NVIDIA Just Came for the MacBook Pro (RTX Spark vs M5 Max)
These are the first NVIDIA RTX Spark laptops: the ASUS ProArt P16 and the HP OmniBook Ultra 16, with a lot more coming into the studio. I put them up against the MacBook Pro 16 with the M5 Max. NVIDIA put a real RTX GPU and up to 128GB of shared memory into a Windows on ARM laptop, and they're aiming it straight at the Mac. But is it worth your money? It depends.
Quick Verdict
If you're buying an RTX Spark laptop just for gaming, you'll probably be disappointed. Same if you're buying it just for video editing, because it doesn't come close to the MacBook Pro there. Where it really competes with the M5 Max is local AI: it reads long prompts faster, it wins image generation, and it's the better machine for huge models with long documents. That's a pretty small group of people, but if it's you, you'll love these laptops.
What I personally love is that this is the first time Windows on ARM feels really good. Snapdragon got it started a couple of years ago, and with NVIDIA behind it now, I think the remaining app gap shrinks a lot faster.
The Machines
- ASUS ProArt P16 (RTX Spark)
- HP OmniBook Ultra 16 (RTX Spark)
- MacBook Pro 16 (M5 Max, 128GB)
For the bigger picture, I also compared them with the ASUS ROG Zephyrus G14 (RTX 5070 Ti, 12GB VRAM), the ASUS ProArt PX13, the Snapdragon X2 Elite ASUS Zenbook A16, and an M4 Pro Mac mini.
Local AI, Explained Quickly
AI works in tokens, which are basically pieces of words, so tokens per second roughly means words per second. People read about 5–8 words a second, so anything over about 10 tokens/s is writing faster than you can read.
AI does two jobs:
- Reading (prompt processing): it takes in your question, a PDF, or a whole codebase. This needs raw GPU power, which is where NVIDIA shines, and it decides how long you stare at a blank screen before the first word appears.
- Writing (generation): it produces the answer one word at a time, running through the whole model in memory for each word. This comes down to memory speed, and Macs have always been good here because of their fast unified memory.
The model also has to fit in memory. Think of the G14's GPU as a small, really fast desk with 12GB of VRAM: quick, as long as the model fits. The Mac, ProArt P16, OmniBook, and PX13 have a huge shared desk of up to 128GB that the CPU and GPU both use. That's why a 120-billion-parameter model only runs on the 128GB machines.
The short version: the RTX Spark laptops read faster, the Mac writes faster, and on big models with long documents they end up very close, with the ProArt and OmniBook ahead on the biggest model. I used llama.cpp for all of these tests.
LLM Results
Small model (9B), runs on everything
| Reading (tokens/s) | Writing (tokens/s) | |
|---|---|---|
| ROG Zephyrus G14 | 3,351 | 83 |
| ProArt P16 | 2,678 | 47 |
| MacBook Pro M5 Max | 2,353 | 86 |
All of them write faster than you can read, but the Mac and G14 are almost twice as quick as the ProArt.
Medium models (27B and 35B)
The G14 sits these out because the models don't fit in its 12GB of VRAM.
- On the 27B, the Mac writes 66% faster (27 vs 16 tokens/s), but the ProArt and OmniBook read a long prompt 28% faster.
- Paste in a 50-page document and the ProArt starts answering in 45 seconds, the Mac in 68, and the PX13 takes over 2.5 minutes.
- By the time the answer is finished, the ProArt/OmniBook and the Mac both land at 117 seconds. The ProArt gets going sooner and the Mac catches up while it writes.
- The 35B tells the same story: 33 seconds on the ProArt, 31 on the Mac.
Big model (120B), the RTX Spark's moment
Only the 128GB machines can load it. The Mac still writes faster, 96 vs 64 tokens/s, and both are way faster than you can read. But feed it a 50-page document and the ProArt finishes in 42 seconds against the Mac's 52, about 20% less waiting. If you work with big models and long documents, this is where RTX Spark really pulls ahead of Apple.
The PX13 runs everything the ProArt runs, just slower on every model; that same document takes it 105 seconds. And the Mac mini shows how much Apple improved this year: the M5 Max reads prompts 5.5× faster than the M4 Pro, because Apple added AI accelerators to the M5's GPU.
