Monday, November 24th 2025
Intel "Nova Lake" NPU 6 Delivers 74 TOPS for Desktop AI PCs
Intel's next-generation "Nova Lake" desktop processor generation will arrive with a 6th-generation Neural Processing Unit (NPU) design. However, we now have preliminary performance details about it thanks to a well-known hardware leaker @jaykihn0 on X. The latest leak indicates that the "Nova Lake-S" desktop configuration will carry an NPU on one of its tiles that deliver 74 TOPS of INT8 performance, which is significantly higher than anything Intel previously shipped in its desktop segment. As of now, Intel ships NPU 3 with "Arrow Lake-S," which came out in 2023 with "Meteor Lake" and offers 11.5 INT8 TOPS. Intel's current highest-performing NPU is the NPU 5 scheduled for "Panther Lake" in early 2026, which will carry up to 50 TOPS for Copilot+ AI PCs.
Intel's decision to upgrade the NPU IP with "Nova Lake" has been known since late last month, when compiler patches confirmed the inclusion of a new NPU 6 in the next-generation desktop chips. However, the exact performance target was unclear until now. Since Intel is enhancing its NPU IP on the same microarchitecture, we can expect the NPU in Nova Lake to have a larger TOPS budget while maintaining software compatibility. This is excellent news for software that incorporated NPU acceleration in previous generations, as the performance improvements will be seamlessly integrated.
Source:
@jaykihn0 on X
Intel's decision to upgrade the NPU IP with "Nova Lake" has been known since late last month, when compiler patches confirmed the inclusion of a new NPU 6 in the next-generation desktop chips. However, the exact performance target was unclear until now. Since Intel is enhancing its NPU IP on the same microarchitecture, we can expect the NPU in Nova Lake to have a larger TOPS budget while maintaining software compatibility. This is excellent news for software that incorporated NPU acceleration in previous generations, as the performance improvements will be seamlessly integrated.
24 Comments on Intel "Nova Lake" NPU 6 Delivers 74 TOPS for Desktop AI PCs
:toast:
People who want AI processing use dedicated GPUs for that, which are hundreds or thousands of times faster for AI calculations than NPUs.
You need like 4 5090s with a $15K build, or you could just pay $1500-$2000 for this.
We know it's a bubble, but pretty much everyone saying they don't care about AI is using ChatGPT or Gemini online anyways.It's an unexpected plus of unified memory configurations for models which require more memory capacity than computing speed to iterate, inference and/or train. dGPUs can't really do that without extremely fast, specialized interconnects, since they're a complete system. If we're to give due credit for this type of architecture on a computer though, Apple's 5 years ahead of everyone else.
i'm aware AI slop is the future regardless of company
I mean after all it's a tool, it could be put to some use if it can actually do something, and based on various apps/benchmarks it seems it can.
So it's up to them (game devs and others) if they can harness its capabilities to enable something superior than before.
But anyway, I think enough/many people expect Nova Lake to be great as a "home user workstation CPU" so this NPU performance news is more in that direction, what people are worried about/unsure of is gaming performance.
Given there will be two CPU breeds apparently, one plain and one with bLLC, we can only guess how they will perform. I would say plain Nova Lake will beat Raptor Lake, I can't see it being worse. There is a chance that they're only equal, although after two gens that would be a disgrace. So modest uplift over Raptor Lake then? And if the bLLC version has strong uplift over plain then it's on! Fist fight with brass knuckles ftw! :nutkick::rockout:
On one model I had, a 20B parameter one, I asked it how to make stuffed peppers, Thai curry, described a common engine problem on a 2009 F-150 and asked what it likely was, and write a C# .Net program to use the Dept of Treasury API to download treasury data for the past 10 years into a .csv file.
It didn't miss a beat. All run locally. I have no idea how they crammed such diverse knowledge into something that fits in an 16GB file.
A lot of people are going to be interested in that I think, especially when you consider what AI can do with information about what you look at and search for. The next weave of AIs are going to be able to control your PC too, use the browser and apps on your computer.
Lot more than just LLMs use NPUs.
There's a lot of interest around these in developer communities, because you can build your own LLM and integrate it into your development environment, instead of paying for one.
lmstudio.ai/
Anyway, they make Xeon-6 chips with no NPU, so you have a choice.