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 What?
 - Today at 11:53:44
A low budget, and I think somewhat popular build, is 2x5060 Ti 16G for a total of 32 GB VRAM. And if using this quant: huggingface.co/unsloth/Qwen3.8-27B-NVFP4, there would be still enough VRAM left for about 128,000 of context tokens (see topic=315954) (8 GB VRAM for 128,000 context tokens): 32 GB - 22.6 GB = 9.4 GB.
But AI agents may go beyond 128k of context, so then you should try UD-Q4_K_XL from huggingface.co/unsloth/Qwen3.8-27B-GGUF, a quant that is smaller and leaves more VRAM for the context: 32 GB VRAM - 17.9 GB quant = 14.1 GB left for the context. 14.1 GB -> ~225,000 context tokens.
But a Q4_K_XL quant may not perform so well beyond 80k-130k of context, so your next try would be one of the 3 (as of this writing) 5-bit quants.
Posted by What?
 - Today at 10:51:40
2*24 GB VRAM GPUs because they have enough VRAM capacity to fit a 27B quant + have enough VRAM left for token context, so all is in the VRAM. This is very important to have actually fast input and output token speeds, which AI agents require. (For simple chat-style conversation where you ask the AI up to a dozen (small-ish) questions per day or something, one GPU (not even 24 GB VRAM required, or even none) + offloading to system RAM would be enough. But, again, AI agents have completely different speed requirements.)

You can also get a single, 48 GB VRAM (or more) workstation, and not consumer, GPU, but it will be much more expensive for what you get, unless you really want to deal with only 1 card.
Posted by What?
 - Today at 10:37:48
Quoteweak GPU, no dGPU option
Indeed, how is it "perfect for AI agents" when AI agents have to handle a lot of context input and output tokens, which requires appropriate compute and memory speed (aka memory bandwidth)?
This laptop has a 128-bit (aka dual-channel (2*64-bit)) RAM memory interface and no dedicated GPU to further increase input (aka prompt processing) and output (aka token generation) token speeds.

Theoretical RAM speed: 89.6 GB/s = 128-bit * 5600 MT/s / 1000 / 8.
Practical RAM speed (is always slower): 80 GB/s ("80270 MB/s").

In this AI speed sense, any gaming laptop upgraded to the same 2*32 GB, 5600 MT/s (or even 4800 MT/s) RAM, which also has a GPU is superior to his.

For (affordable) AI agents it's best to get a desktop PC with GPU(s) that have the densest VRAM per GPU and the highest memory bandwidth you can and want to afford (usually 2 (used) RTX 3090, 4090 and a quant of Qwen3.8-27B, which is currently the best AI model in its size class[1]).

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/open-source?models=qwen3-8-27b%2Cmuse-glimmer%2Cqwen3-6-35b-a3b%2Cgemma-4-31b%2Cgemma-4-26b-a4b
Posted by Redaktion
 - Yesterday at 21:24:59
Mobile workstations like the HP ZBook series are typically meant for GPU-intensive applications. Not so much the HP ZBook 8 G2a. This 14-inch workstation with an AMD CPU forgoes a strong graphics chip, instead offering plenty of RAM and a fast processor - a perfect fit for agentic AI.

https://www.notebookcheck.net/This-laptop-with-64-GB-RAM-is-perfect-for-AI-agents-HP-ZBook-8-G2a-14-review.1350756.0.html