AI Images and Speech-to-Text
Image generation is where NVIDIA's software advantage really shows up. Here's how long you wait for each picture:
| SDXL | Flux | |
|---|---|---|
| ROG Zephyrus G14 | 20s | 35s |
| ProArt P16 | 30s | 24s |
| MacBook Pro M5 Max | 42s | 53s |
| ProArt PX13 | 71s | — |
| Mac mini M4 Pro | 3+ min | — |
In Flux, running the exact same workflow on every machine, the ProArt makes a picture in less than half the Mac's time. The G14 falls behind there because the model overflows its 12GB. NVIDIA's own compressed format, NVFP4, gets Flux down to about 14.5 seconds on the ProArt and OmniBook, and under 13 on the G14. The Mac can't use it.
Speech-to-text goes the other way. Transcribing a 47-minute recording takes the Mac 43 seconds, the G14 80, and the ProArt/OmniBook around 95, so the Mac is about twice as fast as everything else. The PX13 couldn't run it on its GPU at all.
Creative Apps: The Mac Is Clearly Ahead
Outside of AI, the Mac is clearly ahead, and the ProArt sits in the middle of the pack.
- Premiere Pro and DaVinci Resolve: the Mac scores about 1.7–2× the OmniBook and ProArt, and the G14 beats the ProArt too. If video editing is your main thing, I wouldn't buy an RTX Spark laptop yet. A laptop with a dedicated GPU like the G14 does it better for less money.
- Blender: the Mac is about 1.8× the ProArt on the Monster scene, and the ProArt and G14 trade places scene by scene.
- Photoshop: close enough that it doesn't really matter. The Mac wins, the PX13 and ProArt are neck and neck, and the G14 comes last.
CPU: A Little Disappointing
Single-core performance on the ProArt and OmniBook is lower than I'd like. In Cinebench the Mac crushes everything, and even the Snapdragon X2 Elite Zenbook A16 beats the ProArt and OmniBook by about 20% on single-core. Multi-core is actually really good: the ProArt beats the PX13 and the G14, and the X2 Elite is only about 4% ahead. Honestly, a Snapdragon CPU with an NVIDIA GPU would be the sweet spot.
In the Speedometer web-browsing test, the ProArt comes last. It's not terrible, but it doesn't keep up with Snapdragon or Apple. The Snapdragon Zenbook scores 51 against the ProArt's 37 on the same Windows on ARM, so that's on NVIDIA's chip, not on Windows.
Graphics and Gaming
In 3DMark Steel Nomad, the ProArt and OmniBook score around 3,000. That puts them right around a laptop RTX 5070 and about 2.6× the Snapdragon Zenbook, whose Adreno graphics just isn't there yet. They still trail the G14's RTX 5070 Ti, and the MacBook Pro is on top at 4,330, close to an RTX 5080 in a thin gaming laptop.
If a game is built for ARM, it'll run great. The problem is that most PC games are built for Intel and AMD chips. They'll usually still run through Microsoft's Prism translation layer, but that costs performance, and how much depends on the game.
Cyberpunk 2077 (ARM-optimized, so the best case), 2560×1600 High, no upscaling:
| FPS | |
|---|---|
| ROG Zephyrus G14 | 80 |
| MacBook Pro M5 Max | 67 |
| ProArt P16 | 57 |
| ProArt PX13 | 51 |
With upscaling on (DLSS on NVIDIA, FSR on AMD, MetalFX on the Mac), everything gets more playable, and the ProArt goes up to 72 FPS. At 1920×1200 it's very playable even without upscaling at 81 FPS. One odd thing: at 1200p, DLSS only bumped the ProArt to 87, while FSR took the PX13 to 98, so something other than the GPU seems to be holding it back at lower resolutions.
Marvel Rivals isn't optimized for ARM yet, but an ARM version is coming very soon. In the demo, at 2560×1600 medium with DLSS, I was getting about 60 FPS and below. It played fine and never felt slow, but it didn't make use of the high-refresh screen.
Anti-cheat, the big question: in Marvel Rivals it worked with NetEase's own anti-cheat. Valorant doesn't work at all, because its anti-cheat still isn't supported on Windows on ARM and the game won't even load. That was the biggest problem with Snapdragon laptops, and Qualcomm didn't have the leverage to get companies to fix it. NVIDIA does, and Riot is working on ARM support. Once that's sorted, I'd expect a game as light as Valorant to run really well on all of these.
