The GPU spent a decade as the only chip gamers cared about. Then local AI happened, and a second processor quietly became the spec worth reading: the NPU, a dedicated neural engine that runs AI workloads at a fraction of the power the GPU would burn doing the same job.

For portable Windows PCs this matters twice over. Once for the AI itself: local models, private inference, no subscription, no cloud. And once for gaming: every AI task the NPU absorbs is power and GPU headroom handed back to your frame counter. In 2026, three portable machines define what a high-performance NPU setup looks like, including one so new it's still in crowdfunding.

Why the NPU Suddenly Matters

A quick grounding, because "TOPS" numbers are easy to print and easy to misread.

An NPU is silicon purpose-built for neural network math. Its virtue isn't peak power, the GPU usually has more raw compute, but efficiency: it runs sustained AI workloads at single-digit watts, which on a battery-powered device is the difference between a feature and a gimmick. Background tasks like live transcription, image processing, copilots, and small-model inference belong on the NPU precisely so they don't touch your battery curve or your GPU.

The gaming connection follows directly. On a portable PC with a shared power budget, an AI workload running on the GPU steals watts from your frames. The same workload on the NPU costs almost nothing from the gaming budget. A high-TOPS NPU is how one device runs both halves of its life, AI by day, AAA by night, without the two halves fighting.

Platform AI figures on modern chips combine NPU, GPU, and CPU contributions, which is why you'll see both an NPU number and a total below. Both matter: the NPU number for efficient always-on AI, the total for peak inference throughput when everything pulls together.

The New Benchmark: ONEXPLAYER 3

The ONEXPLAYER 3 is the newest entry in this conversation and currently its AI compute leader: a 50 TOPS dedicated NPU inside a platform delivering 180 TOPS of total AI performance, built on Intel's Arc G3 Extreme, the Panther Lake gaming flagship on the Intel 18A process.

The gaming half of its résumé is already independently measured. As tested by ETA Prime, the Arc B390 leads the previous handheld graphics class (Radeon 890M in the Lenovo Legion Go 2, both at 35W) by 97.9% in 3DMark Steel Nomad and 71.9% in Time Spy, and runs Cyberpunk 2077 at around 57 FPS at just 17W. XeSS 3 multi-frame generation takes titles from 20-30 FPS native to roughly 70-90 FPS. The AI half and the gaming half share an 85Wh battery that measured about 10 hours of light gaming at 5W, which is exactly the kind of budget an efficient NPU is designed to protect.

One transparent status note: the ONEXPLAYER 3 is currently funding on Indiegogo and is not yet listed on the official ONEXPLAYER store, with retail pricing not yet announced. It's the benchmark this category is about to have, not the box you can order this afternoon. For that, read on.

Shipping Today: ONEXPLAYER Super V

The ONEXPLAYER Super V is the machine that turned the NPU pitch into a working product you can buy now, and ONEXPLAYER labels it accordingly: a Mobile AI Workstation.

Its Intel Core Ultra X7 358H pairs the same 50 TOPS NPU class with a 122 TOPS Arc B390 GPU for 172 TOPS of total AI compute, and the software side is the differentiator: ONEX AI comes pre-loaded with one-click local model deployment. No environment setup, no dependency hunting; the tablet runs private, offline AI from the first boot. Around that core sits a genuine flagship: a 14-inch 2880x1800 AMOLED at 120Hz with 100% DCI-P3 and stylus support, quad-channel LPDDR5X-8533, and an 85.58Wh battery.

The gaming half holds up its end with the platform's signature efficiency: independently tested at 45+ FPS native in Cyberpunk 2077 at 1800P High, around 110 FPS with XeSS 4x Frame Generation, and 60+ FPS in Resident Evil Requiem at 1800P while drawing only about 40W. Day job and night job, one 12.5mm slab.

The Capacity Route: ONEXPLAYER Super X

The ONEXPLAYER Super X answers the NPU question from a different direction: what if the constraint isn't AI compute, but AI memory?

Its AMD Ryzen AI Max+ 395 carries a 50 TOPS NPU within 126 TOPS of total AI compute, respectable numbers, but the headline is the memory system: up to 128GB of quad-channel unified memory at 256 GB/s, with up to 96GB allocatable as VRAM. That capacity crosses a hard line no conventional portable crosses: 70B-parameter language models, which physically cannot fit on typical 16GB or 32GB discrete GPUs, load and run on this tablet.

For most local AI use, compute is the bottleneck and the Intel machines' higher TOPS totals serve it. For large-model work, capacity is the bottleneck, and the Super X is the only machine in this article, and one of very few portables anywhere, that clears it. Its gaming credentials match the ambition: a Radeon 8060S in RTX 4070 Laptop territory, around 60 FPS in Cyberpunk 2077 at High settings, and a liquid-cooled edition reaching up to 120W with the separately sold Frost Bay unit.

Match the Machine to the Workload

Pricing shown is subject to change; refer to the official ONEXPLAYER store for current pricing.

Conclusion

The portable Windows PC with a high-performance NPU is no longer hypothetical, it's a three-way choice. The ONEXPLAYER 3 sets the new compute ceiling at 50 TOPS of NPU inside 180 TOPS of platform AI, wrapped in the strongest handheld gaming silicon of 2026, currently on Indiegogo ahead of its store debut. The Super V is the answer you can order today, with 172 TOPS and local AI that works out of the box. And the Super X trades peak TOPS for the one capability the others can't offer: enough unified memory to run 70B models on your lap.

The NPU stopped being a spec-sheet footnote the moment local AI became a daily tool. These are the machines built for that moment, and they all game like it never happened.

Check current availability and configurations on the official ONEXPLAYER store！

FAQ

Can Windows handhelds and tablets really play large AAA games?

Yes. ONEXPLAYER devices have the hardware performance to smoothly run demanding titles like Black Myth: Wukong, Cyberpunk 2077, and Forza Horizon. Independent testing across the devices in this article shows current AAA releases at high settings, from 57 FPS at 17W on the ONEXPLAYER 3 to 1800P gaming on the Super V and Super X.

Do ONEXPLAYER devices support external monitors or eGPU docks?

Yes. These devices include full-featured USB4 ports that connect to a monitor, TV, or eGPU dock with a single cable, with the ONEXPLAYER 3 adding Thunderbolt 4 support. Adding an ONEXGPU dock raises image quality and framerates in AAA titles to desktop-class levels.

How long can an ONEXPLAYER device run on a full charge?

ONEXPLAYER devices ship with high-density, large-capacity batteries. A demanding title like Black Myth: Wukong runs for around 3 hours of continuous play on current models, and the ONEXPLAYER 3's 85Wh cell measured about 10 hours of light gaming at 5W in ETA Prime's testing. NPU-powered AI tasks are designed to sip power, so local AI work stretches battery far longer than gaming does.