Microsoft Surface Laptop Ultra RTX Spark Review - The 2000-nits Mini-LED panel is a problem

[Redaktion]

The Surface Laptop Ultra with RTX Spark is supposed to be Microsoft's best laptop. And while it is a good product, there are some disappointing aspects, especially the Mini-LED panel. It is super bright, but it suffers from ghosting and there is no local dimming.

https://www.notebookcheck.net/Microsoft-Surface-Laptop-Ultra-RTX-Spark-Review-The-2000-nits-Mini-LED-panel-is-a-problem.1426103.0.html

[Comparison to alternative]

5094

bucks
Let's compare that to an alternative that exists for over 1 year and 8 months: For AI and gaming:
Let's say you already have a 32 GB RAM desktop PC, adding another 32 GB is a few hundred bucks, and selling your, say, 5070 and getting a 5070 Ti won't also be the full price of the Ti.

Now let's compare an AI score (memory size times memory speed):
This 64 GB N1X Spark laptop:

• 19264 = 64 GB unified memory * 301 GB/s (= 256-bit * 9400 MT/s / 1000 / 8).

Desktop PC with 64 GB DDR5-6000 RAM + 5070 Ti:

• 20480 = 64 GB RAM * 96 GB/s (= 128-bit * 6000 MT/s / 1000 / 8) + 16 GB VRAM * 896 GB/s.

The score is similar, but it's 64 GB vs 80 GB. This means you can fit bigger (quants of) AI models or allow for a larger context.

Now let's compare GPU performances:
3dmark.com/search -> Steel Nomad:
Spark N1X 6144-core: Average score: 2941 or "3134 Points" per this review.
5070 Ti: Average score: 6931.

Agreed, a desktop PC isn't as compact and you might even say that it misses the point, but it's ~1000 vs ~5000 bucks, you get additional 16 GB of the GPU and you'd have roughly over 2 times the FPS in games (2.35 = 6931/2941), and a desktop PC is upgradable and repairable, too.

Communities and more about running AI models locally and privately:
notebookchat.com/index.php?topic=315954.0 ("Your own ChatGPT, offline: AI without the cloud on your laptop")
reddit.com/r/LocalLlama/top
reddit.com/r/LocalLLM/top
artificialanalysis.ai/models/open-source

[Comparison to alternative]

PS: The frontier model, as of this writing, for Spark or any ~300 GB/s unified memory system, is Qwen3.8-Flash-Next (you may go to artificialanalysis.ai and compare it there to other models, it scores the highest per parameter and even NVIDIA has a version of it: huggingface.co/nvidia/Qwen3.8-Flash-Next-FP4-Local (only RTX 50 series, like Spark support NVFP4), but I'm not sure if it will fit in a 64 GB Spark system, even if the 30% filesize n-gram of it can be streamed from a SSD with almost no performance loss). You can also try to run a quant from e.g. huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF on llama.cpp (it has a nice webUI) or type "Qwen3.8-Flash-Next inference engine") in a search engine of your choice.

[Comparison to alternative]

Maybe off-topic or not if you choose a 128 GB Spark + attach a dedicated GPU, or still cheaper than this reviewed laptop: On a AM5 B850 mid-range desktop motherboard you can upgrade to 256 GB RAM (4*64 GB) + GPU and run a 3-bit quant of huggingface.co/unsloth/DeepSeek-V4.1-Flash-1017-GGUF:

Hey guys, made 1-bit Unsloth GGUFs retain ~88.2% accuracy, shrinking size from 508GB to 250GB (−52%). Run on a 128GB+ RAM device for the smallest 1-bit, 3-bit on 256GB or full precision on 320GB. The ngram really came in handy this time - just ensure you have enough SSD. V4.1 Flash 1017 is the strongest open model to date.

Guide: [..]

!