I was in Taipei for Computex 2026 when Nvidia announced the RTX Spark, later having the opportunity to enjoy a showcase with practically every machine announced so far. It's a pretty big deal; after all, Jensen Huang called it the chip that would reinvent the PC, Microsoft treated it as the next big Windows on Arm moment, and every major OEM had either a laptop or a mini desktop ready to show.

On the surface, it actually did sound like a new kind of Windows PC: a 20-core Arm CPU, a Blackwell GPU with the same 6,144 CUDA-core count as an RTX 5070, and up to 128GB of unified memory. However, the RTX Spark isn't new silicon. In fact, it's the same GB10/N1 hardware that's already been shipping inside the DGX Spark and OEM systems like my Lenovo ThinkStation PGX, just repackaged for Windows PCs.

However, that's exactly why the gaming pitch is so interesting to me. Not only did I have a brief experience playing games on RTX Spark hardware at Computex, but I've already been using the GB10 chip for months in the PGX. Granted, that's been mostly for AI workloads like serving large local models and fine-tuning smaller ones, but I also tried gaming on it, too. The results caught me off guard, and they tell me more about the RTX Spark than a few minutes with a demo actually could.

The real changes are platform-level

Same silicon, different wrapper

When Nvidia announced the RTX Spark at Computex, it framed the chip as a consumer Windows-on-Arm platform for creators, developers, and AI enthusiasts. At the top end, though, the hardware lines up with the GB10 I already know: 20 Arm CPU cores co-designed with MediaTek, 48 Blackwell streaming multiprocessors with 6,144 CUDA cores, 192 fifth-generation Tensor Cores, 48 fourth-generation RT cores, and up to 128GB of LPDDR5X unified memory on a 256-bit bus. In practice, that memory bandwidth lands at 273 GB/s, since the shipping product runs the LPDDR5X at 8.533 Gbps. Nvidia has shown a 300 GB/s figure, but that came from an overclocked configuration running at 9.4 Gbps, demonstrated when the GB10 was unveiled at Hot Chips 2025. The 273 GB/s number is what you actually get.

The platform changes that were made are exactly the sort of things you'd expect when a Linux AI workstation becomes a Windows PC. The DGX Spark and similar machines have a ConnectX-7 networking chip and dual QSFP ports allowing you to link two boxes together over 200 GbE for distributed workloads. RTX Spark drops that for Wi-Fi 7 and 10Gb Ethernet, which makes far more sense on a consumer machine. At the flagship tier, the core hardware is unchanged.

On the GB10, you can evaluate the GPU's lineage in software. CUDA reports compute capability 12.1 for GB10, while Nvidia's datacenter Blackwell parts are in the 10.x compute-capability family. For developers, compute capability is Nvidia's public software boundary for supported GPU features and architectural limits, including things such as shared-memory capacity per SM.

As well, 12.x architecture used here is the same branch as Nvidia's RTX 50-series consumer GPUs, including the RTX 5070, which is strong evidence that the GB10 is not just a cut-down datacenter part with gaming features added back in. Instead, RT cores, DLSS 4, and frame generation are part of the platform's design. Obviously, the power budget and memory bandwidth are different from a desktop RTX 5070, but at 1440p, neither one is the wall I expected it to be.

I ran Cyberpunk 2077 at 1440p on an AI workstation

And it held 60 FPS through two translation layers

When I tested the PGX for gaming, I wasn't running games natively, either on an instruction set level or a platform level. The GB10 runs DGX OS, which is based on Ubuntu Linux compiled for aarch64. Steam and practically every game in its library are x86-64 binaries, so every calculation was passed through FEX-Emu to translate x86 CPU instructions to Arm while also translating Windows API calls and DirectX calls to Linux and Vulkan. That's two translation layers stacked on top of each other running on hardware that wasn't meant to run games at all.

The setup involved an autoinstall script published by an Nvidia employee, which configures FEX, downloads a full x86-64 root filesystem, installs Steam directly from Valve's repository, and sets up what FEX calls "thunking", intercepting OpenGL and Vulkan calls from emulated games and routing them directly to the native aarch64 GPU drivers. Without thunking, the GPU is useless and everything falls back to software rendering. With it, the GPU runs at near-native speed. I learned that one the hard way when I tried to set it up manually and Cyberpunk's main menu took several minutes to load while every CPU core maxed out.

