The GPU claim survived contact with benchmarks

Start with what NVIDIA got right, with one caution about the scores themselves. In the TechTimes review roundup, the N1X scored 6,060 in 3DMark Time Spy Extreme, ahead of AMD's Ryzen AI Max+ 395 (Strix Halo) at 4,689 and Intel's top mobile part at 3,226. Trusted Reviews separately measured 11,763 in the standard Time Spy test. These are different benchmarks and the two figures should not be compared directly; together they support strong integrated-GPU performance, not universal leadership — Trusted Reviews also found discrete RTX 5070 gaming laptops faster in some of its tests.

Games still hold up. Cyberpunk 2077 averaged 69.73fps at 1080p on the ProArt P16, and 47.12fps at QHD. Turn ray tracing to its Ultra preset and it fell to 30.63fps at 1080p and 15.31fps at native resolution. That drop is the first place the machine's real limit shows up.

Two bandwidth numbers, and only one of them feeds the GPU

NVIDIA's materials put 600 GB/s in large type. Reviewers confirmed what that number is: the speed of the NVLink-C2C link between the CPU die and the GPU die inside the package.

The pool both dies actually draw from is slower. The shared LPDDR5X memory runs at up to 300 GB/s over a 256-bit interface, and that figure is the one that matters for workloads that repeatedly move large amounts of data between shared memory and the GPU. LLM decoding can be strongly memory-bandwidth-bound, particularly at small batch sizes, because token generation streams model weights from memory continuously — but throughput also depends on quantization, caching, kernels and model architecture. Graphics performance likewise depends on shader throughput, ray-tracing hardware, software optimization and power limits, not on memory bandwidth alone. In neither case can the 600 GB/s interconnect figure be substituted for the memory-bandwidth specification.

Set it against the obvious rival. Apple's M4 Max publishes a memory-bandwidth specification of roughly 540 GB/s, per the roundup's comparison — substantially higher than RTX Spark's 300 GB/s, which gives it a theoretical advantage in workloads limited by memory throughput, including some LLM decoding. Bandwidth alone does not establish which machine generates more tokens per second: model architecture, quantization, inference software, memory allocation and power limits also matter, and comparable model-by-model tests are needed before declaring a winner. The M4 Max is also a bandwidth reference, not Apple's current flagship; October 2026 reviews compare RTX Spark against newer M5 and M6 systems as well. RTX Spark's answer is capacity, within limits: a 70B model at 4-bit precision needs roughly 35GB for the raw weights, and potentially around 40GB or more once quantization metadata and runtime requirements are counted. That exceeds the dedicated VRAM of most conventional laptop GPUs — but RTX Spark is not uniquely able to hold such models, since other unified-memory laptops, including high-memory Apple configurations, can also host large quantized models. NVIDIA claims up to 120B models with up to a million tokens of context on the 128GB configurations. That claim is still NVIDIA's; the reviews verify the smaller models, not the top of that range. One more footnote: the chip's 55 TOPS neural processing unit is reserved for Microsoft Copilot+ features for now, not open to third-party software, per NVIDIA via ComputerBase — and it is separate from the CUDA GPU that produced the inference results in these reviews.

The CPU is competitive, and it is also from 2024

The CPU results refuse to line up into one verdict, which is itself useful information. Trusted Reviews put the ProArt P16 ahead of Intel's and AMD's comparators on Geekbench 6 multi-core, with only Qualcomm's newer Snapdragon X2 Elite Extreme faster. WccfTech found the same chip lagging x86 parts in its 3DMark CPU Profile test, then leading its Cinebench multi-core chart.

Single-thread work is where the results thin out. Cinebench 2026 single-core scores in the mid-500s place the chip roughly between Apple's M3 and M4. The cores are Arm's Cortex-X925 and A725, designs from Arm's 2024 generation, and ComputerBase noted that MediaTek, NVIDIA's partner on the CPU die, has already moved its own phone chips to a newer Arm generation. That age is relevant to a long-term purchase, but the benchmarks are the stronger evidence, not the introduction date alone: the N1X remains capable in multi-core work while trailing newer leaders in some single-thread tests.

Local AI: fits more, writes at a measured pace

The token rates reviewers measured are honest rather than spectacular, and they come from different models and different test setups, so read them as proof that local inference works rather than as a controlled comparison. Trusted Reviews measured Llama 3.1 8B at about 54 tokens per second, on both Windows ML and llama.cpp. WccfTech, running larger models in LM Studio on its tested ProArt P16 configuration, measured 70.27 tokens per second for GLM 4.7 Flash, 37.67 for the 74B Nemotron 3.5 Bolt, and 26.32 for Qwen 3.8 27B. The Bolt figure needs its own footnote: Nemotron 3.5 Bolt is heavily compressed and quantized, so its 74B parameter count is not equivalent to a conventional 74B memory footprint.

Read those two sets together and the product becomes clear. This is a comfortable machine for the 8B class and a usable one for models far larger than a conventional laptop GPU can load, at speeds set in large part by that 300 GB/s pool. CUDA working on production units, after prototype driver problems earlier in the year, matters as much as any single score. Battery needs the same care as the bandwidth figures: the 13 hours 55 minutes in Trusted Reviews' testing was a 1080p video loop on the ProArt P16, and HotHardware's 14 hours 35 minutes was video playback on the Surface Laptop Ultra. Those are media-playback figures, not measurements of endurance during sustained local AI inference or mixed work, and real-world endurance under load will be substantially shorter.

The practical angle

At $2,599 for the 24GB entry configuration and up from there, with the tested 64GB ProArt configuration at $4,399.99, the reviews support one specific purchase: a portable Windows machine for local AI work and serious integrated graphics, bought for memory capacity and CUDA. They do not support buying it as the fastest laptop at any one thing. Apple's machines publish higher memory-bandwidth figures, Qualcomm's newest chip is faster on CPU in several of the reviews, and a discrete RTX 5070 laptop still exists for buyers who only care about games. A narrower pitch than the launch marketing made, but a real one — and October 16 is the first day it rested on measurements instead of a keynote.

Sources

  1. [1] TechTimes, "RTX Spark Reviews Are In: Class-Leading GPU Meets Aging CPU Cores at Launch", October 16, 2026 — review roundup of ComputerBase, HotHardware, Tom's Guide, Trusted Reviews, Windows Central and KitGuru; source for the Time Spy Extreme comparison, the 300 GB/s vs 600 GB/s distinction, the M4 Max bandwidth comparison, the NPU restriction and the $2,599 starting priceRead source
  2. [2] Trusted Reviews, "Asus ProArt P16 (2026, RTX Spark) Review", first reviewed October 16, 2026 — primary review; source for Geekbench 6, Cinebench, the standard Time Spy result, Cyberpunk 2077 frame rates, the Llama 3.1 8B token rate, the video-loop battery result and the $4,399.99 tested priceRead source
  3. [3] WccfTech, "ASUS ProArt P16 'RTX Spark' Laptop Review", October 2026 — primary review; source for the N1X specification table, 3DMark CPU Profile and Cinebench results, and the LM Studio token rates for GLM 4.7 Flash, Nemotron 3.5 Bolt 74B and Qwen 3.8 27BRead source