NVIDIA’s RTX Spark has finally made its debut. I got to see the machine in person at the launch event held at COEX on the evening of the 16th, and it turns out you could buy one right there. I was sorely tempted to bring one home on the spot, but my wallet had other ideas.
To get straight to the point, RTX Spark was a much more interesting machine than I had expected. It combines NVIDIA’s own Grace CPU, a high-performance, 20-core ARM-based processor, with a Blackwell GPU. The GPU shares its roots with the existing GeForce RTX 50 series, and you can think of it as serving the same purposes.
The key to this product is that everything is integrated into one system. For a start, the CPU and GPU are connected not by a conventional PC bus but by NVLink, NVIDIA’s ultrafast interconnect fabric. This lets them exchange up to 600GB of data per second. If that number is hard to put into perspective, the 16-lane PCI Express (PCIe) connection typically used for a PC graphics card transfers around 64GB per second in one direction. Even by simple arithmetic, that is nearly ten times the bandwidth.
Memory is also bundled into the chip package, creating a unified memory architecture in which the CPU and GPU share the same memory pool. Memory bandwidth is around 300GB per second. Compared with the VRAM bandwidth of a dedicated graphics card, that figure might feel a little disappointing, but given the nature of the product, it is a perfectly understandable trade-off. The overall architectural direction is quite similar to Apple’s M-series chips.
DGX Spark’s Twin, with a Very Different Purpose
Some of you may have noticed already, but in hardware terms, RTX Spark is almost the same machine as DGX Spark, the compact workstation NVIDIA introduced for AI developers. They use the same chip, but there are a few crucial differences. DGX Spark runs Ubuntu Linux and can link multiple units together over a high-speed fabric to run in parallel. RTX Spark, on the other hand, runs Windows and operates only as a standalone machine.
That is because the two machines have different goals. DGX Spark is an ultracompact workstation designed from the outset for development environments, while RTX Spark is firmly a personal device. Its focus is on being a GPU-based PC optimized for GPU-centric workloads, from running personalized AI agents to gaming and everyday Windows tasks.
The biggest barrier to entry is, as you might expect, memory and the price that comes with it. With memory so expensive, RTX Spark starts at KRW 5 million, while the model with 128GB of unified memory reaches KRW 10 million. If gaming were your sole purpose, you would probably look at the 32GB model. I could somehow justify that configuration in the KRW 2 million range, or at least under KRW 3 million, but honestly, spending KRW 5 million just to play games is a lot to swallow.
Local AI’s Real Bottleneck: Capacity, Not Performance
But shift your attention to local AI, and the story changes completely. In raw GPU terms, RTX Spark is roughly in the RTX 5070 class, and that amount of compute is actually quite enough to run most AI models in a personal setup.
It is easy to assume that running local AI properly means you absolutely have to get an RTX 5090. Of course, the 5090 is ideal: it has tremendous GPU compute power and enormous memory bandwidth. But the crucial limitation is VRAM capacity. The 5090 has 24GB of memory, and a single card feels painfully cramped when you try to load today’s useful models in the 20B class, meaning 20 billion parameters, or above. You end up needing two cards for a 48GB setup before you finally have some breathing room.
By contrast, the unified memory architecture in DGX Spark and RTX Spark lets you load a model directly into the entire system memory pool. A capacity of 128GB is hard even to imagine on a conventional desktop graphics card. Realistically, if your budget is under KRW 10 million and you want to load a huge model on the scale of 128GB locally, there are currently just two options: Apple’s Mac Studio and NVIDIA’s RTX Spark.
You might find the GPU compute power a little underwhelming. But DGX Spark was conceived as a bridge for designing and prototyping large models before moving them to large-scale data centers. RTX Spark is closer to a personal all-rounder: a Windows machine for games and everyday work that also keeps agents and large language models (LLMs) running all the time.
Besides, the AI ecosystem and model optimization have been advancing at a remarkable pace lately. Models keep getting smaller, while what small models can accomplish improves by the day. We are now in an era when individuals can use AI for vibe coding on a local PC and even generate their own videos. The biggest reason people hesitated to get into this area was not GPU compute speed so much as limited memory capacity. With the arrival of unified-memory computers, that barrier is coming down. It may take a little longer, but things that could not run at all can now run.