Once it was working, I tested three games at 1440p on their highest presets. Counter-Strike 2 averaged 117 FPS and felt smooth throughout, with the translation overhead practically invisible. Granted, that game was running as a native Linux build, but it's a CPU sensitive game, and those calls were still being translated. Cyberpunk 2077 averaged around 60 FPS at 1440p with settings maxed out, which won't set any records but held consistently through combat and exploration. DOOM Eternal was the standout, running between 140 and 170 FPS thanks to id Tech's famously absurd optimization. In that instance, if I didn't know it was running through two translation layers, I wouldn't have been able to tell.

The power draw was especially impressive. The GB10 SoC is rated at 140W for the CPU and GPU combined, inside a 240W total system budget, but during Cyberpunk the GPU pulled well under 100W and total system power stayed absurdly low for what it was doing. For context, the desktop RTX 5070, the card with the same CUDA core count, has a 250W power limit all by itself. The entire PGX, CPU and GPU together, used less power while gaming at 1440p than many discrete graphics cards draw on their own. Thermals stayed reasonable too, with the GPU averaging in the mid-70s and the CPU peaking in the low 90s under sustained load. I didn't spot any meaningful throttling, and the fans stayed quiet.

Worth keeping in mind, though: that 140W is a desktop figure. The flagship N1X laptops are reportedly built for a 45W to 80W package, so they're running this same silicon at roughly half to 60% of the budget the PGX gets... with less cooling headroom on top of that. My power and thermal numbers are a desktop result, and the laptop SKUs will almost certainly be more constrained.

DLSS changes the result completely, and translation penalties still exist

It helps smooth over gaps

The weird thing about GB10 gaming is that drivers, DLSS, and translation layers decide a lot of the experience. Early Cyberpunk 2077 tests on the DGX Spark showed people getting around 50 FPS, similar to what I saw, except those tests were at 1080p medium settings through Box64. That sounds strange until you look at where the bottleneck is; a lot of the translation overhead hits the CPU side, while the GPU is still handling the actual rendering through Nvidia's native graphics stack. Raising the resolution and graphics settings shifts more of the workload onto the GPU, so the performance gap doesn't scale the way it would on a conventional x86 gaming PC. In other words, 1080p medium and 1440p high can end up closer than expected because the limiting factor is not always raw shader throughput or memory bandwidth.

That's also where DLSS changes things, too. It doesn't remove the CPU-side translation cost, but it can make the GPU side much easier to manage once you're targeting higher resolutions or heavier visual settings. On the GB10, the graphics stack, translation layer, and upscaling path are all inherent to its design.

The contrast with AMD's Strix Halo is pretty interesting, too. In Cyberpunk at 1080p with the RT Ultra preset, Strix Halo manages only around 45 FPS even with FSR 3 upscaling in quality mode, because it doesn't have Nvidia's RT hardware or DLSS stack. AMD shipped its version of this hardware class a year before Nvidia, and on CPU-bound or native-x86 workloads, Strix Halo has genuine advantages. For gaming specifically, the lack of DLSS is a pretty unfortunate moat that FSR doesn't totally cross.

Because this is an Arm chip, there's still a translation penalty, no matter how small it may be, as nearly every PC game is compiled for x86. FEX-Emu can only do so much, and in CPU-bound titles it's likely that Strix Halo will compete favorably here. With that said, anything hitting the GPU should perform very well. That's why my 1440p results were as good as they are, because cranking the resolution and settings really only increases the workload of the GPU, where FEX's thunking keeps things running at near-native speed.

I also checked whether the GB10 appears to have anything like Apple's Rosetta-friendly memory-ordering trick, and at least on Linux, it doesn't look like it. Nvidia's own DGX Spark documentation describes the chip as a conventional Arm system with weaker memory ordering than x86, and my ThinkStation PGX reports standard Cortex-X925 and Cortex-A725 cores with no exposed TSO feature. A userspace request for TSO memory ordering was rejected by Nvidia's Linux kernel, and an AArch64 IRIW litmus test produced weak-memory outcomes that x86-style TSO would forbid. That doesn't prove there isn't some RTX Spark-only path somewhere, but there's no visible evidence that the GB10 has Apple-style hardware assist for x86 translation.

Regardless, when it comes to the RTX Spark running on Windows, translation should still improve somewhat. Microsoft's Prism emulator does on Windows what FEX does on Linux, and Microsoft says it tuned Prism specifically for this chip. Oh, and you also lose the Proton layer entirely, so there's only one translation layer instead of two.

