwidth.

The platform supports up to 128 GB of unified memory, although this ASUS ProArt P16 test platform laptop has 64 GB. The GPU allocation was set to 32 GB. That is an allocation within the shared physical memory pool, not an additional 32 GB of dedicated VRAM. You can check and set the allocation via the NVIDIA App in a rather convenient manner. The benefit is straightforward. A workload that exceeds the graphics memory of a conventional laptop may fit into the GPU-accessible allocation here. A local language model also needs room for its context cache and runtime buffers, while Windows and other applications still need memory of their own.

Capacity and speed remain separate considerations. Dedicated laptop GPUs can have different memory bandwidth, power budgets and compute resources. Sharing memory also means that CPU and GPU activity can compete for resources. RTX Spark’s architectural advantage is therefore most persuasive when a workload benefits from both the available capacity and NVIDIA’s acceleration ecosystem.

NVIDIA lists a configurable superchip power range of 45–80 W. And like always, OEMs can tweak it to different power limits. We saw configurations from ASUS, HP, MSI, and others with varying coolling assemblies designed to handle different TDPs of the same RTX Spark chip.

ASUS ProArt P16 H7607BA: the RTX Spark test device

The P16 places the platform in a creator-oriented 16-inch laptop. The particular model configuration combines the larger N1X processor with 64 GB of memory, a 1 TB SSD and a 2880 × 1800 OLED touchscreen.

Specification	ASUS ProArt P16 H7607BA
Platform	NVIDIA RTX Spark N1X
CPU	20 cores: ten Cortex-X925 and ten Cortex-A725
Graphics	Blackwell RTX, 6,144 CUDA cores
Memory	64 GB unified memory
GPU memory allocation	32 GB
Storage	1 TB SSD
Display	16-inch OLED, 2880 × 1800, 16:10, approximately 212 ppi
Refresh rate	120 Hz
Operating system	Windows 11 Pro 26H1, build 28000.3086
NVIDIA driver	617.42
Installed local model	Qwen3.8-27B-UD-Q4_K_M, with Hermes
India price	Not available. Estimated at Rs 1,99,999.

ASUS packs the laptops with a 99.9Wh battery and a 140W USB-C adapter, all within 1.77 kg of weight. Connectivity includes three USB4 Type-C ports, USB-A, HDMI 2.1, an SD Express card reader and an audio jack. It also comes with Wi-Fi 7, Bluetooth 5.4, a haptic touchpad, a backlit keyboard and an infrared camera for Windows Hello. The port selection is well suited to a creator machine on paper. A card reader, external-display connection and several high-speed USB ports reduce dependence on a dock. Onboard memory makes the initial capacity choice consequential, particularly when local AI is part of the intended workload.

The 16:10 display gives more vertical workspace than a conventional 16:9 panel. At approximately 212 ppi, its pixel density balances a detailed image with a lower pixel count than the optional higher-resolution panel. A 120 Hz refresh rate also plays into the interests of folks who like a bit of high-refresh rate gaming. 

ASUS also says that this unit comes with 100% DCI-P3 coverage and up to 1,100 nits HDR peak brightness for the 3K display option. These remain panel specifications rather than measured colour-accuracy or HDR findings. We calibrated the brightness to 200 nits for battery testing. We’ll update the review with more panel measurements. Acoustically, the test device was remarkably quiet in use. Consistent loads did get the fan humming like a jet engine during certain agentic LLM workloads and gaming. 

Windows on Arm: the software path still matters

RTX Spark runs native Arm applications and existing x86/x64 applications through Windows’ Prism translation technology. That creates a broad route to software compatibility, but application architecture remains relevant to performance and reliability. A native ARM64 application avoids translating its CPU instructions. An x64 application uses Prism for that portion of execution. Supported GPU kernels still execute on the GPU, which is why a graphics-heavy application can perform differently from a CPU-heavy program under emulation.

