This $7,000 RTX Spark laptop had no business beating my RTX 5090 — but it did
· pcworld.com · popularity 19/100 · 0 similar stories today · Consumer Electronics & Hardware (Technology & Consumer Electronics)
Read the original at pcworld.com →
Summary
At a glance Expert's Rating Pros Massive 128GB unified memory pool enables huge local AI models Beat my RTX 5090 desktop in a demanding, memory-heavy AI workload Remarkably powerful AI capabilities in a thin, 3.9-pound laptop Gorgeous 4K OLED display and premium hardware Run enormous AI workloads without per-token API charges Cons Eye-watering $7,000 price Battery life craters during sustained AI inference Fans get extremely loud under heavy workloads Local models still trail top frontier AI in some tasks Overkill for gaming, everyday computing, and smaller AI workloads Our Verdict Nvidia’s RTX Spark isn’t going to replace a GeForce RTX 5090 for most workloads, but give it a massive AI model and suddenly its 128GB of unified memory becomes a superpower. Asus’ $7,000 ProArt P16 chewed through my enormous local AI research workloads and even finished one nearly 20 minutes faster than my monster RTX 5090 desktop. That’s genuinely remarkable for a 3.9-pound laptop. The catch? It’s wildly expensive, the fans scream under sustained AI loads, and battery life gets absolutely demolished. Local models still can’t match the best frontier AI at everything, either. But if you regularly work with giant models that overwhelm conventional GPU memory, RTX Spark enables something special — it’s a true game-changer worthy of praise. For everyone else, seven grand buys an awful lot of cloud AI. Price When Reviewed This value will show the geolocated pricing text for product undefined Best Pricing Today Best Prices Today: ProArt 16 Retailer Price Micro Center $6999 View Deal Price comparison from over 24,000 stores worldwide Asus sent me a $7,000 laptop with Nvidia’s new AI-focused RTX Spark chip and 128GB of unified memory. Fortunately, I had something suitably stupid to do with it: run an absurdly compute-hungry AI sports-research project I’ve been building in my spare time. A serious run can chew through hundreds of thousands of tokens, most of which would traditionally happen in the cloud. Nvidia’s RTX Spark chip was built to run enormous AI models locally, combining a 20-core Grace CPU with integrated Blackwell graphics and, in this no-holds-barred ProArt P16 configuration, a monumental 128GB pool of unified memory. So I spent several weeks moving as much of my real workload as possible onto the ProArt. I tried local models, screwed around with prompts and configurations, broke things, fixed things, and then put RTX Spark head-to-head with the GeForce RTX 5090 in my monstrous desktop. I’ve reviewed PC hardware for more than 15 years, and this was easily the most intensive review I’ve ever done. The stakes were simple: Could a delightfully thin 3.9-pound laptop actually handle the giant models and ridiculous research chains I normally run in the cloud? The answer was messy, expensive, and occasionally involved scrambling for a charger while Windows screamed about the battery. It also produced one hell of a surprise: The ProArt finished my biggest AI workload almost 20 minutes faster than my RTX 5090 desktop. Sweet holy moly. But we’ll get to that. This is ostensibly a review of the ProArt P16, but I’m not going to spend much time reviewing the laptop itself. I already covered the design and hardware of this laptop in depth , while Mark Hachman reviewed the RTX Spark chip inside Microsoft’s Surface Laptop Ultra as a productivity and general-purpose laptop, and Michael Crider put HP’s RTX Spark laptop through its paces as a gaming machine . RTX Spark has advantages for both, but at the eye-watering prices these things command, you shouldn’t buy one primarily for either purpose. No, I wanted to find out what happens when you take Nvidia’s ridiculous AI laptop and actually hammer the absolute crap out of it with AI. The results were (A) eye opening. What the hell do you do with 128GB? RTX Spark’s big trick is memory. Even Nvidia’s mighty GeForce RTX 5090 comes with “only” 32GB of dedicated GDDR7, while this ProArt P16 has a whopping 128GB of LPDDR5X in one unified pool shared between its 20-core Grace CPU and integrated Blackwell GPU. You can manually adjust the unified memory split between System and GPU memory in the MyAsus utility, though it won’t let you devote