NVIDIA Project G-Assist: Our Insights

Early this year, NVIDIA finally released Project G-Assist as a real product you can try, the "concept" of which first surfaced way back in April of 2017. The original idea (jokingly) revolved around providing players as much assistance as possible to get them through tricky sections, while the real product relies on AI and is more than just an in-game assistant.

What is Project G-Assist?

Currently, Project G-Assist makes use of Meta’s Llama-3.1-8B small language model (SLM), which runs locally on your PC, more specifically, on your RTX GPU. In NVIDIA’s words: "As modern PCs become more powerful, they are also growing in complexity to operate. G-Assist helps users take control of various PC settings, from optimizing game and system settings, overlaying frame rates and other critical performance stats, to controlling select peripheral settings (like lighting)—all through basic voice or text commands."

The idea isn’t too different from how Google and Apple leverage AI models to enhance their respective digital assistants, allowing them to better understand human language and adjust settings without having to navigate pages-deep menus in different corners of the system. This theoretically would be especially helpful for the casual user: just as those of us who are, say, tech enthusiasts like to fiddle with knobs to our heart’s content, GPU overclocking or tweaking graphics settings might be too intimidating for them – this is where Project G-Assist comes in.

Setup

There are a few things you’ll need to know before installing Project G-Assist, the first of which are the system requirements. The most important of which is that you’ll need to have an RTX 30-series or newer GPU with at least 12GB of VRAM (laptop GPUs are not included at the moment) – unfortunately, this creates a situation where owners of the RTX 3060 12GB can run the model, while owners of the higher-end RTX 3080 (with 10GB VRAM) cannot, due to some strange VRAM configurations of past generations. Ouch.

Assuming your GPU hardware meets the requirements, you’ll also need Windows 10 or Windows 11, along with GPU driver version 572.83 or later; for storage, it’ll need at least 6.5GB of disk space for the system assistant functions to work (voice commands will require an additional 3GB). Currently, only English is supported.

You’ll also need to install the NVIDIA App in order to enable Project G-Assist on your system; for peripheral-related hardware requirements, the current version supports MSI motherboards, along with peripherals from Logitech G, Corsair, and Nanoleaf. Not all models are supported from these brands – check the "System Requirements" tab under the Project G-Assist homepage for more details.

Test System

  • CPU: Intel Core i9-13900K
  • Cooler: Cooler Master MasterLiquid PL360 Flux 30th Anniversary Edition
  • Thermal Paste: Thermal Grizzly Kryonaut
  • Motherboard: ASUS ROG Maximus Z790 Apex
  • GPU: NVIDIA GeForce RTX 5090 Founders Edition
  • Memory: Kingston FURY BEAST RGB DDR5-6800 CL34 (2x16GB)
    • Configured to DDR5-6400 CL32 XMP profile
  • Storage: ADATA LEGEND 960 MAX 1TB
  • Power Supply: Cooler Master MWE Gold 1250 V2 Full Modular (ATX12V 2.52) 1250W
  • Case: VECTOR Bench Case (Open-air chassis)
  • Operating System: Windows 11 Home 24H2

Testing

As noted in the benchmark system specifications above, we will be using an NVIDIA GeForce RTX 5090 Founders Edition to demonstrate this feature. This flagship Blackwell-powered GPU features 32GB of GDDR7 VRAM, 5th-gen Tensor Cores, and 21,760 CUDA cores, all combining to deliver 3,352 TOPS of AI-specific FP4 performance (note that this figure can’t be directly compared with the RTX 4090’s 1,321 TOPS, which uses FP8).

Note: At the time of testing, Project G-Assist is still a pre-release build (version 0.1.9), so some features may be incomplete. Results generated from the tests performed below will only be relevant to this version, as results will vary as the AI model and features are updated over time.

First Use

This is what you’ll first see after enabling the feature via the Alt+G hotkey, and it’ll permanently stay on some part of your screen until you completely disable it (which can be done via quick settings with the Alt+R hotkey). As with AI language models, disclaimers apply – hallucinations may occur (language models may produce incorrect results, often with convincing to the uninitiated), so be sure to fact-check as much as possible.

