xda-developers Task Manager to add NPU support

[UltraMega]

Quote

As reported by Neowin, AMD NPUs will finally show up in Windows' Task Manager. If you're not sure what an NPU is, it's a new component to computers to aid with processing. Much like how the CPU handles regular processes and GPUs are dedicated graphics processors, NPUs are specially used for AI-based processes.

With this agreement, AMD's NPUs will now use the Microsoft Compute Driver Model (MCDM). Part of this implementation means you can keep tabs on how the NPU is doing and how much processing power it's using via Task Manager, which will be particularly important when people begin using their NPUs for strenuous AI-based tasks. As AMD puts it:
 

Being able to track how resources are allocated in real-time and which system components are under load is useful for monitoring application behavior. This kind of tracking is particularly important in notebooks where end users may choose to maximize battery life by controlling where certain workloads run or adjusting global power settings either via the Windows Settings menu or in an OEM-provided application.
 

NPU implementation is going to be vital in the coming months. With AI PCs entering the consumer market, Microsoft will ideally support as many different NPU brands as possible so people aren't left out because of their choice of hardware.

AMD's NPUs will finally get support from Windows' Task Manager (msn.com)

 

This does not apply to Ryzen 7000 CPUs or 7000 series GPUs because neither have NPUs. I think this just serves to help understand what an NPU is and what the roadmap for their implementation will look like. 

 

What is an NPU? | Windows Central

 

Intel unveils Core Ultra, its first chips with NPUs for AI work (engadget.com)

 

Edited February 22 by UltraMega

[pio]

I mean technically speaking, per the definition in the article, "Neural Processing Units are dedicated AI chips that remove some of the work from a computer's CPU or GPU so the device can work better.", that would mean any RX 6000 or 7000 series GPU can function as a NPU (same with any of Nvidia's RTX series cards).

Interesting, I guess we'll see what the future of NPU's will end up looking like soon enough.  Maybe one day, we'll be popping NPU's into our systems to assist with our GPU's.  Like I said, we kind of can NOW, its just people aren't well enough aware that you can run Nvidia and AMD together (or Intel too now).

Edited February 21 by pioneerisloud

[UltraMega]

13 hours ago, pioneerisloud said:

I mean technically speaking, per the definition in the article, "Neural Processing Units are dedicated AI chips that remove some of the work from a computer's CPU or GPU so the device can work better.", that would mean any RX 6000 or 7000 series GPU can function as a NPU (same with any of Nvidia's RTX series cards).

Interesting, I guess we'll see what the future of NPU's will end up looking like soon enough.  Maybe one day, we'll be popping NPU's into our systems to assist with our GPU's.  Like I said, we kind of can NOW, its just people aren't well enough aware that you can run Nvidia and AMD together (or Intel too now).

Definitely a bit of overlap between a GPU and an NPU, but not as much as you might think. An NPU can be up to like 10k times better at certain tasks vs a GPU. 

 

Just like how a CPU technically can do most of what a GPU can do, just not very well. 

 

If AMD GPUs had NPUs or functioned well as NPUs, FSR and DLSS would be of similar quality. It's the NPU that makes the difference. 

Edited February 21 by UltraMega

[Slaughtahouse]

28 minutes ago, UltraMega said:

Definitely a bit of overlap between a GPU and an NPU, but not as much as you might think. An NPU can be up to like 10k times better at certain tasks vs a GPU. 

 

Just like how a CPU technically can do most of what a GPU can do, just not very well. 

 

If AMD GPUs had NPUs or functioned well as NPUs, FSR and DLSS would be of similar quality. It's the NPU that makes the difference. 

 

I want to believe that but my instincts tell me that the NPU will be more like a PhsyX type card. Great for upgrading an old rig but not ideal if you're considering getting faster video card that can simply do this work. Especially if future graphics cards have an NPU. 

 

Thanks for sharing. Curious how "NPUs" will evolve and from a practical standpoint, if they move the needle forward for enthusiasts. 

[UltraMega]

57 minutes ago, Slaughtahouse said:

 

I want to believe that but my instincts tell me that the NPU will be more like a PhsyX type card. Great for upgrading an old rig but not ideal if you're considering getting faster video card that can simply do this work. Especially if future graphics cards have an NPU. 

 

Thanks for sharing. Curious how "NPUs" will evolve and from a practical standpoint, if they move the needle forward for enthusiasts. 

Going forward, pretty much all GPUs and CPUs will have NPUs embedded. 

[UltraMega]

Just wanted to add, I've been trying to find some good info on the technical differences between a GPU and an NPU but it's kind of hard to find a good article or video that explains it in detail. Most of the info I can find about what makes NPUs different from GPUs basically just boils down to "NPUs are much better at AI/Deep learning/Machine Learning tasks".

 

I think a good example would be this: I know from personal experience, you need a GPU about equal or better than a 3080Ti to do Stable Diffusion image generation locally with decent results. That means a big GPU that takes a lot of power and has a huge heatsink. But now there are NPUs that can do the same task even better without the need for a ton of power and heat. There are NPUs that will be able to do things like image generation on a standard smart phone. 

 

Another point that comes up a lot is that things that are typically processed in the cloud will be able to be done locally much more often, which will reduce power consumption for data processing in a big way. 

 

GPUs and CPUs are good at tasks with clear defined answers. NPUs are good at tasks with no defined answer, where the answer need to be created or generated based on existing data. It's kind of ironic that AI isn't very good at math in the same sense that humans are not very good at math, but it makes sense when you think about it. We are very good at coming up with undefined/creative answers to things but not good at very specific detailed answers. Same is true for NPUs to a certain degree. 

 

 

There was a time when simple 3D graphics were rendered on the CPU, and then GPUs came along and did that much faster. CPUs could still do it, but GPUs can do it better. NPUs are to GPUs as GPUs are to CPUs when it comes to AI/DL/ML tasks. The average user isn't doing anything like that today, but presumably once NPUs are ubiquitous we are all going to be using them for a lot more. 

Edited February 21 by UltraMega