more than 96GB to the Nvidia graphics. Brad Chacos/Foundry That confused the hell out of me at first. I fired up Windows expecting to see 128GB of RAM and found roughly half that amount instead. Huh? It turns out Asus ships the ProArt with the unified pool partitioned roughly evenly between normal system memory and memory available to the GPU. That’s sensible for normal laptop use, but normal laptop use isn’t why Asus sent me this thing. You can adjust the split, and I eventually devoted roughly 96GB to the GPU for my heaviest AI workloads. That much memory lets suitably quantized 120-billion-parameter models run locally on a 3.9-pound Windows laptop. Yeah, duh. Look at the specs. I cared a lot more about whether those models could do useful work. I spent hours throwing pieces of my research workflow at different models and configurations. Some were better at reasoning through problems, but more prone to making things up. Others were better at reliably following my complicated instructions and spitting out structured results I could feed into the next stage. Context handling varied. Quantization mattered. Models that looked great during a quick test sometimes looked a lot less impressive when I compared their conclusions against stronger frontier AI. Once I allocated 96GB to the GPU, the lower system memory amount is shown in Task Manager. Brad Chacos/Foundry I also ran into failed runs, truncated outputs, context problems, and plenty of other weirdness. I reworked prompts, changed configurations, and tried different models. At one point I spent entirely too much time optimizing which model should handle which part of the workflow before realizing I was optimizing the test instead of actually reviewing the damned laptop. Eventually I kept coming back to gpt-oss-120b. It wasn’t always the smartest model I tested, but it followed my instructions reliably and didn’t make up nearly as much stuff. So I stopped screwing around and started throwing real work at it. 220,202 tokens in 13 minutes One of my NFL research processes starts by sweeping across the entire slate before deciding which players are worth investigating more deeply. I handed gpt-oss-120b an entire 14-game NFL slate containing 403 potential candidates. Thirteen minutes later, it was done. The model surfaced 198 players for deeper consideration while chewing through 220,202 tokens in 778.885 seconds. On my laptop! Holy shit. Brad Chacos/Foundry And this wasn’t a benchmark I dreamed up for the review. I needed the work done. I’d handed a 120-billion-parameter local model hundreds of players and a ridiculous pile of actual research, wandered off for a few minutes, and came back to find more than 220,000 tokens worth of completed work sitting on my SSD. The ProArt sounded like it wanted to achieve liftoff while doing it — its fans absolutely scream during sustained inference with a model this big — but damned if it didn’t work. I also ran a frozen seven-part benchmark built from different pieces of my actual research workflow. RTX Spark completed all seven locally, generating another 150,995 tokens in 415 seconds, just under seven minutes. The battery dropped from 74 to 65 percent. The results weren’t as good as what I got from the strongest frontier models, but they were useful, so I started throwing nastier workloads at it. Longer chains exposed more problems. Some ran into context or output issues. Some stages needed different prompting. Some runs simply…died. I kept tweaking things until a corrected full NFL chain made it all the way through, starting with another 14-game slate and another 403 candidates. This time it surfaced 208 players while chewing through 216,524 tokens during the discovery stage alone in about 13 minutes, then rolled straight into the later stages. Yeahhhhhhh it’s a beefy AI load perfect for a beefy RTX Spark system. Brad Chacos/Foundry I’d stupidly started that run unplugged with only 38 percent battery remaining. Within minutes Windows started throwing power warnings, Battery Saver kicked in, and I realized this giant model was going to murder the remaining charge. Oh shit! I scrambled for the charger and managed to plug in the ProArt without interrupting the run while its fans continued screaming. Yes, you can run a 120B model locally on a thin-and-light laptop. You’ll want to stay within lunging distance of an outlet. After all this, I wanted to know how RTX Spark stacked up against something much more powerful. My