Intel Reportedly Considering Resurrection of "Arrow Lake Refresh" Processor Family

[T0@st]

Intel is reported to be eyeing a revival of its "Arrow Lake Refresh" desktop processor design—Golden Pig Upgrade disclosed this "strange" revelation via an updated Weibo blog post. Back in 2023, the Chinese hardware reviewer and leaker extraordinaire theorized that Team Blue's much-rumored shelving of "ARL-S/HX Refresh" came down to cost/benefit considerations relating to accommodating an upgraded NPU in the chip's tile-based design. Last September, reports suggested that Intel had put the final kibosh on a 14th Gen Core Ultra 8P+32E "Arrow Lake Refresh" processor series—leaked roadmaps had a launch window marked down for late 2025, going into 2026. Industry watchdogs have picked up on Golden Pig Upgrade's latest forecast—prompting further theorizing.

Intel has its mobile segment's future covered with Panther Lake (later in 2025) and Nova Lake (2026), but a notable gap exists in their desktop world. Nova Lake's desktop S-series is slated for launch at some point in 2026, so this year could be prime territory for a mild refresh of existing Arrow Lake-S processors—on the LGA1851 socket. Golden Pig Upgrade reckons that refreshes of ARL-S (desktop) and ARL-HX (high-end laptop) are back on Team Blue's upcoming product roadmap. Industry moles reckon that an updated NPU design—potentially similar to the one housed in Lunar Lake mobile chips—will be the only major upgrade lined up for the so-called "Core Ultra 300" processor series. Intel's current-gen flagship model—Core Ultra 9 285K—has NPU performance topping off at 13 TOPS. Refreshed Arrow Lake SKUs could be deployed with an improved NPU aspect, perhaps capable of hitting 48 TOPS.



View at TechPowerUp Main Site | Source

 

[phanbuey]

wow...

of all the reasons ARL-S failed, the NPU is probably on the rock bottom of the list.

 

[usiname]

100Mhz higher clock and LGA-1851 will has "upgrade path" and 2 flops per socket

 

[Bobaganoosh]

wow...

of all the reasons ARL-S failed, the NPU is probably on the rock bottom of the list.

^exactly.

Unless they can actually improve latency and I would even suggest pricing, it's not a competitive platform. I wouldn't think they even need to bump clocks much if they can improve the memory performance. I'm pretty sure that would require moving the memory controller to a different tile though, so I don't know how that'd be considered just a "refresh", which means there's no point in doing a refresh at all. It'll sell just like the current Arrow lineup does.

 

[Vayra86]

Looks like Intel is back to business as usual, just pump out product, maximum confusion.

 

[Nostras]

I'm a bit confused, is the 8+32 variant back on the menu or is Intel simplifying the refresh by just doing a clock bump and an NPU upgrade?
If 8+32 is happening it would make for one hell of a productivity machine.

 

[phints]

265K was my disappointment of the year, was hoping for strong competition. I care about 8-12 good p-cores, don't care about power wasting Cinebench accelerator e-cores, and certainly don't care about their weak ass NPU performance.

Intel spent all that money on TSMC 3N lithography when it should be going to AMD for Ryzen 9000 and Nvidia for RTX 5000, then we would actually see some big gains out of CPUs/GPUs that matter. Instead we got mediocrity across the board.

 

[efikkan]

This looks like to be nothing but baseless speculation;
Firstly, if Intel were to do a traditional refresh (partial or complete), will mainly depend on whether the yields have improved significantly after the QS samples or not. (no point in a refresh if the yields are unchanged)
Secondly, a "refresh" with a bigger NPU, in other words a significant change, is not something they could throw together over night. If there are such products in the works, then they should be at the second iteration of engineering samples by now, and the development have been progressing all along.
So either way, this "source" provides nothing that makes sense. Not to mention that nobody is buying the competition because Arrow Lake lacks a bigger NPU.

While bumping the clock speeds 200-300 MHz probably wouldn't do a lot for the user experience, it will probably help the quite a bit in benchmarks where it falls short compared to Raptor Lake. While Arrow Lake is undoubtedly much more powerful than its predecessor, much of the media attention focuses on the things that really matters; like 720p/1080p gaming with high-end GPUs at lower details and various edge cases and synthetic benchmarks. Joking aside, there is no doubt that I would pick Ultra 9 285K over i9 14900K if I had to choose one of them.

What should have been the response to Arrow Lake is not claiming it's bad, as it's evidently not, but that it's too small of an upgrade to justify a new platform. Additionally, obviously the self-inflicted problem of lacking AVX-512. We are also forgetting that Raptor Lake with its super-aggressive burst speeds have much more inconsistent performance, which is noticeable, and has been a valid reason to consider the competition.

 

[Gunbound]

I am not too familiar with the roadmaps and how intel would do things, but if they did a refresh on ARL desktop I can still expect another generation of SKUs coming afterwards for the z890 chipset right? I ask because usually 2 generations per board with Intel, right? So would a refresh count?

 

[freeagent]

Its ok guys, Intel is just spamming the market like Thermalright did

Soon there will be CPUs for all the coolers!

