News:

Willkommen im Notebookcheck.com Forum! Hier können Sie über alle unsere Artikel und allgemein über notebookrelevante Dinge diskutieren. Viel Spass!

Main Menu

Post reply

Other options
Verification:
Please leave this box empty:
Shortcuts: ALT+S post or ALT+P preview

Topic summary

Posted by AI ->normal gaming laptop
 - Today at 21:28:33
Correction 1: Qwen3.8-Flash-Next suggests this bandwidth formula (capacity-weighted harmonic mean):
Bandwidth = (RAM_size + VRAM_size)/(RAM_size/RAM_bandwidth + VRAM_size/VRAM_bandwidth).

The bandwidth of a 32 GB iGPU-only device at 153.6 GB/s remains unchanged.
The bandwidth of a 32 GB at 89.6 GB/s + 8 GB VRAM at 256 GB/s device, according to the new formula, is: 103 GB/s = (32GB+8GB)/(32GB/89.6GB/s+8GB/256GB/s).

Quote from: AI ->normal gaming laptop on September 21, 2026, 13:02:064915 (same as iGPU-only)
Correction 2: Clearly, having the additional 8 GB from the GPU, that allow fitting and running SOTA Qwen 35B-A3B model, is, if you want to run this model, so much better, even if the average memory speed is lower vs the iGPU-only device, but at least you can run it and at decent speeds as well. So let's try a formula that prefers memory size over memory speed:

New AI score = (Memory size)^a * Bandwidth / 10^3 (10^3: makes the numbers smaller).

How high you set the exponent depends on how important you consider the total memory vs memory speed to be (a=1: not important, a=2,3, or even higher: very/extremely important), but being able to fit and therefore run a model, vs not, can make all the difference.

Let's recalc AI scores of the same 5 devices from first post (first 2 rows are the 2 devices that previously scored equally):
Code Select
a=1.75:  |a=2:     |a=2.5:     |a=2.75:    |a=3:
 66      |157      |  890      | 2116      | 5,033
 65   -2%|165   +5%|1,041  +17%| 2621  +24%| 6,592  +31%
 67   +2%|169   +8%|1,073  +21%| 2693  +27%| 6,784  +35%
 85  +29%|219  +39%|1,451  +63%| 3747  +77%| 9,626  +91%
293 +344%|802 +411%|6,008 +575%|16434 +677%|44,958 +793%[1](e.g.:
219 = (32+12)^2*( (32+12)/(32/89.6+12/384) )/10^3
802 = (64-8)^2*( (64-8)/((64-8)/256) )/10^3)

Now, when using a=2, the device with the additional 8 GB (VRAM) scores +5% higher and when a=2.75, then 24% higher than the iGPU-only device.

[1] VRAM allocation of 64 GB Strix Halo is up to 56 GB, on both Windows and Linux: knowledgebase.frame.work/changing-memory-allocation-amd-ryzen-ai-max-300-series-By1LG5Yrll
Posted by Vlado
 - September 21, 2026, 14:23:07
I saw: Editors' Choice Award summer 2026
And I immediately knew it will be any Lenovo laptop.
Posted by AI ->normal gaming laptop
 - September 21, 2026, 13:02:06
QuoteWith a future-proof base configuration of 32 GB of RAM
Not if it comes to AI: It's already not future-proof when it comes to running SoTA Qwen 35B MoE AI LLM model:

Running AI locally (and privately) on a 32 GB RAM/unified memory iGPU-only laptop/device vs a normal gaming laptop:

On a 32 GB RAM / unified memory, iGPU-only, laptop / device, you get about this many context tokens:
Quote from: reddit.com/r/LocalLLaMA/comments/1sq94qx/is_anyone_getting_real_coding_work_done_with.. I've come to the conclusion that (1) 32768 is the biggest context I can get away with in an adequately smart model, and (2) it just ain't enough.

