ASUS ProArt GR1X Tested: RTX Spark Mini PC Runs Cooler, With Two M.2 Slots
- Dylan Bryan
- 841 views
ASUS ProArt GR1X holds sustained CPU speed better than RTX Spark laptops and runs cooler, but its 1TB Gen4 drive trails in model loading.

The ASUS ProArt GR1X puts NVIDIA's RTX Spark N1X into a 150mm-square chassis of about 1.15 liters. It has two M.2 2280 SSD slots, 10GbE, and a dual-fan cooler that ASUS says is built to sustain 140W. Our test unit is the 64GB configuration with a 1TB PCIe Gen4 SSD and Windows 11 Pro. The question we wanted answered is simple: how much does a desktop-class cooler change the results for the same 20-core chip that laptops carry? We ran it against MSI's 128GB EdgeMesa N AI+, the other RTX Spark mini PC in our lab, and the two fastest RTX Spark laptops we have tested, the HP OmniBook Ultra 16 and the Dell XPS 16 Creator Edition.


The Short Version
The GR1X's cooling does most of the work here. In a 30-minute Cinebench 2026 multi-core loop, it held 7,306 points in steady state, 5 percent ahead of the HP OmniBook Ultra 16 at 6,970 and 12 percent ahead of the Dell XPS 16 at 6,550. Its x265 encode ran at 44.1 fps against 40.6 and 38.8 on the laptops. The MSI EdgeMesa matched it on both, so the lead comes from the chassis class rather than from the N1X itself.
Temperatures show the bigger gap. Both laptops averaged 97C through the loop and sat at their 98C limit. The GR1X averaged 83.3C and peaked at 88.7C, while running a higher clock. Storage is the weak point. The 1TB Gen4 drive in our unit read at 6,511MB/s, and a cold load of a 12.8GB model took 10.94 seconds, slower than any other system in the group. That's the part we'd upgrade first, and the second M.2 slot makes that practical.
Where the GR1X Fits in ASUS's Lineup
ASUS sells the RTX Spark systems under its ProArt creator brand. The GR1X is the desktop in a family that also includes the 16-inch ProArt P16 and the 14-inch ProArt P14 laptops. Separately, ASUS sells the Ascent GX10, a DGX Spark system that runs NVIDIA's Linux-based DGX OS for AI development. The GR1X is the Windows counterpart. ASUS aims it at prosumers, AI-focused creators, workflow builders, developers, and gamers who want local agents working inside everyday Windows applications.
The configuration choices are straightforward. ASUS offers 64GB or 128GB of LPDDR5X-9400 unified memory, in black or silver, with either Windows 11 Home or Pro. The bundled software includes ProArt Creator Hub, the StoryCube and MuseTree creative applications, validated ComfyUI workflows, and a three-month Adobe Creative Cloud subscription.
In the US, ASUS sells the GR1X through its own eShop, Best Buy, Micro Center, Central Computers, and Amazon. The 64GB model with a 1TB Gen4 SSD and Windows 11 Home is listed at $4,499 on the ASUS eShop and $4,499.99 at Best Buy, where a 2TB version is listed at $4,999.99. Our review unit matches the 1TB configuration but runs Windows 11 Pro. When we checked, the eShop showed the system as out of stock and Best Buy listed it as coming soon. NVIDIA told us RTX Spark desktops start at $4,399, which puts the 64GB GR1X $100 above that entry price.
ASUS quotes the platform at up to 1 petaFLOP of FP4 AI performance, 120B-parameter LLMs with up to 1 million tokens of context, 3D scenes up to 90GB, and 1440p gaming above 100 fps. ASUS's RTX Spark overview treats the model figures as capacity guidance that assumes suitable quantization, and it ties 70B and 120B-class models to the 128GB configuration. That qualifier matters for this review, because our unit has 64GB.
Design, Ports, and Cooling
The GR1X measures 150 x 150 x 51mm and weighs 1.48kg, which works out to 1.15 liters. ASUS quotes 1.1 liters. The MSI EdgeMesa N AI+ measures 152 x 147 x 53mm, comes to 1.18 liters, and weighs 1.32kg. The two RTX Spark mini PCs occupy about the same footprint on a desk, with the ASUS roughly 160g heavier.
The black version has an anodized finish and the silver version is sandblasted. Both carry an anti-fingerprint matte coating. The Wi-Fi antenna sits under a textured, glass-like panel on the top. The front has one control, a power button with an LED. Power comes in over USB Type-C from a 240W adapter, so the chassis has no barrel connector.