 

[Squared]

An Arrow Lake refresh ought to have been planned for a long time. Arrow Lake uses TSMC's N3B node, which is the worst N3 node. A refresh moving to N3P or N3X would be much better-suited for desktop because it would reach much higher frequencies and game a bit better. Throw in eight additional E cores for productivity workloads, and it would be a good product. Also, it would be very similar to the upgrade Raptor Lake was over Alder Lake. Still no AVX-512 though.

 

[lexluthermiester]

Seems to me that Intel is just leveraging their assets.

 

[efikkan]

An Arrow Lake refresh ought to have been planned for a long time. Arrow Lake uses TSMC's N3B node, which is the worst N3 node. A refresh moving to N3P or N3X would be much better-suited for desktop because it would reach much higher frequencies and game a bit better.

While I haven't kept up to date with TSMC's nodes, if the nodes are similar enough, it should be feasible to port without a major redesign, allowing not only for slightly better clocks, but also tweaks for better timings. But if this is the case then this design would have been completed and taped out somewhere around last summer. But I haven't heard anything yet.

Throw in eight additional E cores for productivity workloads, and it would be a good product.

Nah, having a bunch more of those is mostly for synthetic benchmarks. It's a waste of die space which could have been spent better.

I liked it back when Intel used to do minor architectural improvements between generations, even if it's just ~3-5% IPC, which would be much more interesting than higher peak clocks. But still, if nothing else, higher sustained clocks would be appreciated.

 

[Ravenmaster]

Releasing a refresh of a failed generation of CPU's would do crazy damage to Intel's already rock bottom reputation

 

[Squared]

Nah, having a bunch more of those is mostly for synthetic benchmarks. It's a waste of die space which could have been spent better.

The alternative for highly threaded workloads (which I do have) is a multi-chiplet CPU from AMD, which also has its downsides.

I liked it back when Intel used to do minor architectural improvements between generations, even if it's just ~3-5% IPC, which would be much more interesting than higher peak clocks. But still, if nothing else, higher sustained clocks would be appreciated.

Lion Cove does deliver around 12% higher IPC, just not in games. (But there's also the clock speed reduction.) And Skymont improved a lot more.

 

[efikkan]

The alternative for highly threaded workloads (which I do have) is a multi-chiplet CPU from AMD, which also has its downsides.

If your depiciton of your workloads is accurate, then it sounds to me like you should take a look at Xeon W and Threadripper. (There is also next-gen Threadrippers pretty imminent, and new Xeon Ws probably this fall, however that may fit on your timeline.)

High-end workstations are much more capable than their spec-sheets may indicate, as they tend to perform better without heavy throttling, offer better memory bandwidth and IO, are easier to cool, etc. Not to mention they are much more durable and built for sustained workloads unlike the mainstream platforms, so if you're doing anything where your time is worth money, you should look at building a proper workstation. Depending on your country, the price difference might not be as big as you might think. The biggest difference is the CPU, workstation motherboards tends can at times be found about the price of "well featured" mainstream motherboards.

Lion Cove does deliver around 12% higher IPC, just not in games. (But there's also the clock speed reduction.) And Skymont improved a lot more.

Arrow Lake is undoubtedly much more capable than Raptor Lake, and as the review at Phronix clearly shows, it's actually about ~12% faster, despite the deficit in rated clock speed. This is due to the fact that Raptor Lake can boost very high in either short bursts or when the computational load is low (which is typical for games), but if you do heavier and longer benchmarks, the benefit from this boosting will decrease.

I haven't tried Arrow Lake yet, but what I can tell from doing work on Raptor Lake vs. Comet Lake, Coffee Lake, Skylake-X, etc. Is that it's certainly faster in peak speed than the Skylake-family (in some workloads is quite significant), but the constant fluctuations in performance is quite noticeable to those who are sensitive responsiveness. This is why I wouldn't be eager for just boosting the upper clock speed limit for an Arrow Lake refresh.

But as for Skymont; These E-cores are incredibly weak compared to what synthetics would lead you to believe. Especially when it comes to computationally heavy workloads and due to their shared L2 cache which makes them perform worse than in isolation.

 

[lexluthermiester]

Releasing a refresh of a failed generation of CPU's would do crazy damage to Intel's already rock bottom reputation

Do explain.

 

[Squared]

If your depiciton of your workloads is accurate, then it sounds to me like you should take a look at Xeon W and Threadripper. (There is also next-gen Threadrippers pretty imminent, and new Xeon Ws probably this fall, however that may fit on your timeline.)

I keep an eye on Xeon and Threadripper. They're really far beyond my needs though. And they tend to come out like a 8 months after the microarchitecture they contain is available in mainstream desktops, which means paying more for what in most of my uses is less.

But as for Skymont; These E-cores are incredibly weak compared to what synthetics would lead you to believe. Especially when it comes to computationally heavy workloads and due to their shared L2 cache which makes them perform worse than in isolation.

They work well in media encoding, which is the workload I'd like to do more but it takes forever and I can't justify upgrading to do it.

