Nvidia RTX Spark Review: Hello, Agentic AI

Remember when the world was busy obsessing over ChatGPT and how the AI chatbot could respond to any question within seconds? It has been four years since that moment, and we are now at the brink of another revolution that will change the way we interact with our PCs forever. Yes, I am talking about agentic AI. In the simplest of words, it is when an AI agent isn't just limited to answering a few questions but can actually take action on your behalf. And Nvidia's RTX Spark chipset helps you run these agents locally, on your device.

Now even though I have been using AI tools for a while, my interactions with them have not really been interesting. I just ask a question, get an answer, and then do the actual work myself. But when I got the chance to try out a device powered by the Nvidia RTX Spark chip, I realised that I didn't really have to do the actual work at all. How? Read on.

What is Nvidia RTX Spark?

First things first, the Nvidia RTX Spark is not a discrete GPU or your conventional laptop chipset. It is a superchip, an Arm-based SoC that combines an efficient Grace CPU with up to 20 cores with an RTX Blackwell GPU, both built on TSMC's 3nm process. The two halves sit side by side on the package and talk to each other over 600GB/s of bidirectional NVLink C2C bandwidth, which is fast enough that the GPU can treat a large portion of the system's unified LPDDR5X memory as its own.

There are two memory configurations of up to 128GB. Now on a normal laptop, a local AI model is limited by the few gigabytes of VRAM its GPU can see. But here, a 128GB machine can hand a large chunk of that memory to the GPU, and suddenly models that would usually live on a workstation or in the cloud fit right onto your desk.

The platform comes in two variants, with more GPU and CPU cores in the top tier. I have been using the higher-spec variant of the RTX Spark with 6,144 CUDA cores.

Nvidia has already announced the first devices to be powered by the RTX Spark chipset. These include the Asus ProArt P16, Lenovo Yoga Pro 9, Microsoft Surface Laptop Ultra, Dell XPS 16, HP Omnibook Ultra 16, and MSI Prestige N16 Flip AI.

Now, on paper, this all sounds impressive. The real question, of course, is what it feels like when you actually use it. Without further ado, let’s get to it.

Running local LLMs

Any AI chatbot can tell you how to install a browser, import a video, or edit an audio track. But an AI agent can actually perform these tasks for you. To experience this myself, I installed Hermes Agent on a device powered by the Nvidia RTX Spark chipset and spent some time with it.

I decided to begin with something easy and asked Hermes to download Google Chrome and set it as my default browser. The agent opened the settings, made the change, and confirmed it. There was no copy-pasting or clicking through endless menus. Now, even though this was a small task, after using the RTX Spark I actually felt the difference between a chatbot that talks and an agent that acts And boy ain't that cool!

From there, I thought of getting even more experimental. I asked the agent to organise a folder of my downloads into subfolders by file type, and it went through the whole pile without me having to touch anything. The model first thought for 30 seconds, identified how many files it needed to organise, and then got on with the task. As someone who is super organised and is always decluttering her PC, this was a dream come true.

Also, this is where the 'running local’ part really starts to matter. The agent runs on models that are downloaded to your device. This means that your data doesn't leave the machine, you don't pay per-request cloud API fees, and after the models are installed, the whole thing works offline. That is a very different experience from the cloud AI tools most of us are used to.

The RTX Spark also changes what you can build around these agents. Because the GPU-accessible memory pool is so large, you can fine-tune a model on the same laptop. Basically, a model can learn your documents, terminology, and response style. You can also run image and video generation locally through ComfyUI, chain upscaling, frame interpolation, and HDR in one pass, and iterate without a single cloud generation fee.

That said, memory capacity is only one part of the equation. Model performance also depends on the model itself, its quantisation, the software stack and the workload. RTX Spark gives you the hardware to experiment, but the experience still depends on how well the different pieces work together.

Using creator tools with Nvidia RTX Spark

The RTX Spark pairs a full set of RTX-accelerated creative applications with local generative AI, and both sides get a boost from the unified memory.

On the generative side, ComfyUI ships with new nodes that use RTX Video Super Resolution, AI frame interpolation, and real-time SDR-to-HDR conversion, so you can generate a clip at a lower resolution and upscale, interpolate, and tone-map it in one local pass.

Adobe is also re-architecting Photoshop and Premiere for the RTX Spark to be fully GPU-accelerated, and the beta versions already show what that looks like. Premiere Pro runs TensorRT-accelerated features on-device, including AI object masking that tracks a subject through a shot, Enhance Audio for separating dialogue and background noise locally, scene edit detection, and auto-reframe for repurposing footage across aspect ratios.