The disclaimer message also appears when you first enter a message/command, again declaring that the AI-generated results cannot be completely guaranteed. After you see this message, the chatbot is ready to respond to commands with natural language – that is, only a limited set of commands (natural language or otherwise) is available in this version, which you can reference on the website.

System Information and Monitoring

Starting off with simple questions such as the nature of the system, G-Assist appropriately responds with all the important hardware information listed in the response. However, it did seem to have difficulty getting the valid resolution of our BenQ 4K monitor (i.e. 4K 60Hz), but other than that, it passes our initial sniff test.

Next, another (presumably) common use case is monitoring the GPU’s power consumption. We have more traditional telemetry up in the top-right corner, but it doesn’t provide a complete graph unless you have a third-party tool like HWiNFO64; therefore, in this case, the casual user might want to ask the chatbot for the information they need.

We posed three different questions to the Project G-Assist chatbot, with the first two being responded to without issue; that said, the third question did seem to be beyond its capabilities, as we initially wanted it to provide real-time monitoring when available. Instead, it gave us the current GPU power consumption.

It’s also worth noting that when the GPU is working hard to generate a response, it will use the vast majority of its available power, in which case our RTX 5090 FE was momentarily drawing over 350 Watts every time we issued a prompt to the chatbot. On older or weaker hardware, the time it takes to generate a response might take even longer (worst-case scenario is the RTX 3060 12GB as it is the lowest-end model with enough VRAM to access this feature), but in this case, we observed about half a second of "thinking" time before the response was generated.

Gaming and Performance

Let’s switch gears and take a look at gaming. If you have too large a game library in Steam to sift through, you can launch a game straight from the chatbot – assuming you somehow don’t have a game shortcut on your desktop or the Start menu (in this case, we didn’t even need to spell out the full name of Forza Horizon 5 for it to figure out which game to launch, though it is the only Forza game in our system).

Coincidentally, a driver update may have messed with the settings in-game, causing FH5 to be stuck at a paltry 15 FPS. A beleaguered casual gamer might immediately slam the Alt+G hotkey and start asking G-Assist "what’s going on," but this is where G-Assist’s limitations lie: it lacks the ability to read in-game settings and instead provides a generic response providing some basic directions for the user to diagnose the problem.

Through manual diagnostics, we did find that the game had somehow switched its internal framerate limit to only 15 FPS, which G-Assist had no way of detecting. Its response showing "frame rate limiter disabled" presumably refers to NVIDIA’s driver-level settings in the NVIDIA App, but it’s very likely that a casual user wouldn’t be able to resolve this issue on their own and would likely be misled by this less-than-ideal response.

Next, let’s take it to Counter-Strike 2 to see if NVIDIA can figure out how to improve PC latency – a metric that competitive gamers have to keep an eye on, but not everyone can readily understand. Asking G-Assist to provide an average latency report is easy enough, but it failed to provide any specific recommendations to further improve this metric (and it gave us the same response we just saw in Forza Horizon 5).

That’s still fine since we are assuming that NVIDIA has marketed its features well enough that NVIDIA Reflex is a feature most FPS gamers are likely aware of. So, what happens if they can’t find the location of the option in CS2’s rather complex in-game settings, and choose to ask the chatbot? Unfortunately, it had absolutely no awareness that Reflex was actually enabled, instead telling us that it was disabled. I guess that’s why we were reminded to fact-check its errors.

Other Scenarios

In the next scenario, we’re probing the chatbot to see if it can figure out a way to enable RTX Video Super Resolution (RTX VSR), a video upscaling technology designed to increase the effective resolution and reduce compression artifacts in online videos such as YouTube and Twitch. Now, if you’re familiar with League of Legends, you’d know that sometimes a team fight can make the screen extremely cluttered and cause all sorts of visual artifacts to exist in blocky pixels; or in other cases, you want to upscale a 1080p stream to your 4K monitor.

To be fair, Project G-Assist did manage to figure out the feature we were looking for despite us not explicitly mentioning the name of the feature; but it had no ability to detect if the feature was already enabled. (Wouldn’t it be trivially simple for G-Assist to check NVIDIA App’s settings?)