ARM64EC adds another possibility, combining native Arm components with compatible x64 components. A native application label therefore does not necessarily establish that every plug-in or dependency follows the same path. NVIDIA and Microsoft have been working on scheduling, thermal management, GPU-accessible memory and Windows ML integration for RTX Spark. These are important supporting changes because the architecture asks Windows to manage heterogeneous CPU cores, a substantial integrated GPU and shared memory under one power envelope.

CPU performance

The P16’s measured CPU results establish a capable foundation. Cinebench 2026 returned 557 in the single-core test and 7,117 in multicore. Geekbench 7.1.0 produced 2,699 single-core and 24,155 multicore.

Benchmark	Measured P16 result
Cinebench 2026 single-core	557
Cinebench 2026 multicore	7,117
Geekbench 7.1.0 single-core	2,699
Geekbench 7.1.0 multicore	24,155

Cinebench 2026 provides a consistent CPU rendering workload for comparing RTX Spark with competing Intel, AMD and Qualcomm processors. 

Processor	Cinebench 2026 single-thread	Cinebench 2026 multicore
NVIDIA RTX Spark N1X, tested P16	557	7,117
Qualcomm Snapdragon X2 Elite Extreme X2E-96-100	668	7,412
Intel Core Ultra 9 290HX Plus	569	8,592
AMD Ryzen 9 9955HX	527	8,018
AMD Ryzen AI Max+ 395	467	7,268

N1X finishes close to the Core Ultra 9 290HX Plus and ahead of the Ryzen 9 9955HX in single-thread rendering. Qualcomm’s X2 Elite Extreme is faster, scoring 668 against the P16’s 557. In multicore rendering, the P16 finishes approximately 4% behind the X2 Elite Extreme and 2% behind the Ryzen AI Max+ 395 average. N1X is competitive with both, although laptop cooling and power limits can influence margins this small. Intel’s 290HX Plus and AMD’s 9955HX retain a clearer multicore advantage. The P16 trails their listed results by approximately 17% and 11%, respectively. For work dominated by CPU rendering, a powerful HX laptop remains a relevant alternative.

AMD’s Ryzen AI Max+ 395 is a close rival, combining an x86 CPU, Radeon graphics and shared memory. N1X delivers higher single-thread performance in Cinebench and finishes just behind it in multicore. CUDA support gives RTX Spark a different route to accelerating rendering and local AI, while AMD retains native compatibility with established x86 software. In Geekbench 7.1.0, N1X scored 2,699 single-core and 24,155 multicore. Its 20-core design has ample parallel processing capability, although the Cinebench results leave Intel’s and AMD’s HX processors ahead for CPU rendering.

Speedometer 3.1

The ASUS ProArt P16 scored 22.8 in Firefox and 36.1 in Microsoft Edge in Speedometer 3.1.

Browser	Speedometer 3.1 score
Firefox	22.8
Microsoft Edge	36.1

Edge scored approximately 58% higher than Firefox on the same P16. For browser-based work, that makes Edge the stronger performer in this test. Speedometer measures interactions with simulated web applications, including updates to page content and interface elements. The gap points to a substantial difference in how the two browsers handle those tasks on RTX Spark, rather than a uniform improvement in page loading or every website. Compared to the competition, the AMD Max+ 395 typically scores around 37 and the Intel Core Ultra 9 290HX Plus usually hits 39.4. 

Graphics performance: RTX Spark reaches discrete-laptop territory

When it comes to synthetic benchmarks, the NVIDIA RTX Spark doesn’t hold back any punches. It’s been long described to be nearly similar as an RTX 5070. 

3DMark benchmark	RTX Spark	FPS
Port Royal	9,220	42.69 FPS
Solar Bay	55,882	212.48 FPS
Solar Bay Extreme	11,428	79.92 FPS
Speed Way	3,038	30.39 FPS
Steel Nomad	3,226	32.26 FPS
Time Spy	13343	NA

In Steel Nomad, average scores place the RTX 5060 Laptop at 2,582, the RTX 5070 Laptop at 2,944 and the Radeon 8060S at 2,338.