Every port sits on the rear panel. You get four USB Type-C ports at 20Gbps, HDMI 2.1, a 10GbE RJ45 port, and a Kensington lock slot. Three of the Type-C ports support DisplayPort Alt Mode. The fourth is the power input under the USB PD 3.1 EPR specification. Wireless is Wi-Fi 7 with a 2×2 radio and Bluetooth 5.4. The system drives up to four displays, using HDMI plus the three DisplayPort-capable Type-C ports. ASUS lists no USB Type-A port, no audio jack, and no front-panel USB. Most desks will need a hub or dock, and that's a real cost on top of the $4,499 sticker.

The cooler is a dual-fan design. Air enters through vents in the base, crosses 57 fins fed by four heat pipes, and exits at the rear. ASUS specifies 109 blades per fan. It says the system is built to sustain 140W and that owners will rarely notice thermal throttling, even at 25C to 30C ambient. That's a claim about real behavior, so we tested it directly, and the sustained results below are where it either holds or doesn't.

Storage Layout
Two full-length M.2 slots are rare in a chassis this size, and that's the design choice we'd point to first. ASUS lists two M.2 2280 slots with support for PCIe 5.0 and 4.0 drives up to 4TB. The 2280 length is the most common retail size for Gen4 and Gen5 SSDs, so a buyer can add a second drive for models and datasets, or replace the boot drive with an off-the-shelf part.

ASUS's factory options are a 4TB Gen5 drive and 1TB or 2TB Gen4 drives. Our unit shipped with a 1TB Micron drive, which sits in the primary slot on the motherboard beneath the cooler structure. The second M.2 2280 slot is on the underside of the board. It was empty in our unit and is reachable once the bottom plate is off. Limited storage has been a recurring shortcoming in the GB10 systems we've tested, which typically ship with a single short M.2 drive. An open second slot changes that calculation.
Test Setup and Memory Behavior
We tested the GR1X as it arrived. It had the RTX Spark N1X, 64GB of LPDDR5X-9400, the 1TB Micron SSD on a PCIe Gen4 x4 link, and Windows 11 Pro 26H1 on Arm. The power mode was set to Best performance, with a 4K external monitor attached. Windows assigned 30.5GB of the unified memory to the GPU as dedicated memory and another 16.2GB as shared.
The memory is one physical pool. Windows presents the GPU allocation as dedicated GPU memory and the rest as system memory, which the GPU can spill into when it needs more. A slider in the NVIDIA app sets the allocation. The default sits close to an even split on 64GB and 128GB RTX Spark systems, and the app warns that extreme minimum or maximum settings may cause issues. NVIDIA advises keeping the default unless a workload needs more, since a bigger GPU share leaves less memory for Windows and other applications. It recommends leaving at least 16GB for the CPU.
NVIDIA says agents and LLMs see little performance loss when they spill into system memory. It suggests raising the allocation for image and video generation models that need more than the default provides. We tested every system at its default. Our suite sizes its model tiers to the dedicated amount, so a model the suite skips on a 64GB system may still run there with spillover or a larger allocation.
All results come from sr-client 1.13, our Windows client suite for local AI, content creation, CPU, graphics, and storage. LLM serving and MLPerf Client run through llama.cpp on CUDA. Model loading, the coding agent, and RAG run through Ollama on CUDA. Image generation uses stable-diffusion.cpp on Vulkan. We report the median of repeated passes and the newest valid result where a test ran more than once. An asterisk marks results the suite flagged for a failed pass or high run-to-run variance. The two laptops were tested earlier on sr-client 1.11.
The comparison set is the MSI EdgeMesa N AI+, a 1.18-liter desktop with 128GB of memory and a 4TB Samsung Gen5 SSD, plus the 128GB HP OmniBook Ultra 16 with a 2TB Sandisk Gen5 SSD and the 64GB Dell XPS 16 Creator Edition with a 1TB SK hynix drive. The XPS 16 is the direct memory match for the GR1X. The EdgeMesa and the HP show what the larger configuration adds. All four use the same 20-core N1X. The two desktops run it at a higher power limit under larger coolers, so power, cooling, memory, and storage are the variables. An x64 build of Cinebench crashed under emulation on all four machines, so no emulation results appear here.
Sustained Performance and Thermals
We started with the sustained test, because it's the one a mini PC has to win. The loop renders the Cinebench 2026 multi-core scene continuously for 30 minutes. The GR1X finished at 7,306 points in steady state, 1.007 times its first pass, so it didn't throttle. Mean steady-state clock on the GR1X was 3,385MHz, compared with 3,239MHz on the HP and 3,083MHz on the XPS 16. The EdgeMesa held 7,276 points at 3,372MHz, level with the ASUS.
Temperature separates these systems more than the scores do. The GR1X averaged 83.3C and peaked at 88.7C, nearly 10C under the 98C limit that both laptops hit. The EdgeMesa produced the same score at an 88.4C average, 5C warmer than the ASUS, and its peak reached 107.3C.
| Cinebench 2026 loop | ASUS ProArt GR1X (64GB) | MSI EdgeMesa N AI+ (128GB) | HP OmniBook Ultra 16 (128GB) | Dell XPS 16 Creator Edition (64GB) |
|---|---|---|---|---|
| Steady-state points | 7,306 | 7,276 | 6,970 | 6,550 |
| Mean CPU temperature (C) | 83.3 | 88.4 | 97.4 | 97.3 |
| Peak CPU temperature (C) | 88.7 | 107.3 | 98.2 | 98.3 |
| Steady-state clock, mean (MHz) | 3,385 | 3,372 | 3,239 | 3,083 |

