Brief content overview
Generated as an editorial, reader-visible overview from the article content.
The ASUS ROG GR70 combines a compact design with powerful notebook hardware and, in the tested configuration, is suitable for gaming at WQHD resolution and local AI tasks. However, the tests also show a clear limitation: 8 GiB of graphics memory requires careful workload allocation when using larger models and parallel services. AI results vary depending on the task and do not replace a review of the outputs.
Key facts
Technical specifications
Measurements and observations
In the AI microbenchmark with Gemma 4 E4B Q4KXL, CUDA was ahead of Vulkan in prompt processing and token generation. The test series comprised three short runs per backend and does not support conclusions about sustained continuous load or efficiency.
A recorded proofreading run with Gemma 4 E4B took around 7.17 seconds on the server side. This is not an end-to-end time from submitting a request in the user interface to the fully displayed response, and by itself it does not demonstrate text quality.
The test system scored 923 points in the AI Center benchmark. Among other results, 14 of 16 reasoning tasks were solved correctly; 90 percent of the AI agent workflows were completed in full. The values refer to the documented tasks and their evaluation criteria.
In the thermographic measurements on the open motherboard, the surface temperature in the marked area near the power supply was 73.38 °C. The measurement point near the CPU was at 63.86 °C, and the marked PCB area was at 57.96 °C. These surface readings do not determine the temperature of individual components or the processor’s core temperature.
For the RTX 5070 Laptop GPU, the author cites their own telemetry measurements of an average of around 29 W during text generation and reasoning, and 69 W during image generation. The maximum recorded GPU temperature during image generation was 75 °C. The recorded text tasks involved 56,311 tokens and 12.2 Wh of GPU energy consumption; this is not a measurement of total power draw at the wall outlet.
Test conditions and limitations
The documented AI test system used Windows 11 Home, 64 GiB of memory, llama.cpp, and CUDA and Vulkan. The language model tests used specified tasks, settings, and model variants. The results apply to this configuration and cannot readily be extrapolated to other models, software versions, or GR70 configurations.
In addition to model weights, the 8 GiB of graphics memory must accommodate runtime buffers, context data, and other services. The file size of a model therefore does not indicate its total memory requirement during operation. A sandbox was also unavailable for the programming test; passing functional tests does not demonstrate secure, isolated code execution.
Assessment
For gaming at WQHD resolution, the results at medium settings seem consistent with a compact notebook graphics solution. In the more demanding games tested, the average frame rate remained below 60 FPS; however, averages alone say nothing about the consistency of individual frame times. The article’s desktop graphics card comparisons serve as a point of reference and are not performance requirements.
The GR70 is capable for local AI work in the tested workflows, but it is not without limits. CUDA proved faster than Vulkan in the short backend tests. When building a knowledge base, the dedicated GPU was significantly faster than the processor and integrated graphics. If the language model needs graphics memory, moving other services to the processor or integrated graphics can free up memory, but at the cost of performance.
The benchmark evaluates specific tasks, not general intelligence or error-free unattended operation. Checkpoints, dates, numbers, text revisions, and image captions must be reviewed. The author recommends Gemma 4 26B-A4B for certain German-language text tasks only if the available memory configuration does not force unfavorable offloading; no completed 26B hybrid run on the GR70 was available.
Strengths
Limitations
Anyone navigating the AI landscape at the moment might occasionally get the impression that artificial intelligence only begins where graphics cards cost as much as a decent used car and the power meter whimpers for mercy during reasoning. Apparently, anything below 24 GB of graphics memory is out of the question anyway; less than 48 GB makes things difficult, and anyone who really wants to work seriously with local language models should just buy half a workstation right away. That is, of course, wonderfully convenient, at least for those selling such hardware. The calculation looks a little different for the user. After all, there is often a pretty big difference between what is theoretically possible and what you actually need every day. That is exactly the difference I want to explore today.
That is why the ASUS ROG GR70 is on my desk: at first glance, it does not seem to fit this discussion particularly well. A compact gaming PC with notebook technology, a powerful Ryzen 9 9955HX3D and a GeForce RTX 5070 Laptop GPU whose 8 GB of graphics memory initially sounds more like a polite refusal in light of current AI ambitions. Of course, you could simply put the system through a typical gaming test suite, compare a few frame rates and then discuss the price. That would not be wrong, but it would overlook a substantial part of what this kind of hardware can actually do nowadays. So I want to find out how far you can get with a comparatively small local AI system if you make sensible use of the resources available instead of focusing solely on model size.