desktop packs a GeForce RTX 5090. RTX Spark versus my GeForce RTX 5090 My desktop is a ferocious monster in its own right: a beastly Falcon Northwest Fragbox with a Core i7-14700K, 96GB of system RAM, and a GeForce RTX 5090 with 32GB of screaming-fast GDDR7. (I run it with a 4090, but swapped in an Nvidia-supplied 5090 for this testing.) The RTX 5090 has vastly more raw GPU horsepower than the integrated Blackwell graphics inside RTX Spark. I figured this would be a bloodbath. I froze one of my large research workloads and ran the same gpt-oss-120b model with the same prompts, inputs, stage boundaries, and inference settings on both systems. Live web and state acquisition were kept out of the timed portion so both machines were chewing through the same work. This picture is of a 4090 sitting on the Asus ProArt 16, because the 5090 was in my test desktop. Brad Chacos/Foundry The ProArt processed 392,181 prompt tokens and generated another 79,903 completion tokens — 472,084 tokens total — in 33 minutes and 13 seconds. My RTX 5090 desktop took 52 minutes and 51 seconds. Well damn, Sparky. That’s a hell of a flex. The ProArt finished almost 20 minutes earlier, making it roughly 1.6 times faster at completing this particular monster workload. The 5090 run generated somewhat more completion tokens, 92,270 versus 79,903, so don’t read this as a laboratory-perfect tokens-per-second comparison, and definitely don’t read it as “RTX Spark is faster than the RTX 5090 at AI.” It isn’t. Give the 5090 a workload that fits comfortably inside its 32GB of VRAM and its vastly greater horsepower can run wild. But gpt-oss-120b doesn’t fit comfortably inside 32GB. My desktop can lean on another 96GB of conventional system RAM, but the model is working across separate pools of memory. RTX Spark’s Grace CPU and Blackwell GPU share one gigantic pool, and I’d configured the ProArt to make roughly 96GB available to the GPU, though my test crossed over into system memory there, too. For this oversized workload, that architectural difference mattered enough for the much slower laptop to finish first. By a lot . The ProArt paid for it in battery life. I started the run at 100 percent; 33 minutes later it sat at 54 percent. But holy shit, that’s interesting. Nvidia stuck 128GB of unified memory in this thing so it could tackle enormous AI models, and I’d found a real workload that genuinely put it to use. This 3.9-pound laptop finished the job almost 20 minutes faster than the monstrous RTX 5090 desktop sitting beside it. The $7,000 question Of course, local AI isn’t free. This laptop costs 7,000 freaking dollars before we even start talking about electricity, depreciation, or the frankly stupid number of hours I’ve spent screwing around with models and prompts. But the scale of my workload had gotten pretty stupid, too. During one particularly insane week while I was deep in the weeds building and testing this system, I reconstructed roughly 90 million input tokens plus another 8 million output and reasoning tokens. Call it 98 million tokens in one week. That’s an outlier, and simply multiplying 98 million tokens by the price of some top-shelf frontier model would be disingenuous. A sane cloud version would route simple mechanical jobs to cheaper models, heavier routine research to more capable ones, and save the expensive frontier stuff for problems that actually need it. Brad Chacos/Foundry I don’t expect to burn 98 million tokens every week. Something in the range of 15 to 30 million per week looks much more plausible during normal active use, with roughly 20 million as a decent middle-of-the-road estimate. My earlier modeling put an intelligently routed cloud version of that workload somewhere around $400 to $500 per month at then-current API pricing, give or take quite a bit depending on the model mix, caching, and other optimizations. Those aren’t historical bills, and API prices change constantly. More importantly, spending $7,000 to avoid a few hundred dollars per month isn’t exactly a financial masterstroke. You can buy a hell of a lot of cloud inference for seven grand. But after I got the local workflow humming, I stopped caring so much about how many tokens I was using. If my initial sweep could look at 403 candidates instead of 50, why not? Challenge an analysis three different ways instead of once? Go nuts. Take repeated snapshots as new information comes in? Sure. The fans will howl and the battery will