A single multi-core run shows the same order. The GR1X scored 7,304 against 7,001 on the HP and 6,631 on the XPS 16, leads of 4 and 10 percent, with the EdgeMesa at 7,284. Single-core results were 553 on the GR1X, 548 on the EdgeMesa, and 561 and 565 on the laptops. That 3 percent spread favors the laptops slightly, which fits a test that never loads the cooler.
The temperature pattern holds across the rest of the CPU suite. The GR1X averaged 70.0C in Blender, 67.2C in the LLVM build, 66.5C in HandBrake, and 55.4C in 7-Zip. The laptops averaged between 78C and 97C in those tests. In Blender, HandBrake, and 7-Zip the ASUS delivered throughput equal to or better than the HP while running 21C to 27C cooler on average. Stockfish averaged 80.8C on the GR1X against roughly 97C on both laptops.
The Closest Comparison: EdgeMesa
The EdgeMesa is the fairest comparison because it's the same class of chassis. The GR1X ran 3C to 8C cooler than the MSI on average across the seven CPU tests in our table, at 70.0C against 76.0C in Blender and 66.5C against 74.1C in HandBrake. The gap widened to 11C to 18C in LLM serving, RAG, and image generation, where the GR1X averaged 73.4C, 65.3C, and 63.9C. Throughput in those tests was essentially identical between the two desktops.

Peak readings on the GR1X stayed below 90C in nine of the ten tests we logged, including 88.7C in the loop, 88.6C in Stockfish, and 88.4C in LLM serving. The exception was the standalone Cinebench run, which logged a 103.8C maximum against a 65.3C mean. That was a short excursion that didn't repeat during the 30-minute loop. The EdgeMesa logged peaks above 100C in three tests. GPU-heavy work ran coolest of all, with image generation averaging 63.9C on the ASUS.

