Seems the Youtube reviews are out.
2 Likes
Show me the links! 
I was looking for those for the past couple of days
Found 1!
1 Like
None of the reviews test the NPU at all.
I mean, is AI still not really a thing?
I can understand nobody tested the new maybe improved wifi antennas, that is niche as hell but how is there still no mention of the pixel response times of the new display?
Idk about the intel world, but I know in the AMD world NPUs are largely just a waste of silicon still.
The intersection of:
- Needs matrix multiplication
- Needs to be more power power efficient than iGPU
- OK to be slower than iGPU
- Has software support
Is nonexistent in my experience.
I’m sorry, I have to disagree here. AMD XDNA2 is very real, works well, and I actually use it every day for a wide range of tasks.
Maybe you personally don’t need it, and that’s fair. But calling it a “waste of silicon” ignores the fact that, it’s significantly more power-efficient than GPU inference while also being faster.
Your new “existent” experience will be FLM. Try it for yourself: https://fastflowlm.com/
And xdna1 is pretty much entirely abandoned and can be rightfully called a waste of silicon.
That’s awesome, I’m happy to be wrong about this one, I’ll give fastflowlm a try!
1 Like
100% agree. XDNA1 is a vaporware full of waste. and mostly a lie 
1 Like
1 Like
In the Level1techs there’s already a thermal problem
It reached 100C with only 33.8W, on my AMD 7840U I got 85-95C under 35W. Also, uneven CPU thermals, a whole 25C difference. idk about the 400MHz but I guess at least one of the cores got throttled. I prefer range on
btop
The default setting (shown in the review) can be a bit random.
I guess that’s why users can never get to advertised 60W PL2, but it still better than the AMD models, performance wise
Also, I see this as a red flag
Please, do not bring connected standby to Linux
I tested GitHub - aweussom/NoLlama: NPU Ollama - An Ollama/OpenAI compatible API for Intel OpenVINO compatible computers · GitHub for intel NPU.
However the models you can run are limited to openvino AND npu-optimized ones which are all pretty old. I tried with non-npu-optimized openvino models and they can only be run on GPU.
There’s also little benefits of running models on NPU. The power draw is around 10W on NPU, so not that low either. On top of that GPU token rates are much higher.
Hitting 100C with 33W is about the same as I observed.
Linux uses s2idle which is not connected.
Kinda understandable since it is different kinds of cores, the temperateure difference within the p-cores and within the e-cores seems quite even.
Well shucks looks like the pixel response times are still pretty bad
2 Likes
Actually if you look closer C0-C3 are P cores and 3 of them at 80C, C3 at 100…
How did they manage to get 55W for 20 seconds while most users in this forum can barely reach 35W?
Good catch, though unless you got better info it may also be one of the low-power e-cores.
That C4-C11 are the e-cores is a relatively safe bet, if the p-cores are 0-3 or 12-15 is less sure (probably still 0-3).
Still there is one hotter core in a cluster where all cores are fully loaded is odd indeed.
The e-cores all being exactly the same temperature is also a bit odd, do they share a sensor or are they all individually fine controlled to stay at the red line?
Pretty sure the reviewers outnumber the “most users on this forum” sample size at this point by quite a bit. Different testing method + windows I suppose. If you look at the hwinfo screenshots the cyberpunk test never goes above 30 but the prime95 does (it is possible they cleared the min/max sometime during the cyberpunk test though).
This intrinsic uneven thermal is probably a limiting factor. It has better cooling (one big heat pipe plus factory PTM), but has worse temperature than 7840U. In fact, if you check Notebookcheck’s 7840U review, the X7 358H has better iGPU and RAM performance but the 7840U still has better CPU performance. I guess the AMD AI 300 series is underwhelming that standards have been lowered
I think it might be necessary to pull up the Intel equivalent of amdgpu_top to find out which core has which temperature.
What happens during thermal throttling? reducing overall power on redirecting power from hotter to cooler cores?