ASUS ProArt GR1X review: RTX Spark finds a desktop role
A 140W thermal design targets long creative jobs and continuously running AI agents in a compact desktop.
CUDA acceleration and up to 128GB of unified memory give the GR1X a distinct role in local AI and memory-heavy creative work.
RTX Spark’s existing results show capable graphics and responsive smaller language models, while software compatibility remains workload-specific.
The ASUS ProArt GR1X takes the RTX Spark out of a laptop and puts it somewhere that makes considerable sense, the all too familiar small form-factor desktop. And the benefits are as expected, you get a device that’s intended to stay connected, remain available and keep working. Pair it with NVIDIA PAIR (excuse the pun) and you can run local agents, render queues and image-generation workflows which do not need a battery or a built-in display. They need memory, supported acceleration and cooling that can handle a long session. And memory, this one has oodles of it. The GR1X combines NVIDIA’s 20-core RTX Spark N1X processor with a 6,144-core Blackwell RTX GPU and up to 128 GB of unified memory. Its 140 Watt thermal design shifts the emphasis towards sustained operation. The compact chassis and 10 GbE connection also make it an interesting fit for a studio, development desk or shared local AI installation.
Its strongest argument is the combination of resources. A conventional desktop can offer more graphics compute or broader component upgradeability, but accessing a large memory pool through NVIDIA’s CUDA ecosystem changes which workloads fit. That is particularly relevant to local models and complex creative projects.
ASUS ProArt GR1X Design and specifications
Coming in a compact 150 × 150 × 51 mm chassis, the GR1X occupies approximately 1.15 litres of chassis volume. That is small enough to sit beside a monitor without requiring the space of a conventional tower. Its external power adapter adds to the complete installation, but the computer itself suits a desk already occupied by creative equipment.
| Specification | ASUS ProArt GR1X |
|---|---|
| Platform | NVIDIA RTX Spark N1X |
| CPU | 20-core Arm CPU |
| GPU | NVIDIA Blackwell RTX, 6,144 CUDA cores |
| Memory configurations | 64 or 128 GB unified LPDDR5X |
| Listed memory bandwidth | Up to 300 GB/s |
| Thermal design | Up to 140 W TDP |
| Cooling | Dual-fan system with heat pipes and fin stack |
| Included SSD options | 1 or 2 TB M.2 NVMe PCIe 4.0 |
| Storage expansion | Two M.2 slots, PCIe 5.0/4.0 support as specified by ASUS |
| Wired networking | 10 GbE RJ45 |
| Wireless connectivity | Wi-Fi 7, Bluetooth 5.4 |
| Display connections | HDMI 2.1 and three USB-C DisplayPort Alt Mode connections |
| USB data connections | Three USB 3.2 Gen 2×2 Type-C connections |
| Power input | Separate USB-C connection supporting EPR PD 3.1 |
| Supplied adapter | 240 Watt |
| Dimensions | 150 × 150 × 51 mm |
| Listed weight | 1.48 kg |
| Operating-system options | Windows 11 Home or Pro |
The 10GbE connection is a useful desktop choice. A creator working from a suitably equipped NAS can keep footage, project archives and model libraries on shared storage. The connection does not guarantee a particular transfer rate, but it avoids making ordinary gigabit Ethernet the obvious ceiling. Also, unlike the NVIDIA DGX Spark the desktop RTX boxes don’t come up with any ConnectX-7 ports for daisy chaining the devices. The USB selection is less conventional. The available ports are all USB Type-C ports so you’ll have to get peripherals with Type-C ports or get adapters. Some omissions such as an SD card reader, an analogue audio jack and the Type-A ports do take away from making the GR1X a well-rounded desktop.