Even with the translation hit, gaming is one place these beat the Mac, because you have a much bigger library of games to choose from. So this is a very capable 1080p–1200p gaming machine, but it's not a gaming laptop. If gaming is your priority, get the G14. You'll save a lot of money and get better performance. (And yes, it detected my printer and it worked, which used to be hit or miss on Windows on ARM.)
Battery, Power, and Heat
PCMark 10 doesn't run on macOS, so the Windows laptops ran PCMark's video test and the Mac looped a 1080p video until it died.
| Video battery life | |
|---|---|
| MacBook Pro M5 Max | 27h 39m |
| ProArt P16 / OmniBook Ultra 16 | ~13h |
| ProArt PX13 | 9h 12m |
| ROG Zephyrus G14 | 6h 21m |
So you're getting about half the battery life of a MacBook Pro.
I also ran a 27B model for 10 minutes straight and measured power at the wall. The Mac pulled 92W, while the ProArt pulled 145W, right at the limit of its 140W charger. The Mac does more than twice the AI work per watt; it's just a more efficient laptop.
Do the RTX Spark laptops lose performance under sustained load? Not at all. The ProArt and OmniBook didn't slow down over those 10 minutes, while the Mac dropped about 20%, from 30 to 24 tokens/s. But the Mac starts so far ahead that it's still faster after slowing down: 24 against the ProArt's 16.5.
Unplugged, nobody slowed down. The ProArt kept full AI speed on battery and used 12% of its battery in 10 minutes. The Mac used 13% while writing about 45% faster, and the PX13 used about 17%. You get full performance on battery, which is great to see. They also stay pretty quiet, far quieter than thicker gaming laptops, they don't get super hot, and the cooling is more than adequate.
64GB vs 128GB
There's no performance difference between them: every test came out within about 3%, so for editing and gaming it doesn't matter. Only buy the 128GB model if you want to run something like the 120B model, which needs about 63GB on its own. You can change how much memory the GPU gets with a slider in the NVIDIA app (it needs a restart), which gives you more room for bigger models. On a Mac you don't have to touch a setting. For everyone else, the 64GB model is the same laptop for less money.
Price and Build
These aren't cheap Mac alternatives. At 128GB, the configs I tested all cost about the same: the OmniBook Ultra 16 fully specced is $6,229, and the 128GB MacBook Pro runs about $6,700–$7,000 depending on where you buy it and whether you get the 14- or 16-inch.
I don't think everyone's going to ditch their MacBooks for one of these, but they're some of the nicest Windows laptops I've ever used. I've seen the OmniBook Ultra, the Lenovo, the Dell, and the ProArts, and they're gorgeous. The ProArt P16 is just over 4 lb, lighter than the MacBook Pro. The OmniBook is a bit heavier, but also really beautiful. They all come with gorgeous OLED displays, CNC aluminum bodies, haptic touchpads, and lots of ports. This is where Windows on ARM should have been two years ago, and I'm glad it's finally here.
Who Should Buy One
I'm not convinced everyone is going to run local AI on their laptops. Personally, I'd probably run it on a box in my studio and connect to it from my laptop, and most developers I've talked to keep their open-source models on a cloud server or a machine in their office. If you do want local AI on the go, right next to everything else you're doing, NVIDIA has made it really easy to install tools like Hermes and OpenClaw. For an agent that should always be on, though, an always-on machine is still the better home.
- Gamers first: skip it. There are better Windows laptops for a lot less money.
- Video editors: skip it. The G14 gives you better video performance and much better gaming for less.
- Local AI enthusiasts: you have to really love and want to use local AI to spend this kind of money, but as a first try, this is really good, and I'm surprised how polished it is.
I'd rather carry one of these and take the hit on video performance just to use a laptop this nice. I just hope NVIDIA scales it down into something cheaper, because a lot of people would love it. As of today, if you don't care about gaming, you're still better off with a MacBook Pro 16. But if AI matters to you at all, you want to game a little on the side, and you don't mind the performance trade-offs, these RTX Spark laptops are incredible.
The big winner out of all this might actually be Qualcomm. With NVIDIA's leverage, more apps are going to work on Windows on ARM, so even Snapdragon laptops from two years ago are about to get better.
Where to Buy
- HP OmniBook Ultra 16 — Buy here
- ASUS ProArt P16 — USA · Canada
- ASUS ProArt P14 — Buy here
- MacBook Pro 16 — Buy here
Affiliate Disclosure: This article contains affiliate links. If you purchase through these links, we may earn a commission at no extra cost to you.
Published: October 2026