Graphics platform	Steel Nomad score
RTX Spark N1X, ProArt P16	2,991
GeForce RTX 5070 Laptop	2,944
GeForce RTX 5060 Laptop	2,582
Radeon 8060S	2,338

The Steel Nomad score puts RTX Spark close to the RTX 5070 Laptop average in this workload. That is a meaningful level of graphics performance for a platform that also offers a large unified memory pool. Speed Way gives a slightly different ordering. RTX Spark’s 3,038 compares with the 2,806 score for the RTX 5060 Laptop and 3,403 for the RTX 5070 Laptop. RTX Spark sits between those two averages. RTX Spark consequently sits around the RTX 5060 to RTX 5070 Laptop class across these two workloads. Its shared memory capacity adds room for larger GPU workloads, but these scores alone reveal little about power consumption or battery performance.

Gaming performance

The gaming results cover demanding titles at 1440p. Most runs use DLSS Super Resolution and 3× frame generation at 1440p, with ray tracing enabled in several titles.

Game	Average frame rate
Cyberpunk 2077	138 FPS
DOOM: The Dark Ages	141 FPS
Gears of War: E-Day	135 FPS

At these settings, RTX Spark delivers triple-digit displayed frame rates in demanding games at 1440p. DLSS upscaling and frame generation are central to that performance, alongside reduced quality settings in several titles.

Super Resolution reconstructs the output from a lower internal resolution, reducing the conventional rendering workload. Frame generation inserts additional displayed frames, while Ray Reconstruction replaces supported denoising paths. Each serves a different purpose.

Reflex can reduce queuing latency in supported games, but the displayed frame rate alone does not establish responsiveness. Base rendering speed, frame pacing and input latency remain important, particularly in competitive games. The average frame rates alone do not reveal those characteristics. On the P16, we ran all the games at 1920×1080 to compare with other configurations.

Anti-cheat support remains game-specific

Anti-cheat remains a potential obstacle for competitive gaming. Support varies by game, even when the anti-cheat provider supports Windows on Arm. Its listed available examples include games using Easy Anti-Cheat, BattlEye and Valve Anti-Cheat. The same material places titles associated with RICOCHET and Vanguard in a later-support category. Availability can vary by title even where the underlying provider supports the platform. For a mixed creator and gaming laptop, the gaming performance is encouraging. For a purchase centred on one competitive title, confirmed support for that particular game remains essential.

DLSS 5 adds a different kind of rendering

DLSS 5’s 3D-guided neural rendering is distinct from resolution reconstruction and frame generation. It uses rendered information and motion vectors to alter lighting and material appearance within a supported game’s integration.

NBA 2K27 is the relevant early example for RTX Spark. NVIDIA’s October update claims roughly twice the DLSS 5 model execution speed, a 24% reduction in model size to 118 million parameters, and 20% lower 4K memory usage. NVIDIA reports a game performance improvement of more than 30% over the September implementation in NBA 2K27. The feature is relevant to a power-constrained laptop because its own inference cost consumes part of the graphics budget. Reducing that cost can make the visual treatment more practical. DLSS 5 also changes the appearance of supported games, making character detail, lighting consistency and temporal stability central to its appeal. 

AI performance

The RTX Spark N1X generated text locally at between 27.85 and 73.75 tokens per second across Procyon’s four language models. Phi 3.5 delivered the highest output rate, while Mistral 7B and Llama 3.1 both exceeded 50 tokens per second. Average time to first token remained below one second across all four models.

Procyon AI Text Generation	Score	Average time to first token	Average output speed
Phi 3.5	2,749	0.25 seconds	73.75 tokens/s
Mistral 7B	2,352	0.44 seconds	52.97 tokens/s
Llama 3.1	2,277	0.39 seconds	50.35 tokens/s
Llama 2	2,162	0.81 seconds	27.85 tokens/s

These speeds give local writing and coding assistants enough throughput to produce sustained responses, with a short initial wait. Longer prompts place more demand on processing before generation begins, so responsiveness also depends on the amount of material passed to the model.