Encoding, Compute, and Desktop Work
Video encoding gave the GR1X its largest application lead. HandBrake's x265 encode ran at 44.1 fps against 40.6 on the HP and 38.8 on the XPS 16, a 9 and 14 percent lead. SVT-AV1 ran at 16.4 fps against 15.4 and 14.0. The EdgeMesa was level at 43.7 and 16.3 fps. Blender's Classroom scene rendered at 95.7 samples per minute, 3 and 10 percent ahead of the laptops, and the Monster scene at 215.6 was 2 and 8 percent ahead. The Junkshop scene was effectively a tie with the HP, at 127.2 against 127.7, while the EdgeMesa led it at 129.7.

The all-threads 7-Zip result was 153.8 GIPS on the GR1X, 152.6 on the EdgeMesa, and 150.6 and 147.0 on the laptops. That's a smaller spread, which fits a test limited as much by memory as by clock speed. The LLVM build was the one all-core test where the ASUS didn't lead. It finished in 225 seconds, behind the EdgeMesa at 217 and the HP at 218, and ahead of the XPS 16 at 232. A compile mixes single-thread phases and file I/O with parallel work. The one-thread 7-Zip score of 8.92 GIPS on both desktops sat just under the laptops at 9.10 and 9.13, which helps explain it.


Shorter and lightly threaded tests land close together across the group. Stockfish with all threads ran at 15.13 Mnps on the GR1X, level with the HP at 15.19 and 9 percent ahead of the XPS 16 at 13.87. primesieve to 10^12 finished in 9.39 seconds, against 9.53 on the HP and 10.39 on the XPS 16. The EdgeMesa ran Stockfish at 14.42 Mnps and primesieve in 9.40 seconds. The single-thread Python tests and Speedometer 3.1 were within about 3 percent of the HP on both desktops. The XPS 16 led the browser test at 36.5 against 34.3 on the GR1X. For everyday desktop responsiveness, the four machines are effectively equal.
Memory-bound work is a wash, as you'd expect with identical LPDDR5X-9400 in all four machines. STREAM Triad measured 124,529MB/s on the GR1X, 122,423 on the EdgeMesa, and 120,783 and 122,872 on the laptops. Latency on 4KiB pages ranged from 140.5ns to 143.6ns. GROMACS benchMEM produced 17.41 to 17.48 ns/day across the group. OpenBLAS DGEMM at 356.8 GFLOPS trailed the HP at 374.0 by about 5 percent, and the LU solve landed within about 2 percent across the group.
Local AI Performance
The serving test sends one, four, and eight simultaneous requests to each model through llama.cpp. That approximates a few people sharing one system. A single request produced 58.4 tokens per second on Ministral 3 3B, 42.7 on the 8B model, and 13.1 on Qwen3.8 27B. Those single-request numbers sat within 6 percent of both laptops and within 1 percent of the EdgeMesa. The first token from the 27B model arrived in 8.6 seconds on the GR1X and 8.7 on the EdgeMesa, against 10.8 and 10.9 seconds on the HP and the XPS 16.
The systems separate once requests stack up on the larger model. At four requests on Qwen3.8 27B, the GR1X produced 31.4 tokens per second against 28.1 on the HP and 27.5 on the XPS 16. At eight requests it produced 40.0 against 33.8 and 31.9, which is 18 percent ahead of the HP and 25 percent ahead of the XPS 16. Ministral 3 8B at four requests ran at 109.3 tokens per second, compared with 104.1 and 96.3. The desktop's higher power limit and cooler help the chip stay well under its ceiling, and the gap grows with the load.

The EdgeMesa followed the GR1X within 2 percent across the serving table. Its 128GB of memory made no difference in these tests, because every model in the table fits on our 64GB unit. The two desktops serve models of this size at the same rate.