Blackwell’s DLSS Advantage, and How Fast It Feels in Practice
In fact, watching it run at the event, I found the perceived speed quite impressive too. Generating a four- to five-second 1080p video locally normally takes somewhere in the 800-second range, so the job finishes in under 15 minutes. Anyone who has generated video locally will know that this is already quite fast.
But this chip’s roots are in Blackwell. That means it can use NVIDIA’s signature DLSS technology as it stands. Generate quickly at a low resolution, upscale with DLSS, interpolate frames, add lighting, expand the color range with HDR, and finally apply DLSS 5, and the story changes completely. Put it through that entire post-processing pipeline, and producing a 4K HDR video of the same length takes just three minutes. This is where the real power of a hardware-accelerated ecosystem becomes clear.
I had worried that the relatively low memory bandwidth of 300GB/s might make LLMs slow, but coding with an agent was more responsive than I expected. Of course, I will need to borrow a machine and test it more thoroughly later. At the event, though, I entered a prompt asking it to build a dashboard that reads email, assesses its importance, and organizes what needs doing today. It used the Qwen 3.8 27B model, and the result took less than four minutes.
There is no question that this is a top-tier machine for personal local AI. It may not be breathtakingly fast, but it is comfortably fast enough for practical use, and, above all, it imposes far fewer limits on which models you can run than an RTX 5090 does. With local AI, whether a model fits in memory at all, whether it can run or not, is a much more fundamental question than whether it runs quickly or slowly. With 128GB, there is practically nothing you cannot run.
There is no denying that it is expensive, but if you want to get seriously into local AI and explore it in depth, it is an excellent option. Getting this much memory and this level of GPU ecosystem in a single system is no easy thing. Ironically, beginners who are unfamiliar with model quantization and memory optimization stand to benefit most from having an absurd amount of memory: they can load and experiment with whatever models they want without running into so many restrictions.
The Economics of Local AI: Token Costs and Running Costs
Even so, you probably still think it is expensive. As I mentioned when I covered the Mac Studio before, this price sits in a curious place: depending on how you compare it with cloud AI token costs, it can look expensive or actually quite cheap.
If all you do is chat casually with AI and ask simple questions, a ChatGPT or Claude subscription at KRW 30,000 a month is enough. If you occasionally generate an image or a few seconds of video for fun or to explore an idea, there is nothing wrong with using free credits or waiting for your token allowance to reset.
But things change if you are a heavy user who really puts AI to work: repeatedly generating longer videos, producing lots of images, or putting an agent to work coding all day. Plans costing a few tens of thousands of won a month hit their limits almost immediately, and even a KRW 300,000-a-month Pro plan comes with restrictions. In professional settings, spending more than KRW 1 million a month on API and enterprise plans is quite common.
With local AI, by contrast, the initial hardware purchase becomes a sunk cost, and after that there is no stress about token billing at all. If you are a heavy user, try adding up just one year of cloud token costs. You will begin to see where the hardware might pay for itself.
Local AI also offers considerable peace of mind when you are working with code that raises security or privacy concerns, or when you want to escape cloud services’ censorship and generation restrictions. Running costs are very low too. RTX Spark has fairly tight power limits: the whole system runs at under 80W, while even a fully loaded GPU hovers around 60W. Run it at full load 24 hours a day for an entire month, and the calculation still comes to less than 60kWh. That works out to only around KRW 16,000 a month in electricity.
Being able to run the work you want for as long as you like without worrying about token costs or security, freeing yourself from model-size restrictions, and even getting portability with a laptop or small-form-factor machine are advantages unique to local AI that are hard to replace. Judged purely as a gaming PC, the price is clearly disappointing. But once AI’s role in your own workflow passes a certain threshold, this could actually be a tool that reduces your overall spending right now. It is expensive, but there is no real substitute, which is why the market is finding its own reasons to accept the price.
RTX Spark gives me plenty to look forward to. I briefly spoke at the launch with Lablup CEO Jung-kyu Shin, who became the first buyer there and paid for two units on the spot. He told me, “It basically means I can put a DGX in my bag and use it wherever I go, so there is no reason to hesitate.” I think that might be the most accurate assessment of this machine right now.
That brief taste at the event was not enough for me either, and I have a mountain of tests I want to run myself, so I have asked NVIDIA to lend me a review sample. If I get a chance to put one on my desk and give it a proper workout soon, I will return with a deeper look combining benchmarks and hands-on use. If you have any questions about RTX Spark, or a workload you would really like me to test, leave a comment or send me an email!