Of course, the GR70 cannot escape its origins either. So first comes the obligatory unboxing, followed by a brief detour into the supposedly idyllic world of gaming, where graphics cards have always been judged by how many images they can throw onto the monitor in a second. A little gaming is part and parcel of a ROG system, after all, and anyone testing a gaming PC should also show whether it can play games properly. I will take care of that with the appropriate seriousness, but without making it the main focus of this article. Because immediately afterward, I will take a sharp turn, leave the familiar FPS charts behind and head where the system interests me much more: my local AI workspace, with real applications, actual workflows and tasks where a correct answer matters more than an impressive token rate. Because agents are no longer found only in the movies; they are also in your own office.
That is where my own AI Suite comes in, which I developed specifically for scenarios like these. Of course, there are limits, and I would be the last person to tell you that a small language model suddenly has the same capabilities as a substantially larger system. Some tasks require more context, more knowledge or simply more computing power than it makes sense to provide locally. But the key point for me is that this does not have to impose a fundamental limitation on the entire workspace. My AI environment can use local models for everyday work and switch to more powerful cloud models when needed. However, only when such escalation is actually necessary and explicitly permitted or requested by the user. Anyone who wants to keep their data local should be able to do so. Anyone who consciously allows external computing power for a particularly difficult task should also have that option. Intelligent division of labor instead of a blind hardware arms race—that is my approach, at least.
Accordingly, you will not be getting any glossy images today featuring a decoratively illuminated desk, artificial fog and the usual marketing magic. I am interested in the GR70 as a work and gaming system packed into a tiny body. That is why my actual AI workflow, reproducible measurements, language models, office tasks, document processing, agent workflows and practical performance take center stage. The benchmarks come from my own AI environment and are based on what a computer like this is really supposed to do in everyday use. Synthetic measurements are welcome to help explain technical relationships, but they simply cannot replace completed work. After all, what good are 200 tokens per second if the model gets the calculation wrong in the end or schedules an appointment right in the middle of an hour that is already booked?
And don’t worry, the screwdriver will not be left in the drawer to cultivate mold spores. Naturally, I will also take an in-depth look at the GR70’s internals, including the PCB layout, power delivery, cooling and how ASUS managed to fit so much technology into such a small case. But only after the actual work is done. First, the computer has to show what it can do; then it can get undressed. Given the test objective, that order seems entirely appropriate to me. Oh, I can give you one preview without revealing half the results: the little ASUS can do more than its specifications might suggest at first glance. Of course, the 8 GB of graphics memory occasionally puts a damper on the fun, and not every task can be solved by cleverly switching models or tweaking a few settings. But that is not really the point. If you know your tools, accept their limitations and choose the right approach, you can often get surprisingly far without lugging around the biggest possible setup.
That is an old mountaineering rule, by the way: you do not have to climb every peak with the heaviest backpack. You should, however, know where the trail leads and when it makes more sense to turn back after all. With the GR70, there is a third option, too: you can get help from the cloud when things get difficult. And that is exactly what makes this interesting to me.
Unboxing and First Impressions: Lots of Computer, Little Packaging Romance
ASUS has opted for packaging for the ROG GR70 that initially takes up somewhat more space than the computer itself might lead you to expect. The color-printed box is fairly generous in size, which seems a little excessive for a mini PC that measures around three liters. Inside, rigid plastic inserts hold the system in place and protect it during transport. That is practical, although I would have welcomed a somewhat more restrained choice of materials for a product in this price range. After all, the box is supposed to get the computer to your desk in one piece, not win a design award itself. ASUS keeps the included accessories pleasantly simple. The GR70 comes with an external 330-watt power supply and power cable, as well as the usual documentation, including the user manual and warranty information. No extra cables for video output, elaborate accessory bundles or gaming gimmicks are included. The stand is part of the case design and comes pre-installed, so the system can be set up straight away. That means getting started initially involves nothing more than connecting the power supply, a monitor and the desired input devices. Given the available connectivity, this should not pose much of a challenge for most users.
The power supply itself is a reminder, however, that this little computer should not be mistaken for an ordinary office mini PC. With a rated output of 330 watts, it provides enough power headroom for the Ryzen 9 9955HX3D, the dedicated GeForce RTX 5070 Laptop GPU and the rest of the system hardware. The output voltage is 19.5 volts at a maximum of 16.9 amps. Of course, actual power consumption is substantially lower during normal operation and depends heavily on the workload. The external power supply also takes up space, even if it naturally does not appear in the system’s official volume specifications. The famous three liters apply only to the part of the system that actually sits on the desk.