Nobody wants to call it a gaming machine

And that might be the right call for now

I couldn't help but notice that, at Nvidia's Computex showcase, the branding choices made by the companies were interesting. Every single RTX Spark machine wore a creator or productivity badge. ASUS went with ProArt instead of ROG, Dell chose XPS over Alienware, HP used OmniBook rather than Omen, Lenovo picked Yoga instead of Legion, and MSI stuck with Prestige rather than any of its gaming lines. Nvidia spent plenty of time talking up 1440p gaming, but nobody looked at a 5070-class GPU sitting in their first-wave hardware and decided to call it a gaming laptop.

There are two ways to read that. The pessimistic read is that the OEMs don't trust Windows on Arm gaming to be fully baked on day one. Prism emulation, driver maturity, and anti-cheat support are all moving targets, and the last thing anyone wants is a gaming badge on a machine that struggles through the first wave of compatibility patches. The more pragmatic read is that these machines are going to be expensive. The DGX Spark hit $4,699 after its price hike, top-spec RTX Spark laptops are estimated at a minimum of $3,000, and the early adopter audience for a $3,000-plus Arm laptop skews creator and AI developer, not gamer. Either way, the branded gaming push may come in a later wave, especially given that Nvidia has committed to several generations of these chips.

The anti-cheat situation reinforces the cautious approach. Riot is bringing its games to Windows on Arm, PUBG and Fortnite will be too, and Nvidia is working with developers across the usual anti-cheat and DRM stack. That's promising, but "working with" is not the same as "everything works at launch." If I'm an OEM, I'm not putting an Alienware or ROG badge on a machine until I know the games people expect to play on those brands will actually run.

Still, none of that changes what the silicon can do. The GPU doesn't care what the branding of the machine is, and the fact that these machines are launching under creator brands doesn't mean they can't game. But it does imply that the companies selling them would rather under-promise in that department and let the performance speak for itself.

Gaming performance isn't the open question

Price and compatibility still are

The RTX Spark arrives in the fall with no announced pricing to speak of yet... but if the DGX Spark's trajectory is any indication, it won't arrive cheap. The DGX Spark launched at $3,999 in October 2025 and was raised to $4,699 four months later. OEM clones like the Dell Pro Max with GB10 run closer to $5,000. The 128GB LPDDR5X memory that makes these machines interesting for AI workloads is the same memory that's been constrained enough to trigger massive price hikes, which likely explains why Nvidia is also planning RTX Spark configurations that start at 16GB of memory. Unfortunately, though, you lose a big part of what makes the platform interesting at that tier: the ability to run large local models while still having a large unified memory pool for everything else.

On the Linux side, there are still open questions. Nvidia couldn't comment on whether the RTX Spark will support Linux, but my read is that the silicon appears to be practically identical to the DGX Spark, and that runs Ubuntu natively. Booting Linux on RTX Spark hardware should be possible, though it may require some work, and whether Nvidia actively supports it or leaves it to the community is unknown. Given that I ran all of my gaming benchmarks on Linux, and Canonical is now building an official Arm64 Steam Snap with FEX-Emu bundled in, the Linux gaming path is there even if Nvidia doesn't officially bless it.

When Nvidia says the RTX Spark is "Game Ready," I believe it. I've run Cyberpunk 2077 at 1440p on it, Counter-Strike 2 at 117 FPS, and DOOM Eternal north of 140 FPS, all on a machine drawing less power than a mid-range desktop GPU. You can even get DLSS 4 and path tracing up and running, and that's all on Linux, not on Windows where it's designed to run in the first place.

With that said, there's an important caveat: desktop RTX Spark mini-PCs use essentially the same GB10-class silicon and power envelope as the PGX I tested, so for those, the gaming performance isn't a promise I'll have to wait and see the results of. For those, it's a chip I've already gamed on, just wearing a new name and a Windows license. The laptops are a different story. It's the same architecture, but squeezed into a smaller thermal package with less cooling, so my expectations are grounded more in extrapolation rather than something actually measured. I expect them to game well, and from my brief time testing some games out in a controlled environment at Nvidia's showcase at Computex, it seems they do.

Either way, the price, the anti-cheat timeline, and the Prism quality are still open questions. For desktop Spark, gaming performance isn't one of them. For the laptops, it's the most answerable unknown on the list.