The 64GB unit loaded the 27B model with room for eight concurrent requests. It did not run the 35B coding-agent model or Qwen-Image at Q8. ASUS ties 120B-class models to the 128GB configuration, and we haven't tested that size on any RTX Spark system. Our Windows client suite doesn't include it. Models of that size belong in our Linux suite, run on larger unified-memory desktops and GPU workstations. If you plan to go well past the 27B class, the 128GB GR1X is the one to order.
MLPerf Client issues one prompt at a time, and the results are a four-way tie. The GR1X generated 56.8 tokens per second on Llama 3.1 8B and 96.3 on Phi-4-mini. The EdgeMesa managed 56.3 and 96.8, the HP 56.5 and 97.0, and the XPS 16 55.7 and 96.2. Time to first token on Llama 3.1 8B was shortest on the ASUS at 0.24 seconds, and the ASUS tied the EdgeMesa at 0.11 seconds on Phi-4-mini. For one person working with a small model, none of these four machines will feel faster than the others.
Model Loading: Where the GR1X Finished Last
Cold model loads are the one AI result where the GR1X finished last, and the drive is the cause. It took 10.94 seconds to load the 12.8GB gpt-oss:20b model from disk to first token. The HP needed 6.70 seconds, the EdgeMesa 7.00, and the XPS 16 7.73. Read throughput during that load averaged 1,788MB/s on the ASUS, against 3,446MB/s on the EdgeMesa's Gen5 drive, 3,211MB/s on the HP's, and 2,372MB/s on the XPS 16. Warm loads were 3.68 to 3.69 seconds on the GR1X and the laptops and 3.96 on the EdgeMesa. That confirms the gap sits in storage, not in the processor.
The coding agent asks a model to write and test code across several rounds, three times per configuration. With gpt-oss:20b at a 64K context, the GR1X passed one of three attempts, the EdgeMesa and the HP passed two each, and the XPS 16 passed all three. Generation rates were level, between 57.1 and 58.5 tokens per second, so the pass counts reflect run-to-run variation in the model's output. The suite flagged several of the ASUS, MSI, and HP runs for failing at least one test, so treat those pass counts with caution. The 128GB EdgeMesa also ran the 35B agent tier, which the suite doesn't run on the 64GB systems. When Speedometer ran alongside the agent, the browser slowed by 1.79x on the ASUS and 1.78x on the MSI, against 2.35x and 2.79x on the laptops. That's a useful sign of how well each desktop keeps a foreground app responsive during background AI work.
RAG and Image Generation
Our RAG test embeds a set of passages, retrieves against them, and answers questions with phi4-mini. The GR1X ingested 115.6 passages per second, 4 percent ahead of the HP at 111.5 and 11 percent ahead of the XPS 16 at 104.4. It generated answers at 104.2 tokens per second against 103.4 on the HP and 95.0 on the XPS 16. The median question was answered in 0.22 seconds. The EdgeMesa matched it at 114.7 passages per second and 103.4 tokens per second. A document assistant at this scale fits comfortably within the 64GB configuration.
Image generation through Vulkan gave mixed results. FLUX.2 klein 4B at Q4 produced 10.07 images per minute on the GR1X, against 11.01 on the HP and 11.17 on the XPS 16. Z-Image Turbo 6B at BF16 matched the HP at 4.17 against 4.19, and Qwen-Image 20B at Q4 edged both laptops at 0.53 images per minute. The laptops ran the earlier suite release, while the EdgeMesa ran the current one and produced 9.95 on FLUX.2, level with the GR1X. That suggests the suite release accounts for some of the gap, so we wouldn't read too much into the laptop numbers on that test.

