Comment by throwa356262
16 hours ago
The sad part of this is that all those powerful NPUs are basically paperweight.
35-50 low power TOPS taking 1/5 of the die area and not being used by anyone...
16 hours ago
The sad part of this is that all those powerful NPUs are basically paperweight.
35-50 low power TOPS taking 1/5 of the die area and not being used by anyone...
Local speech recognition is something useful that runs well on a (power efficient) NPU.
Dragon Naturally Speaking used to cost real money and wasn't as accurate as an open model like Whisper.
These things are a waste of sand.
Could’ve added more gpu or cpu or cache and people would’ve been happier.
When these came out, I thought, what is the average PC user going to get out of this? Faster image touchups, and...? You can't run an LLM chatbot of any comparison to free cloud models.
Oh yeah, and Recall. That worked out well.
Yep… massive waste. A dsp coprocessor hobbled by npu workflows.
[dead]
> Could’ve added more gpu or cpu or cache and people would’ve been happier.
I actually wonder about this - my regular dev workstation is a Qualcomm ARM laptop - by-and-large my bottleneck for CPU-bound tasks is thermal. Would adding more CPU or cache instead of the NPU help with performance/watt?
We cantnrun local models on them? They must have some sort of api exposed right? Qualcomm has the Snapdragon Neural Processing Engine iirc
You can do cool experiments with them. But nothing a typical consumer is likely to care about.
Yes, you can. Of course relatively limited, but you absolutely can run local models.
Based on my own testing, the Windows drivers are a bit restricted to do some more advanced stuff, and it may be due to some security issues in the past? at least that is what I gathered from exploring this issue with GLM 5.3