Meanwhile, Photoshop's on-device remove tool can paint out unwanted objects using a local model, with no cloud round trip. DaVinci Resolve also has a native Arm build, and once you point it at the Nvidia GPU and let it run its first-time AI model optimisation, the AI effects run on the RTX hardware.

I wanted to see how this works in practice, so I installed DaVinci Resolve on my device (well, actually got Hermes agent to do it). I then asked the Hermes Agent to import a video into Resolve and then import a music track. Once both were available in the project, I instructed the agent to trim the music to match the video's length. Hermes managed to complete that task easily, even though the process took around five minutes.

Now, I already know how to import files into Resolve and adjust an audio track on the timeline and would have probably done it faster than the agent (no offense). But, it is still amazing to be able to describe what I wanted and watch the agent carry it out. For creators who are new to the world of video editing, this can be incredibly helpful.

I also tried out the Nvidia Broadcast tool, and it really can make a whole lot of difference in how you appear and sound on-screen. With tools like background blur, Auto Frame, Video Noise Removal, Video Super Resolution, Video Frame Generation, and Virtual Key Light, there is a lot you can do here. Virtual Key Light brightened my face while Auto Frame helped keep me in the frame. Background blur helped separate me from my surroundings, and Video Frame Generation made the video appear much smoother.

You can also improve your audio through this tool. For example, the Noise Removal tool can reduce unwanted background sounds, and Studio Voice will make your voice sound cleaner. So, if you are someone who records videos at home or regularly appears on camera, these tools could help improve the final output without expensive audio equipment.

(Representative image)

I also attended a demonstration from Nvidia, where the Hermes Agent tackled a far more ambitious video-editing task in DaVinci Resolve. The agent was given a 90-minute recording of the recent Microsoft Surface event. Its task was to find the portions where Nvidia RTX Spark was mentioned and bring those moments together into a shorter video.

Hermes produced a clip of approximately six minutes and finished the job in a short time. This means that if you are a video editor, you can simply describe what you want shortlisted from a video and leave everything else up to the agent.

Gaming experience on the Nvidia RTX Spark

Now, a machine this powerful can obviously run your favourite AAA titles. And it can run them quite well. During an initial demo of the RTX Spark at Computex earlier this year, I got a chance to test Pragmata and Alan Wake 2 on the devices. And I couldn’t believe my eyes. Both of these fairly demanding games were running like a breeze on the super thin and light laptops.

Now that I have a device for testing, I fired up the recently-launched Control Resonant. Running at 2880x1800 resolution with graphics set to Ultra (yes, you read that right) and Ray Tracing off, the game was giving me a stable 50 FPS. It not only looked gorgeous, but was also resulting in a super smooth experience. Do note this is without any frame gen technologies.

When I turned frame gen on and set it to 2x, I started getting 100 FPS and an even smoother experience. When I decided to turn Ray Tracing on, the FPS did drop to 50 but the game was still very much playable.

I then fired up 007 First Light launched a couple of months back. At 2560 x 1440 resolution with high graphics settings, DLSS Super Resolution set to Balanced and 3X frame generation enabled, I was getting around 139 FPS. That is a solid number for a laptop this thin, and the game actually felt super smooth. One thing to note is that I did leave Path tracing off in this scenario.

Then I moved to Cyberpunk 2077, the one game that can bring the best of gaming laptops to their knees. With the same 2560x1440 resolution, path tracing turned off, ray tracing turned on, and field of view set to 80, I was getting around 135 FPS and a smooth experience throughout.

Here is the gaming experience, at a glance:

So as far as gaming is concerned, the Nvidia RTX Spark can do wonders in a thin and light machine that you can just carry around anywhere.

PS: These results are from just one particular RTX-powered machine and may vary as per the machine you’re using. Moreover, I tried out all of these games on battery, without the device being plugged in.

Verdict

The NVIDIA RTX Spark is likely to introduce agentic AI to many users. Its unified memory architecture gives compatible AI models access to a large memory pool, making it easier to experiment with local AI workloads that can be difficult to run on conventional PCs. As Nvidia says, this chip brings together AI development, creative applications and gaming in a compact platform designed with local AI at its core.

So for AI enthusiasts, it enables them to run AI models locally. For creators, there are a number of AI-powered tools at their disposal, and for gamers, there is a possibility of running their favourite AAA titles on a fairly thin and light machine.

But this is still a first-gen chipset, which means that the experience will also depend on the machine you are using it on. As for me, I was overall pleased and look forward to tinkering with the device as I spend more time with it.

Perhaps the most important thing to note is that the RTX Spark chip isn’t just interesting for what it can do today, but also for the possibilities it opens up for the future. For developers, creators, and anyone who really wants to experiment with local AI, it is something that will keep them busy for hours.