Intel Arrow Lake will debut later this year, but only some chips will feature neural processing units (NPUs). Currently, only Arrow Lake-H/U mobile processors will boast an NPU. However, it appears that a refresh of sorts, could see the entire stack, desktop included, packing dedicated AI chops at the expense of die space.
As things stand, we expect Core Ultra 200V chips to offer 48 TOPS of AI computing power. This would put them near AMD’s Ryzen AI 300 series and its 50 TOPS. All in all, this should make Arrow Lake-H/U processors some of the best CPUs for workloads involving artificial intelligence. S and HX SKUs, meanwhile, will need to lean on GPU acceleration. For now, anyway.
According to Jaykihn0 on X, Intel may be planning an Arrow Lake Refresh which would include an upgraded NPU on the SoC tile of every CPU. The latter would thus measure 2.8mm longer than Arrow Lake’s die. This change seems to concern both desktop and mobile variants of the architecture. That said, the HX and S variants would maintain the same package size.
The leaker also adds that motherboard compatibility will depend on power mode. For example, if Fast Voltage Mode (FVM) is enabled on the VccSA rail. This means that support will depend on each vendor. As usual, though, take all of this with a grain of salt. We are still far from the release of this rumoured Refresh. Futhermore, Intel is yet to launch its first Arrow Lake series of chips, after all.
NPUs are more prevalent than ever and they’re only going to continue to advance with each passing processor generation. They should alleviate the load on major AI providers as more basic requests can run locally rather than clog up cloud infrastructure, improving privacy for the user too. Whether the silicon cost is worth it remains to be seen. What is sure, though, is that many would prefer to see the silicon put to use for extra cache or other performance enhancing specifications. So, these NPUs better bring some meaningful value.
It’s worth remembering that consumer AI and NPUs are largely still in their infancy. Don’t forget, image upscalers like DLSS and FSR were both relatively poor on release compared to where we are now. Personally, as long as the generational processing improvements are noticeable, Intel and AMD can keep packing NPUs all they want.