MLPerf’s results show how that behaviour changes across inference engines. With Llama 3.1 8B Instruct, Windows ML delivered slightly higher overall throughput at 56.7 tokens per second, against llama.cpp’s 55.0 tokens per second. However, llama.cpp started responding sooner, with a time to first token of 0.27 seconds against 0.34 seconds.

MLPerf workload	Windows ML output speed	llama.cpp output speed	Windows ML time to first token	llama.cpp time to first token
Llama 3.1 8B Instruct	56.7 tokens/s	55.0 tokens/s	0.34 seconds	0.27 seconds
Phi 4 Mini Instruct	103.3 tokens/s	95.7 tokens/s	0.17 seconds	0.12 seconds

Phi 4 Mini Instruct produced faster responses through both engines, exceeding 100 tokens per second through Windows ML. Here too, llama.cpp returned the first token sooner, while Windows ML maintained the higher overall generation rate.

The difference became more apparent with longer summarisation workloads. Llama 3.1 took 1.24 seconds to begin responding through Windows ML and 1.03 seconds through llama.cpp, with output rates of 51.7 and 52.0 tokens per second respectively.

MLPerf SWE Agent, Llama 3.1 8B Instruct	Windows ML	llama.cpp
Average expected duration	55.195 seconds	46.025 seconds
Average tools duration	6.363 seconds	6.750 seconds

In the SWE Agent workload, llama.cpp recorded a lower average expected duration, although its tools duration was slightly higher. The choice of inference engine therefore affects more than raw token throughput, particularly when an agent combines generation with tool use.

Computer vision performance

Procyon’s AI Computer Vision 2.0 benchmark returned an overall score of 3,948, using the Blackwell GPU with TensorRT at FP16 precision.

Workload	Measured result
ConvNeXt throughput	1,959.43 images/s
BLIP encoder	355.755 iterations/s
BLIP decoder	348.035 iterations/s
SAM2.1 encoder	60.196 iterations/s
SAM2.1 decoder	795.751 iterations/s
DETR	233.629 iterations/s
ESRGAN	16.875 iterations/s

The results cover image classification, image-and-text processing, segmentation, object detection and upscaling. Alongside the language-model tests, they demonstrate the GPU’s ability to handle several distinct local AI workloads.

Blender: standard Arm results and a separate accelerated reference

The Blender benchmark used the latest alpha release to ensure compatibility with the new platform.

Scene	Measured samples per minute
Monster	181.758
Junkshop	121.858
Classroom	90.160

The ASUS ProArt P16 scores 181.758 in Monster, 121.858 in Junkshop and 90.160 in Classroom with accelerated rendering. Blender’s native Arm execution and supported NVIDIA rendering acceleration give RTX Spark several processing options. The unverified backend in the P16 run leaves its accelerated rendering performance unresolved. Blender 5.3 is due to add DLSS 4.5 Ray Reconstruction to the Cycles viewport. Cleaner interactive previews could make lighting and material adjustments easier, extending RTX acceleration beyond the final render.

Creative applications

Native Arm creative applications and preview integrations bring CUDA, TensorRT and RTX acceleration to editing, image processing and rendering. In Premiere Pro, these include Object Masking, Enhance Audio, Scene Edit Detection, Auto Reframe and a GPU-accelerated colour workflow. Photoshop’s preview Remove Tool can use a local model. DaVinci Resolve provides another important route to native Arm creative work with supported GPU acceleration.

Local processing has a practical benefit for interviews, unreleased footage and confidential assets: an operation can run without sending its content to a remote inference service. That benefit applies to the supported local operation, rather than automatically extending to every feature in the application.