A Notebook Without a Screen, but with a Character of Its Own
The unusual case shape is the first thing that stands out after unboxing. Measuring 282.4 × 187.7 × 56.6 millimeters, the GR70 is considerably flatter than a classic mini tower, but much larger than the small office NUCs that you can occasionally hide behind a monitor. When placed upright on its wide stand, the system looks more like an especially slim game console. Nevertheless, it takes up pleasingly little desk space compared with a full-size gaming PC. The case weighs around 2.75 kilograms and stands correspondingly solidly thanks to its substantial aluminum base. The design is reminiscent of the ROG NUC 2025 for a reason: ASUS uses largely the same case design. The front of the GR70 is finished in a slightly lighter gray, which ASUS calls Star Grey, while the other surfaces are predominantly dark. The exterior is largely made of plastic, with a metal inner frame providing the necessary rigidity. The build quality seems appropriate for the technical concept, without ASUS having to use excessive amounts of material. In any case, the panel gaps and mechanical design are in keeping with the expectations for a system that might be transported more often than a large desktop tower.
Naturally, a ROG product cannot do entirely without visual flourishes. One side features an illuminated ROG logo behind a transparent cover, and there are also lighting elements on the front. These can be customized using ASUS Armoury Crate or Aura Sync. Those who would rather use the system as an unobtrusive AI workstation can tone down the lighting accordingly. Personally, I can live with the illuminated logo, as long as the actual hardware does not dictate the fan curve in return. In the end, I am more interested in what is working behind the cover than in what is blinking in front of it. The numerous ventilation openings are also eye-catching. The side panels and the top of the case are generously perforated, with ASUS integrating some of the openings into the ROG design. This gives the system a distinctive look, but above all serves a technical purpose. Three fans work inside, drawing in cool air from the sides and expelling the heated air through the rear. Two fans handle most of the CPU and GPU cooling, while a third radial fan takes care of an additional cooling path. I will take a closer look at how ASUS has designed the airflow and thermal contact surfaces during the later teardown.
Plenty of Ports, but Not Much Room for Them
ASUS has made pleasingly consistent use of the available case space when it comes to connectivity. On the front, or rather along the side-accessible port panel, there are two USB-A ports rated at 10 Gbit/s each, a USB-C port also rated at 10 Gbit/s and a combined 3.5 mm audio jack. There is also the power button and the lighting elements. That is enough for typical peripherals, external storage devices or a headset, without having to reach behind the system every time.
The rear is where things really get crowded. Two HDMI 2.1 outputs, two DisplayPort 2.1 connectors and an additional USB4 port supporting up to 40 Gbit/s provide unusually versatile video and data connectivity. DisplayPort 1.4 is also available over USB4. Together, the platform supports up to five simultaneously connected displays, although the specific resolutions and refresh rates naturally depend on the chosen combination. Four more USB-A ports rated at 10 Gbit/s each, a 2.5-Gbit Ethernet port and the socket for the external power supply complete the port panel. A Kensington lock slot is also present.
This setup is particularly practical for the AI workspace I am planning. Multiple monitors, additional drives, audio hardware and various input devices can be connected without immediately needing an external USB hub. The integrated network connection also provides a solid 2.5-Gbit/s foundation for accessing large data collections or a local NAS. Wireless connectivity comprises Wi-Fi 7 and Bluetooth 5.4, although the MediaTek module used here, with a maximum channel width of 160 MHz, does not use Wi-Fi 7’s highest theoretical configuration. That is no disaster for everyday use; you just should not confuse the actual specifications with the wireless standard’s maximum capabilities.
Easy to Service Without a Major Screwdriver Saga
It is good to see that ASUS has taken future access to the memory and SSDs into account despite the compact design. After removing or releasing the relevant case fastener, the side cover can be taken off to expose the accessible side of the motherboard. Depending on the task, the stand may need to be removed first. Overall, this is much more straightforward than with many gaming laptops, whose bottom panels are sometimes secured with numerous screws and plastic clips. Beneath it are two SO-DIMM slots and two M.2-2280 connectors. This means both the RAM and the storage drives can generally be replaced or upgraded. ASUS specifies a maximum of 96 GB of DDR5-5600. The M.2 configuration includes one PCIe 5.0 x4 interface and another with PCIe 4.0 x4. The 64 GB of RAM installed in the test sample is therefore by no means the technical maximum. The GeForce RTX 5070 Laptop GPU is a different matter: its 8 GB of graphics memory are part of the permanently integrated graphics solution and cannot be upgraded like a SO-DIMM module. This limitation will become relevant later when we look at larger local language models.
However, I’ll save the detailed look at the circuit board, power delivery and cooling for the teardown. The unboxing is first about placing the computer in context as a complete system. And the GR70 makes a pretty sensible impression: a small footprint, a powerful mobile platform and an unusually extensive selection of ports for a system of this size. The illuminated ROG elements are a reminder that ASUS originally positioned the device as a gaming computer. However, I’m about to put it to a few other tasks as well.
ASUS ROG GR70 specifications
The following overview refers to the configuration used in the test, with a Ryzen 9 9955HX3D, GeForce RTX 5070 Laptop GPU and 64 GB of DDR5. Memory and SSD configurations may differ in other retail variants. And after all those specifications, it’s time for a quick game before things get serious and really interesting.