The 140W thermal design: meant for sustained workloads
The desktop’s cooling system is designed for a 140 Watt TDP. That comes into the picture most when a workload keeps the processor occupied for long enough to expose thermal and power limits. This includes rendering queues, repeated exports and in the case of AI PCs such as the RTX Spark - prompt processing or a sequence of generated images. A larger thermal budget does not translate into the same percentage improvement everywhere. A single-threaded application may already be close to its frequency ceiling. Token generation can be sensitive to memory bandwidth. A game can become CPU-limited even when the GPU has additional headroom.
| Workload | Potential benefit from greater sustained power | Other limiting factors |
|---|---|---|
| CPU rendering | Higher sustained multicore throughput | CPU frequency limits and renderer behaviour |
| GPU rendering | More consistent compute performance over a long render | Backend, scene structure and memory traffic |
| LLM prompt processing | Faster repeated processing where compute-limited | Model, context length and runtime |
| LLM token generation | Workload-dependent improvement | Memory bandwidth and model structure |
| Image-generation queues | Better maintained throughput across repeated jobs | Model precision, resolution and memory use |
| Mixed CPU/GPU workloads | More room for simultaneous activity | Shared power allocation and memory contention |
The appeal is consistency across a long working session. A desktop that maintains its performance through a queue can finish more useful work than a machine that is quick only at the start. TDP also differs from whole-system power draw. The 140 W specification is not a measured wall-consumption figure, and the 240 W adapter rating is not an indication that the computer continuously consumes 240 W. Actual energy use depends on the workload and time taken to complete it.
128 GB of unified memory
RTX Spark combines an Arm CPU with NVIDIA graphics and a shared physical memory pool. That arrangement gives a supported GPU workload access to substantially more memory than many compact PCs with conventional graphics configurations. The practical difference is between fitting a workload and running it quickly. A model that fits comfortably in a smaller dedicated VRAM pool can still run faster on a more powerful discrete GPU. A larger model that does not fit can make capacity the deciding factor.
| Memory configuration | Practical role |
|---|---|
| 64 GB unified memory | Local models and creative projects with a moderate GPU allocation, while retaining memory for Windows and applications |
| 128 GB unified memory | Larger models, more demanding scenes or additional capacity for concurrent work |
| 32 GB GPU allocation on the tested 64GB P16 | Existing platform configuration used with Qwen3.8-27B-UD-Q4_K_M |
A 32 GB GPU allocation is part of the physical memory pool, not an additional 32 GB of dedicated VRAM. Model weights, context cache and runtime buffers all compete for space, while Windows and other applications need memory of their own. The 128 GB configuration consequently buys room rather than an automatic doubling of speed. It is most valuable when the intended workload would otherwise be constrained by capacity. Over-allocating memory to the GPU can leave ordinary desktop work short of resources and introduce paging. Then again, in the NVIDIA App you can only take it up to 96 GB. So that always leaves 32 GB for the system.
N1X CPU performance
The 20-core N1X scored 553 single-core and 7,103 multicore in Cinebench 2026. Geekbench 7.1.0 returned 2,611 single-core and 23,103 multicore.
| Benchmark | N1X |
|---|---|
| Cinebench 2026 single-core | 553 |
| Cinebench 2026 multicore | 7,103 |
| Geekbench 7.1.0 single-core | 2,611 |
| Geekbench 7.1.0 multicore | 23,059 |
Those results establish enough CPU capability for a serious mixed-use system. N1X can handle parallel work while also supporting the interactive stages of editing, development and ordinary productivity. Its strongest distinction remains the GPU and memory architecture alongside that CPU.
| Processor | Cinebench 2026 single-core | Cinebench 2026 multicore |
|---|---|---|
| ASUS ProArt GR1X | 553 | 7,103 |
| NVIDIA RTX Spark N1X, ProArt P16 | 557 | 7,117 |
| Qualcomm Snapdragon X2 Elite Extreme X2E-96-100 | 667 | 7,410 |
| Intel Core Ultra 9 290HX Plus | 568 | 8,590 |
| AMD Ryzen 9 9955HX | 525 | 8,015 |
| AMD Ryzen AI Max+ 395 | 466 | 7,264 |
The laptop N1X result sits approximately 4% behind Qualcomm and 2% behind the Ryzen AI Max+ 395 in multicore rendering. The Intel HX processor leads it by a wider margin, as does the Ryzen 9 9955HX. In single-core rendering, N1X finishes close to Intel and ahead of both listed AMD processors, while Qualcomm retains a clear lead. Surprisingly, the desktop variant lagged a bit behind the laptop variant. However, context matters and the GR1X’s higher sustained-power design gives it an opportunity to improve on the laptop’s behaviour, particularly in lengthy parallel work, and single run benchmarks don’t clearly indicate that.