Okay, so be it – perhaps we’ll just ask the chatbot to take us directly to the settings page to enable the feature, just to give it the best chance possible. That also didn’t work, with the chatbot providing no further suggestions and leaving any casual user to ask Google instead (which, given the state of things, will likely give them another AI-generated result).

Diving Deep into Project G-Assist: Can NVIDIA’s AI Assistant Hit the Mark?

NVIDIA’s Project G-Assist promises to leverage artificial intelligence to simplify PC management and enhance the gaming experience. Powered by Meta’s Llama-3.1-8B SLM, which runs locally, it aims to optimize system settings, monitor performance, and control peripherals through voice or text commands. While the concept is promising, the actual performance falls short of perfection.

Setup Hurdles: Hardware and Software Barriers

Setting up Project G-Assist presents several obstacles. First, the requirement for an RTX 30-series or newer GPU with at least 12GB of VRAM significantly limits its potential user base. This exclusion leaves out a substantial number of gamers with less powerful GPUs, including many RTX xx60 series owners. Additionally, the reliance on specific operating system versions and drivers adds complexity.

The supported peripherals are also limited to MSI motherboards and devices from Logitech G, Corsair, and Nanoleaf, further restricting the utility for users without these specific brands of hardware.

Real-World Performance: Mixed Results

In real-world testing, Project G-Assist demonstrated inconsistent performance across various tasks. While it could accurately retrieve system information and monitor GPU power consumption, it struggled with more complex queries. For example, it failed to recognize the correct resolution of a BenQ 4K monitor and had difficulty providing specific guidance on optimizing game settings.

Regarding gaming, Project G-Assist could launch games in Steam, but its usefulness in resolving performance issues was limited. When Forza Horizon 5 experienced framerate problems, G-Assist failed to diagnose the root cause and instead offered a generic response that wasn’t particularly helpful to the user. Similarly, in Counter-Strike 2, it failed to provide specific recommendations for reducing latency and even incorrectly reported the status of NVIDIA Reflex.

Missing Features and Limitations

The limitations of Project G-Assist extend beyond its inconsistent performance. It also lacks crucial features, such as the ability to read in-game settings and detect the status of RTX Video Super Resolution (RTX VSR). These omissions significantly restrict its utility as a comprehensive PC assistant.

Furthermore, G-Assist relies on a locally run language model, which means it requires substantial computing resources. During testing, the RTX 5090 FE consumed up to 350 Watts of power whenever the chatbot generated a response. This could pose performance issues for users with older or less powerful hardware.

Better Communication and Expectation Management

Given its current state, NVIDIA would do well to better communicate that Project G-Assist is still in a beta phase. Its limited functionality and inconsistent performance could lead to frustration for users who expect a more polished experience. By being transparent about G-Assist’s current capabilities, NVIDIA can set realistic expectations and avoid unnecessary negative feedback.

Future Potential: A Work in Progress

Despite its limitations, Project G-Assist still holds future potential. As AI technology continues to evolve, NVIDIA can improve the language model, expand its features, and optimize its performance. By addressing the current limitations and adding new capabilities, Project G-Assist has the potential to become a valuable tool for casual users. However, it has a long way to go before it reaches that potential.

As it stands, Project G-Assist feels more like a fancier, natural language version of the command console than a comprehensive PC assistant. While it may be capable of handling some basic tasks, it’s not yet polished enough to reliably resolve advanced issues or provide personalized guidance. Only through continued development and improvement can Project G-Assist truly deliver on its promise of simplifying PC management and enhancing the gaming experience.

Another important issue to address is the system requirements. Unless you have a fairly high-end GPU with 12GB or more of VRAM, you simply can’t use the feature - this effectively excludes all RTX xx60-series owners (unless you have an RTX 3060 12GB, RTX 4060 Ti 16GB, or RTX 5060 Ti 16GB), who make up a large portion of NVIDIA-powered PCs in many of the Steam hardware surveys we’ve seen in recent years. I do hope that the language model can be scaled down to fit 8GB or even 6GB of VRAM, or it won’t see widespread adoption unless NVIDIA starts installing more VRAM in GPUs from now on.