The additional 8 to 12 GB (VRAM) memory of gaming laptops make all the difference in being able to fit a SoTA AI model[1] huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF 4-bit UD-Q4_K_XL quant at decent context:
Calculating available context: 16,384 [KV cache context tokens per GB]*(32 [GB total memory] - 6 to 8 [GB for the OS] - [quant filesize])[2]:
Qwen3.6-35B-A3B-MTP-GGUF -> UD-Q4_K_XL.gguf (22.9 GB):
  • On a 32 GB RAM / unified memory, iGPU-only, laptop/device:
    Approx. 34,400 context tokens = 16,384*(32-7-22.9) (confirms the quote).
  • On a 32 GB RAM + 8 GB VRAM GPU laptop/PC:
    Approx. 165,500 context tokens = 16,384*(32+8-7-22.9).
  • On a 32 GB RAM + 12 GB VRAM GPU laptop/PC:
    Approx. 231,000 context tokens = 16,384*(32+12-7-22.9).
(1 token = 0.75 words)

(The biggest "innovation" I see, when it comes to AI, is the switch from 2 GB to 3 GB GDDR7 memory density: Compact and light 8 GB VRAM gaming laptops become 12 GB VRAM ones. (And 12 GB VRAM become 18 GB VRAM ones (no mention of them so far).)

Let's compare the AI scores of the fastest 32 GB RAM at 9600 MT/s iGPU-only laptops vs normal gaming laptops (32 GB at 5600 MT/s + a 8 and 12 GB VRAM, RTX 40 and 50, GPUs; and these normal gaming laptops also have upgradable RAM:
  • AI score of 32 GB RAM iGPU-only devices: 4915 = 32 GB RAM  * 153.6 GB/s (=128-bit * 9600MT/s / 1000 / 8).
  • AI score of a RTX 4060 Laptop, 8 GB VRAM, GPU + 32 GB RAM at 5600 MT/s: 4915 (same as iGPU-only) = 32 GB RAM * 89.6 GB/s + 8 GB VRAM * 256 GB/s.
  • AI score of a RTX 5070 Laptop, 8 GB VRAM, GPU + 32 GB RAM at 5600 MT/s: 5939 = 32 GB RAM * 89.6 GB/s + 8 GB VRAM * 384 GB/s.
  • AI score of a RTX 5070 Laptop, 12 GB VRAM, GPU + 32 GB RAM at 5600 MT/s: 7475 = 32 GB RAM * 89.6 GB/s + 12 GB VRAM * 384 GB/s.
  • AI score of a 64 GB Strix Halo (not in the article) (4060 Laptop type of performance, see 3dmark.com/search (Radeon 8060S)): 16384 = 64 GB RAM/unified memory * 256 GB/s. But mind that only on Linux you may be able to allocate most of the memory (search for 'strix halo linux gtt') (this is why you maybe should not get a Strix Halo-based laptop/device, unless you are fine with using Linux).

This is to sd  on a normal/typical/classical gaming laptop, than on the iGPU-only laptops.

So, if you care about running AI locally (and privately(, because some companies may demand that (for certain tasks))), simply choose a normal gaming laptop, as the prompt processing (input) and token generation (output) speeds will be faster than on mentioned iGPU-only laptops.
(And generally, a desktop PC build is going to offer 2-3 more AI score / performance per buck than a laptop.)

More about running AI models locally: notebookchat.com/index.php?topic=315954.0 ("Your own ChatGPT, offline: AI without the cloud on your laptop") and comments.

[1] artificialanalysis.ai/?models=muse-glimmer%2Cqwen3-6-35b-a3b%2Cqwen3-8-27b%2Cg9v3-39a5b%2Cgemma-4-31b%2Cgemma-4-26b-a4b%2Cqwen3-8-27b-non-reasoning&intelligence=artificial-analysis-intelligence-index#artificial-analysis-intelligence-index
(Qwen3.8-27B scores much higher, but due to being a dense model, will run many times slower)
[2] reddit.com/r/Qwen_AI/comments/1vo8pjz/qwen3827b_kv_cache_works_out_to_64_kibtoken_so/
Posted by Redaktion
 - September 20, 2026, 21:29:01
What laptop tested in summer 2026 performed the best with its swiftness, features, and endurance? We thoroughly tested 49 newly-released notebooks and can now present the winners, that is, the most exciting models, in this article.

https://www.notebookcheck.net/Best-laptops-for-fall-2026-49-notebooks-tried-and-tested.1398522.0.html