Everyday performance
Speedometer 3.1 returned 34.2 in Microsoft Edge. Speedometer measures responsiveness in simulated web applications, so the gap does not imply that every website loads 40-50% faster. It does show that the platform’s everyday performance depends on the application as well as the processor.
| Speedometer 3.1 browser | N1X |
|---|---|
| ProArt P16 | 36.1 |
| ProArt GR1X | 34.2 |
The Procyon Essentials screenshot broadens the picture to app startup, browser work, file operations and background activity.
| Procyon Essentials component | N1X score |
|---|---|
| Overall | 4,516 |
| App Startup | 5,090 |
| Browser, foreground | 4,031 |
| File Operations | 4,505 |
| Browser Tabs, background | 5,659 |
| Video Call, background | 5,013 |
The run used workload version 1.1.2835 and Chromium 145.0.7632.162. These scores cover the less glamorous work that continues around a render or an agent session. They establish completed productivity workloads on N1X, without turning the individual component scores into directly comparable measures of different tasks.
Graphics performance
The ProArt GR1X graphics results put N1X right in the vicinity of the RTX 5070. Steel Nomad returned 3,221, while Speed Way scored 3,082. This would be a tad better than the ProArt P16 and it’s where the higher TDP of the desktop variant seems to show prompt results.
| 3DMark workload | N1X, ProArt P16 score | N1X, ProArt GR1X score |
|---|---|---|
| Port Royal | 9,112 | 9,118 |
| Solar Bay Extreme | 11,522 | 12,021 |
| Speed Way | 3,072 | 3,082 |
| Steel Nomad | 2,991 | 3,221 |
| Graphics platform | Steel Nomad | Speed Way |
|---|---|---|
| RTX Spark N1X, ProArt GR1X | 3,221 | 3,082 |
| RTX Spark N1X, ProArt P16 | 2,991 | 3,072 |
| GeForce RTX 5070 Laptop, published average | 2,944 | 3,403 |
| GeForce RTX 5060 Laptop, published average | 2,582 | 2,806 |
| Radeon 8060S, published average | 2,338 | Not included |
Its larger GPU-accessible memory capacity then gives it a different balance of strengths from a conventional laptop GPU. Compute throughput remains important, but a scene or model that exceeds a smaller memory allocation changes the comparison.
Blender and creator performance
The standard Blender 5.3.0 Arm benchmark returned the following GR1X results using its automatic device selection.
| Blender scene | ProArt GR1X |
|---|---|
| Monster | 183.479 |
| Junkshop | 134.847 |
| Classroom | 90.504 |
For a studio, the GR1X’s appeal is supported acceleration combined with a small, permanent desktop installation. The deciding detail can still be a plug-in, codec or capture-device driver. A native editor alone does not establish support for the entire workflow.
Procyon text generation
The supplied AI measurements give a more tangible account of local inference than a peak FP4 compute specification. Procyon’s text-generation run used llama.cpp b10635 and workload version 1.1.228 on the 6,144-core N1X GPU.
| Model | GR1X | v/s P16 | GR1X | v/s P16 |
| Phi 3.5 | 2,562 | −6.8% | 69.34 tok/s | −6.0% |
| Mistral 7B | 2,027 | −13.8% | 51.13 tok/s | −3.5% |
| Llama 3.1 | 2,044 | −10.2% | 48.15 tok/s | −4.4% |
| Llama 2 | 2,284 | +5.6% | 27.96 tok/s | +0.4% |
Phi 3.5 delivers the highest output throughput in this run, at 69.34 tokens/s. Mistral 7B and Llama 3.1 both near 50 tokens/s, while Llama 2 is slower at 27.96 tokens/s. The recorded average first-token times remain below a second across these workloads.
That supports responsive streaming output for these benchmark prompts. Longer documents and multi-stage agent tasks can behave differently, particularly when the model must repeatedly process substantial context. The load-time figures are also those reported by the benchmark, not a guarantee of cold-start performance from storage in every application.