The larger Q8 and BF16 versions of Qwen-Image 20B did not run on the 64GB unit. The 128GB EdgeMesa and HP completed the Q8 version at 0.51 and 0.47 images per minute. FLUX.2 klein at Q8 was flagged for a stalled image, and the BF16 version failed on both attempts when the Vulkan device was lost. We left those rows out of the table.
Graphics and Storage Results
SuperTuxKart at the ultra preset averaged 274.0 fps at 1080p on the GR1X. That was ahead of the HP at 269.7 and the XPS 16 at 248.6. The GR1X also averaged 114.1 fps at 4K on our external monitor, a resolution we didn't run on the laptops, and the EdgeMesa managed 111.8 fps at 4K. FFmpeg's hardware HEVC encode through Media Foundation reached 65 fps on the GR1X, against 54 on the EdgeMesa, 53 on the HP, and 61 on the XPS 16. The hevc_nvenc encoder isn't present in the Arm build of FFmpeg that the suite uses, so that's the path we could measure.
Our unit's 1TB Micron drive links at PCIe Gen4 x4 and performs like a Gen4 drive. It delivered 6,511MB/s sequential read and 5,235MB/s sequential write on the 25GiB footprint. That's level with the 1TB drive in the XPS 16, at 6,502 and 5,670MB/s, and about half the rate of the Gen5 drives in the EdgeMesa at 13,315 and 12,910MB/s and the HP at 13,598 and 11,899MB/s. ASUS offers a 4TB Gen5 option for the GR1X, and the cold-load results show what a faster interface is worth for model files.

Random reads were the strongest part of the Micron drive's results. At queue depth 32 across four jobs, it delivered 552,652 IOPS, ahead of the HP at 449,043 and the XPS 16 at 300,063. Random writes were the weak point at 143,165 IOPS, against 386,814 on the EdgeMesa and 234,888 and 256,856 on the laptops. A buyer can keep this drive for Windows and add a second drive for models. That's the first upgrade we'd make.

Verdict: Who Should Buy the GR1X
The GR1X makes its case through cooling. It ran our 30-minute loop 5 to 12 percent faster than the fastest N1X laptops while averaging 14C lower, and it stayed under 90C in nine of the ten logged tests. That margin showed up as 9 to 14 percent more x265 throughput and 18 to 25 percent more output with eight requests on a 27B model. ASUS's claim that owners will rarely see thermal throttling matched what we measured in these tests.

The storage layout is the other reason to choose this chassis. Two M.2 2280 slots mean the 1TB Gen4 drive in our unit, which was the slowest part of the system in cold model loads and random writes, is a starting point an owner can build on with standard retail drives. It's a fixable weakness, and that matters.
Memory is the decision you have to get right at purchase, because unified memory can't be added later. The 64GB configuration handled models up to the 27B class with eight concurrent requests, RAG, and the smaller image models. It stopped short of Qwen-Image at Q8 and the 35B coding-agent model. For work inside that envelope, the 64GB GR1X matched or led the 128GB systems in LLM serving and most CPU tests. Its widest deficits were about 4 percent in the LLVM build and about 5 percent in OpenBLAS DGEMM. Larger models call for the 128GB GR1X.
Against the MSI EdgeMesa N AI+, performance is effectively a tie. The two were within 2 percent of each other in the Cinebench loop, Blender, x265 encoding, and LLM serving. The GR1X ran 3C to 8C cooler on average in CPU tests, and it has two M.2 2280 slots where the EdgeMesa offers one SSD. The EdgeMesa we tested carried 128GB and a 4TB Gen5 SSD, so it loaded models faster and ran the larger ones our 64GB unit couldn't. NVIDIA told us that EdgeMesa configuration carries a $7,499 MSRP, which is $3,000 above the 64GB GR1X.
Within ASUS's own range, the GR1X is the stationary Windows option beside the ProArt P16 and P14 laptops. The Ascent GX10 remains the pick for Linux and DGX OS development. Against the laptops we tested, the GR1X trades the built-in display, keyboard, and battery for lower temperatures, higher sustained output, and a second SSD slot. At $4,499 with 64GB and a 1TB SSD, it lands about $300 below the 64GB XPS 16 Creator Edition, which we reviewed at $4,799.99 with the same memory and storage. Once you add a display, keyboard, and whatever hub the GR1X needs for its missing Type-A and audio ports, that saving shrinks quickly.
The next thing to watch is the 2TB and 4TB Gen5 options. If the drive gap closes, the cold-load penalty goes away, and the GR1X's thermal lead looks even stronger for a desk that runs sustained local models all day.
Leaderboard: The ASUS ProArt GR1X holds the Best RTX Spark Mini PC spot on our Best Mini PCs leaderboard.