For a creator considering the platform, the more consequential question is whether the exact project, effects and plug-ins are supported. A native editor with an unsupported specialist plug-in can still interrupt an established workflow. Preview availability improves the platform’s immediate reach, but it also makes application-version checking part of adoption.

Battery life

The ProArt P16 consumed seven percentage points of battery charge during one hour of UL Procyon Video Playback. The test began at 98% and ended at 91%, with the screen calibrated to 200 nits using the SpyderPro 5 and refresh rate set to 120Hz.

Battery test condition	Recorded value
Workload	UL Procyon Video Playback
Duration	1.0 hour
Display brightness	200 nits, instrument calibrated
Refresh rate	120Hz
Starting charge	98%
Ending charge	91%
Observed consumption	Seven percentage points

That is an encouraging observation for a machine carrying substantial graphics capability. It is also a video-playback observation rather than a complete endurance result. Hardware video decoding represents a different workload from CPU rendering, gaming or continuous local inference.

A straight-line extrapolation from seven percentage points per hour produces approximately 14.3 hours from a full charge. In an actual run, we ended up getting the laptop to run for 13.4 hours. 

Where RTX Spark stands against Intel, AMD and Qualcomm

Qualcomm’s X2 Elite Extreme is faster in Cinebench single-thread rendering and narrowly ahead in multicore. RTX Spark adds Blackwell graphics, CUDA and a substantial GPU-accessible memory allocation, giving it greater appeal for supported GPU-heavy creative and AI workloads.

Against Intel’s Core Ultra 9 290HX Plus and AMD’s Ryzen 9 9955HX, N1X delivers credible CPU performance but trails the listed multicore rendering results. An HX machine paired with a powerful discrete GPU remains a strong option for maximum throughput and established x86 workflows.

Ryzen AI Max+ 395 results sit close to N1X in multicore Cinebench, and its shared memory approach also addresses workloads that challenge small dedicated graphics allocations. RTX Spark’s distinction is NVIDIA’s software and graphics ecosystem; AMD retains the appeal of an x86 application environment.

Against conventional RTX 5060 and RTX 5070 laptops, the graphics results place Spark in a familiar performance class. The shared memory arrangement then changes the capacity discussion. A workload that fits comfortably within an ordinary GPU’s VRAM may favour that GPU’s particular implementation; a larger model or scene can make memory capacity the overriding factor.

RTX Spark is most compelling for GPU-accelerated work that needs substantial memory. CPU-heavy rendering still favours the HX rivals, while software compatibility can make an x86 system the more practical choice for an established workflow.

Verdict: a credible new option for demanding Windows work

RTX Spark gives Windows on Arm a more convincing proposition for graphics, creative software and local AI. The ProArt P16’s measured CPU performance provides a capable baseline, while the graphics results indicate performance around the RTX 5060 to RTX 5070 Laptop class, depending on the workload.

Its most consequential feature is the combination of that graphics capability with unified memory and CUDA. The 64 GB test configuration, with 32 GB allocated to the GPU, offers room for local workloads that can be constrained by the dedicated graphics memory of conventional laptops. And that slider to change allocation as per your needs adds to the platforms flexibility.

The platform still demands informed software choices. Browser results vary substantially, creator acceleration depends on exact application builds, and game compatibility includes anti-cheat and launcher support. The frame-generation-enabled gaming figures show a supported experience, while leaving base rendering speed and latency as separate questions.

The P16 is a suitable vehicle for the idea: a 16-inch 120 Hz OLED workspace, substantial battery capacity and useful connectivity. For creators and developers who can use NVIDIA acceleration and need more GPU-accessible memory, RTX Spark deserves serious consideration. For workloads dominated by CPU rendering, a particular x86 dependency or a game without confirmed support, established Intel and AMD systems remain relevant alternatives. NVIDIA has made the Windows laptop market more interesting by adding a different balance of compute, graphics and memory, and the early evidence gives that balance substance.