Agent performance: tool time counts too
The MLPerf SWE Agent entries report average expected duration and average tools duration for Llama 3.1 8B Instruct.
| MLPerf SWE Agent metric | WindowsML | llama-cpp |
|---|---|---|
| Average expected duration | 55.195 seconds | 46.025 seconds |
| Average tools duration | 6.363 seconds | 6.750 seconds |
Llama-cpp’s average expected duration is approximately 16.6% lower, despite its slightly lower overall text-generation throughput. Tool duration is slightly higher. These are separate exported metrics and should not be added together without the benchmark’s aggregation rules.
The result reinforces the value of measuring a workflow rather than only its output stream. An agent alternates between model processing, tool calls and inspection of the response. Improving one stage does not necessarily improve the others by the same amount.
The GR1X’s desktop format fits this work well. A document indexer or coding agent can remain available on a fixed machine rather than depending on a laptop that may be asleep or away from the desk. Its usefulness still rests on correct outputs and successful tool execution, neither of which can be established from throughput alone.
Computer vision: acceleration beyond chat
Procyon AI Computer Vision 2.0 returned an impressive set of scores. These workloads broaden the platform’s role to image understanding, segmentation, detection and enhancement. They are relevant to creative pipelines that process visual material as well as text. Compared to the P16 laptop, we’re seeing improvements as high as 10.2% and in some benchmarks, the scores lag by 3.5%.
| Workload | ProArt GR1X | Change |
| ConvNeXt | 1,980.40 images/s | +1.1% |
| BLIP decoder | 353.302 iterations/s | +1.5% |
| BLIP encoder | 392.013 iterations/s | +10.2% |
| SAM2.1 decoder | 805.602 iterations/s | +1.2% |
| SAM2.1 encoder | 60.898 iterations/s | +1.2% |
| DETR | 225.346 iterations/s | −3.5% |
| ESRGAN | 17.709 iterations/s | +4.9% |
The numbers use different operations and units, so ConvNeXt’s image throughput cannot be ranked directly against ESRGAN’s iteration rate. SAM2.1’s encoder and decoder also have markedly different costs. The speed of a complete application depends on which stages it runs and how often. For a dedicated desktop, batch processing is a natural fit. A queue of images can be analysed or enhanced while other work continues. The benchmark establishes supported acceleration, while actual project completion time depends on resolution, batching and the surrounding software.
Software and daily ownership
Windows on Arm runs native Arm applications and translates supported x86/x64 software through Prism. GPU acceleration adds another layer: an application’s CPU components and GPU kernels can follow different execution paths. For creative work, the complete dependency chain makes quite the difference. A supported editor may still rely on a plug-in, codec, capture card or audio interface without the required Arm support. Developers face similar considerations with packages and compiled dependencies.
The GR1X also differs from ASUS’s DGX OS-based Ascent machines. Its purpose is a Windows environment for everyday applications alongside accelerated local workloads. That can be valuable for users who want their editor, browser, development tools and models in one familiar installation. Storage deserves attention from the outset. Model weights, cached media, application environments and generated outputs can consume a 1TB drive quickly. The second M.2 slot and fast network connection provide useful expansion routes.
For continuously available operation, idle consumption and fan behaviour matter as well as performance under load. No GR1X wall-power logs, sound measurements or temperature traces accompany the available results, so quietness and energy efficiency remain unmeasured rather than inferred from the chassis size.
Verdict: a focused alternative to a tower
The ProArt GR1X has a clear role. It combines a compact desktop footprint with NVIDIA acceleration, substantial unified memory and a cooling design intended for sustained work. That combination suits local models, creative queues and background agents better than a specification sheet focused only on peak CPU speed would suggest.
The platform evidence is substantial. N1X delivers across creator workloads and more importantly, AI workloads. Those capabilities give the GR1X a credible foundation. Its 140 W thermal design then targets the next practical requirement: keeping the hardware productive through long jobs. The size of that sustained-performance benefit still needs desktop measurements.
The strongest fit is a creator or developer whose workloads benefit from both CUDA and a large GPU-accessible memory pool. The 10GbE connection also makes sense in a studio with shared storage. Less suitable workflows are those tied to unsupported x86 dependencies or a requirement for replaceable graphics hardware.
The GR1X is therefore a focused desktop proposition rather than a universal tower replacement. The architecture already explains why it deserves attention: substantial local AI capacity and RTX-accelerated creative work in a small Windows desktop.