beg for mercy, but there isn’t an API bill attached every time I get curious. Right now I might get a snapshot in the morning and another later in the day. Running locally means I could potentially take one every hour instead. Instead of two snapshots, I get the whole movie of the available market. How quickly does information go stale? When does waiting for more information help enough to offset prices moving against you? How often does new information actually change a decision? Answering those questions well could mean running a lot more analysis than I’d ever consider economically reasonable in the cloud. Will any of it make the project better? I have no idea. More computation can just produce more AI bullshit if you’re not careful . That’s why you run the experiments, and now I can. So what do I actually use locally? Not everything. I spent a lot of this review trying to determine how much frontier AI I could replace with local models. Turns out I don’t really care. Brad Chacos/Foundry The best frontier models are still better at some things, sometimes a lot better. If I’ve just run some enormous experiment and I’m trying to figure out whether I discovered something useful or merely found an exciting new way to fool myself, give me the biggest brain available. Ditto for poking holes in an argument or figuring out what experiment to run next. I don’t need that brain grinding through hundreds of thousands of routine tokens first. So I throw the huge, well-defined jobs at local models, then bring the important results back to frontier AI when I need heavier reasoning. Sometimes that sends me right back to the ProArt with another ridiculous batch of work. Funnily enough, having this much local compute around might make frontier AI more useful to me, not less. Local models can lose the plot. But weirdly, getting a 120-billion-parameter model running wasn’t even the hardest part. Feeding it trustworthy information, keeping giant piles of research organized, and making sure nothing quietly broke between stages caused me so much more pain than the actual inference. Give a model bad data or bad instructions and it’ll very efficiently produce a mountain of garbage for you. None of this is turnkey. But it works. I’ve kept doing it. Should you buy an RTX Spark laptop? For most people? Hell no. The ProArt P16 itself is fantastic. I already went deep on the gorgeous 4K 120Hz OLED, premium chassis, surprisingly good speakers, and other everyday hardware , and cramming all this stuff into a 3.9-pound laptop remains impressive as hell. This thing feels like Darth Vader’s MacBook Pro. It’s the single most impressive Windows laptop I’ve ever personally touched, despite a few quirks with the space bar. But 7,000 dollars is 7,000 dollars. If you just want an awesome laptop, buy something much cheaper. If you mostly want to play games, definitely buy something much cheaper; Michael’s RTX Spark gaming tests found that quadrupling the memory from his 32GB HP test system to this 128GB ProArt made essentially no difference in games (and we don’t recommend Spark for gaming even in cheaper versions). And don’t spend seven grand simply because running a 120-billion-parameter model on a laptop sounds cool. It is cool. But then what? Brad Chacos/Foundry I had something for it to do. I threw entire NFL slates and hundreds of players at the ProArt. I watched it churn through more than 220,000 tokens in 13 minutes. I stupidly kicked off one huge run with 38 percent battery remaining and had to scramble for the charger when Windows started freaking out. I watched another run eat 46 percent of a full charge in 33 minutes. I dealt with busted outputs, models that weren’t good enough, and prompts that needed reworking. And then the damned thing beat my RTX 5090 desktop by almost 20 minutes. Not because RTX Spark is magically more powerful than an RTX 5090. Nvidia built this chip differently, for a different sort of work, and I finally found the sort of work where that difference really mattered. Turns out 128GB of unified memory isn’t just a ridiculous spec. Sometimes ridiculous is exactly what the job calls for. And if you deal with ridiculousness daily, the Asus ProArt 16 with its 128GB and RTX Spark inside might just be worth every penny — all 700,000 of them in this laptop’s case. Just make sure you keep your charger handy.
Found in: PCWorld. Rating basis: single outlet; outlet tier 2 (unlisted); 1 feeds; feed position 7.