Post reply

Name
Email
Subject
Message icon

Other options

Return to this topic
Don't use smileys

Verification:

Please leave this box empty:

Shortcuts: ALT+S post or ALT+P preview

Topic summary

Posted by Alice84

- Today at 12:34:33

Can someone explain to me why almost top of the line Intel Core Ultra 9 with 16 cores and 1.8nm architecture cant keep up in terms of performance with 3 year old R7-7840U with only 8 cores and build on 4nm? Also iGPU performance is almost double on Ryzen despite having slower memory.

What the hell is going on?

www.notebookcheck.net/index.php?id=127065&specs[]=341381&specs[]=360204

Posted by Worgarthe

- Today at 11:35:41

Quote from: Not special for AI on Today at 10:40:54This doesn't really deserve the "AI" / "AI+" in its product name:
AI requires these 2-3 things:
...

40+ TOPS NPU is necessary to be branded as "Copilot+" and/or "AI+." It is comparable to "Ultrabook" back in the day.

QuoteWhat is a Copilot+ PC and what sets them apart?

Copilot+ PCs represent a cutting-edge category of AI-enhanced Windows 11 devices. In addition to the CPU and GPU available on conventional computers, Copilot+ PCs feature high performance neural processing units (NPUs). These NPUs—capable of executing over 40 trillion operations per second (TOPS)—are designed to handle demanding AI workloads efficiently, enabling high-performance features such as real-time language translations. By handling certain AI tasks, the NPU frees up the CPU and GPU to focus on other operations, resulting in improved overall system performance and can deliver longer battery life 2 due to optimized power distribution.

40+ TOPS is here in this MSI. It's a dumb marketing name, but 40+ TOPS is certainly here. This is a thin and light laptop, not a dedicated local LLM machine (which you, for some reason, expect every existing laptop to be). That's all.

Posted by RobertJasiek

- Today at 11:27:57

"GPU performance is determined by the memory bandwidth"

Only if memory bandwidth is a bottleneck for the used AI software.

"The number of CPU threads doesn't matter for running endconsumer AI (aka inferencing)"

Only if the used AI does not profit signficantly from more CPU threads. And think again: would you run on just one thread?

As always, your statements about AI are heavily prejudiced because you only consider LLMs but pretend all AIs would behave like LLMs. They do not!

Posted by Not special for AI

- Today at 10:40:54

This doesn't really deserve the "AI" / "AI+" in its product name:
AI requires these 2-3 things:
1. Memory size to fit a decently capable LLM model + its context.
32 GB RAM gives you a limited context of ~32000 tokens (= ~24000 words) for a quant of the new 35B MoE SOTA LLM Qwen3.6-35B-A3B-UD-Q4_K_M. And agentic workflows often require more than 100000 tokens, so it will not work.

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 only 32 GB RAM/unified memory is the biggest issue with this "AI+" laptop, had it at least 48 GB, then it would deserve the "AI+" in its product name.
Since there's no dedicated GPU, there's also no fast memory/VRAM, to partially or fully offload an AI LLM model to.
2. Memory speed, also known as memory bandwidth: Relevant for token generation (output) speed.
The practical/measured memory bandwidth is 97068 MB/s. Therefore, the memory is running at 7500 to 8000 MT/s:
96 GB/s = 128-bit * 8000 MT/s / 1000 / 8 * (0.7 to 0.8) (theoretical -> practical).
This is ok/expected for a (non-top of the line) 2026 128-bit(/dual-channel PC) system (the vast majority of all PCs/laptops are 128-bit systems). Some 2026, 128-bit, systems run at 9600 MT/s.
3. GPU 3D/FPS performance / compute: Relevant for prompt processing (input) speed. But GPU performance is determined by the memory bandwidth.
Here, the iGPU scores only 2391 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"
this/Intel Graphics (Panther Lake) (4-core) (~128 GB/s = 128-bit at 7500-8000 MT/s): "Average score: 3226"
Intel Arc B390 (12-core) (153.6 GB/s = 128-bit at 9600 MT/s): "Average score: 7353"
RTX 4070 desktop GPU (504 GB/s): "Average score: 16568" (correct, you'd me much better off building a desktop PC for AI and get a much better AI performance bang for the back)

(The number of CPU threads doesn't matter for running endconsumer AI (aka inferencing) and 1-4 threads top out the inferencing performance.)

Posted by Redaktion

- Today at 08:33:37

With the Prestige 13 AI+, MSI is targeting mobile users who are often on the move and value a compact, lightweight laptop. Equipped with a brand-new Panther Lake processor, the 13-inch device is designed to win over customers looking for maximum portability.

https://www.notebookcheck.net/MSI-Prestige-13-AI-laptop-review-The-perfect-companion-for-on-the-go.1318523.0.html