Mesure NPU Enhance/impact

[Igor_fagundes]

Can anyone confirm that Adobe applications that feature I.A. Do require local processing (CPU/NPU)? (and not only GPU). I'm looking forward to understanding what is the impact of using NPU based Processor.

Is there any benchmark database or even a benchmark that I could run in both platforms to help me to understand these possible gains?

Besides Generative AI, are there other use cases where workloads have shifted from execution on the CPU to execution on the NPU (even partially)? Any examples of application optimization independent of Generative AI?

[D Fosse]

Generative AI is processed on Adobe's servers and returned to the user.

Other non-generative AI-based algorithms use various processing. Some use the GPU, some use the CPU, and some a mix.

Nvidia RTX-series GPUs have specialized Tensor cores for AI tasks.

[Conrad_C · Accepted solution]

I am not aware of many Adobe applications using the NPU yet. I can only think of two examples so far:

Adobe Lightroom Classic/Lightroom and I think Adobe Camera Raw began to use the Apple Neural Engine NPU to help accelerate AI Denoise further. Denoise performance is usually completely dependent on the power of the GPU (adding CPU cores, RAM, and storage doesn’t change the results much). After Adobe enabled the Apple Neural Engine for AI Denoise, processing times dropped by a significant percentage depending on the Mac model. However, unwanted visual artifacts were found that caused Adobe to suspend use of the Apple Neural Engine for Denoise, and it is not clear to us users when that will be re-enabled.

At last fall’s Adobe MAX conference, it was announced that Premiere Pro would begin using the NPU on Copilot+ PCs for the Audio Category Tagger feature. The linked press release calls Premiere Pro “the first Adobe app to leverage NPU on Copilot+ PCs.” That feature is now available in the public beta of Premere Pro, named “Intel NPU-supported Audio Category Tagger”.

There may be more examples but those are the only two I know about. It would be interesting if anyone else can add to that list of Adobe app features that actually take advantage of the increasingly common NPU. If no one else can add more examples, then it’s possible that an NPU might be largely unused by Creative Cloud apps.

Regarding benchmarks, as a Mac user I don’t know as much about Windows tools. On the Mac side, it’s disappointing that macOS Activity Monitor (similar to Windows Task Manager) doesn’t provide a way to monitor NPU activity in the same way that it will show you an activity history for the CPU and GPU cores. The only Mac tool I know about for monitoring NPU use is the free utility macmon, which does have an Apple Neural Engine history display, although it looks like it’s more about power consumption than activity. It’s the graph marked “ANE” at the bottom right corner of the picture below, which is from the macmon page on Github.