(Not claiming to be 100 % correct.)
Depending on one's definition, this laptop's configuration makes it not really deserve the "AI" in its product name.
My definition is that I want to run/fit into RAM+VRAM at least this very popular SOTA AI LLM model: huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF UD-Q4_K_XL quant (17.9 GB). Some may prefer huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF UD-Q4_K_XL quant (22.9 GB), as it runs multiple times faster, but it also requires a bit more memory.
In laptops/consumer desktop PCs, the GPU's VRAM is much faster than system RAM (12-channel RAM servers can be faster (614 GB/s)). As such, if an AI LLM model + its context fits into the VRAM completely, without offloading to RAM, it will run much faster.
AI requires these few things:
1. Memory size to fit a decently capable LLM model (the mentioned ones) + its context.16 GB RAM + 8 GB VRAM (remember, the OS requires 6-8 GB for itself) doesn't allow to fit the mentioned 27B quant.
Adding a second 16 GB RAM stick will give you 40 GB (32 GB RAM + 8 GB VRAM). This should also allow for large context (100,000+ tokens) that is often required in agentic workflows.
2. Memory speed, also known as memory bandwidth: Relevant for token generation (output) speed.RAM: The practical/measured memory bandwidth is only 43921 MB/s (single-channel speed). As mentioned, adding a 2nd 16 GB RAM stick will double the memory bandwidth by enabling dual-channel and give you:
89.6 GB/s = 128-bit * 5600 MT/s / 1000 / 8.
8 GB VRAM: 5070 Laptop GPU: 384 GB/s.
(The vast majority of all PCs/laptops are 128-bit (2 * 64-bit or 8 * 16-bit) systems.)
Ideally, sellers should warn/inform their customers if they buy a single-channel RAM configuration and how to fix it (populating both RAM slots). Or not sell such configurations in the first place (but then the 16 GB RAM configuration would be 2 * 8 GB sticks and that wouldn't be cool). Fortunately, the single-channel speed should be easily fixable in this laptop by adding a second RAM stick.
3. GPU 3D/FPS performance / compute: Relevant for prompt processing (input) speed. But the GPU performance is, again, determined by the memory bandwidth, so AI requires really only 2 things.Here, the 5070 dGPU scores 13753 Points in 2560x1440 Time Spy Graphics. For comparison:
3dmark.com/search - Time Spy:
- Strix Halo's Radeon 8060S iGPU (256-bit at 256 GB/s): "Average score: 10034"
- RTX 4070 desktop GPU (504 GB/s) (has 12 GB VRAM): "Average score: 16568"
- RTX 5070 Ti desktop GPU (896 GB/s) (has 16 GB VRAM): "Average score: 24455"
- RTX 4090 desktop GPU (1008 GB/s) (has 24 GB VRAM and will fit the mentioned Qwen 27B quant and run it much faster): "Average score: 30487"
13753 Points is pretty good. If you need faster prompt processing and want to get the most AI performance bang for the buck, you'd be better off building a desktop PC.
4. The number of CPU threadsThe number of CPU threads matters, but it's rarely a hardware limitation, as 3-4 threads top-out the inferencing AI performance in my 7800X3D, dual-channel, PC + GPU (and only 1 thread tops-out the performance on my Air 15 2026, this could be due to Arm architecture or I need to take a look at it again). The new MTP LLMs require more threads (approx. 2 times as many in my case) (see also reddit/"PSA: Test your "threads" argument in llama.cpp (+80% performance in my case)").
The GPU's GB/s numbers are taken from en.wikipedia.org/wiki/GeForce_RTX_50_series#Mobile and 40_series.
Deutsch
English
Español
Français
Italiano
Nederlands
Polski
Português
Русский
Türkçe
Svenska
Chinese
Magyar