Intel Core Ultra 9 285K Review

Finally! Intel's new Arrow Lake architecture is launched. The new CPUs are full of design changes, like removal of Hyper-Threading, new Lion Cove P-Cores, an improved Thread Director and more. In our review we got surprising results that were both impressive and disappointing.

www.techpowerup.com

I haven't tried Arrow Lake yet, but what I can tell... is that it's certainly faster in peak speed than the Skylake-family (in some workloads is quite significant), but the constant fluctuations in performance is quite noticeable to those who are sensitive responsiveness. This is why I wouldn't be eager for just boosting the upper clock speed limit for an Arrow Lake refresh.

I'm not sure, but I think using the N3P or N3X node would allow the sustained boost clock speed to be a bit higher.

 

[efikkan]

They work well in media encoding, which is the workload I'd like to do more but it takes forever and I can't justify upgrading to do it.

The advantage of increasing the E-core count would at best yield a few percent extra for select workloads, and do little or even worsen performance for others (unless you increase the TDP as well).
There is a danger in trying to push the mainstream platforms too far, as they are constrained by TDP etc. This is why we get these platforms with low power numbers on paper, yet the motherboards needs to be designed to handle extreme bursts of power and heat, driving up costs for everyone. Mainstream platforms will always be a compromise, and pushing e.g. core counts too high to make them appear better in synthetic benchmarks may actually hurt real world performance, as it leads to more throttling and inconsistent performance (with stutter etc.). Many buyers "solve" some of these constraints by overclocking, especially raising power-limits and all-core clocks, but this should be avoided for productive works, as it not only shortens product lifespan, but also will lead to lots of file corruption and application crashes.

When it comes to media(video?) encoding, there are many different codecs, encoders/applications and quality settings, and you should consider the characteristics of the specific ones you need. For some, buying a mid-range GPU will do the trick and give you plenty of acceleration, but it also comes at a cost of becoming obsolete very quickly as new codecs/versions arrive.

Another option which I often suggest to people who like to have the latest mainstream platform for gaming etc. and buys the new one every ~3-5 years, yet be productive and do a mix of workloads, is to keep the old one side-by-side (instead of a dual-mointor setup on a single PC), then use them simultaneously, e.g. surfing on the old and the heavy workload on the new, and whenever there is a large batch workload (e.g. encoding), then just switch to the old to continue, whether it's work or play. I switched to dual-PC setup at home many years ago, and find it way more productive than a single PC with dual monitors at work.

There is also the option of offloading batch workloads to a "server", whether it's actually server grade hardware, or just a slightly older desktop PC. Some get some great deals on some used server hardware, but there are noise and power consumption to be considered here.

So I'm just saying you have many options, consider which ones are the best use of your money, time and skills.

 

[Squared]

The advantage of increasing the E-core count would at best yield a few percent extra for select workloads, and do little or even worsen performance for others (unless you increase the TDP as well).

Using a better N3 node means higher possible clock speeds or lower power.

TSMC Details 3nm Evolution: N3E On Schedule, N3P and N3X To Deliver 5% Performance Gains

www.anandtech.com

Raptor Lake uses 10% more power with 100% more e cores.

Intel Core i9-13900K Review - Power-Hungry Beast

With the Core i9-13900K, Intel delivers impressive performance. Our in-depth review confirms: Raptor Lake is the world's fastest CPU for gaming. Even in applications the processor is able to match AMD's Zen 4 Ryzen 9 7950X flagship. If only power consumption wasn't so high...

www.techpowerup.com

This would be a 50% increase in e cores. And if it doesn't move to a better node then it won't solve Arrow Lake's biggest shortcomings so there's little reason to make it.

 

[Darmok N Jalad]

Arrow Lake was an unfortunate but necessary step backward for Intel. It's hard to point to Raptor Lake as a baseline, because that product has not exactly been problem free. It's part engineering marvel (6GHZ) and part engineering disaster (fast and loose with the power management from Intel down to board makers). A refresh for Arrow Lake may not offer much gain for everyday users, but it would extend the life of the socket and chipset (which probably aren't selling all that well), and it would give Intel's partners a new generation to market. A real problem that was created, in big part by Intel, is this expectation of annual refreshes of product lines, even when there's not much new to offer. A faster NPU will hardly matter to anyone, but it creates a justification for new products. We've all seen how everyone is using AI to help sell their wares, so it's a hollow upgrade means to achieve that end.

 

[N/A]

Using a better N3 node means higher possible clock speeds or lower power.

So refresh is just N3P for 200 Mhz higher clocks forgoing the double EUV patterning and cheaper. Hopefully not the N3X with 350% leakage.

 

[efikkan]

Raptor Lake uses 10% more power with 100% more e cores.

If you want to operate it at stock, then at some point throwing extra cores at it will mean more throttling. While it may look great in reviews where benchmarks are short or sometimes without power limits, or where synthetic benchmarks are heavily weighted, in real operation it will usually come at a cost.

It's much smarter to prioritize any room in the "power budget" or "transistor budget" on making exsisting cores faster, as it will scale better than adding more weak cores.