Panther Lake unveiled: A deep dive into Intel's next-gen laptop CPU

Intel’s next-generation mobile processor, “Panther Lake,” builds incrementally on the excellent “Lunar Lake” chip populating laptops right now. But there’s something odd afoot: a “16-core, 12 Xe graphics cores” version that could be Intel’s answer to AMD’s Ryzen AI Max, complete with multi-frame graphics generation powered by AI.

Intel has officially revealed its new Panther Lake architecture, and it gives enthusiasts a lot to chew on: a return to the performance (P-cores), efficiency (E-cores), and low-power efficiency (LP E-cores) cores of Intel’s first Core Ultra chip, Meteor Lake. Intel has returned with a fifth-generation NPU capable of 50 TOPS and an image processing unit (IPU) that actually uses AI for some functions.

Intel’s “Xe3” GPU is worth some discussion all by itself, with its awkward branding and powerful multi-frame generation that will potentially elevate frame rates three or four times what they were before. Can Panther Lake address mainstream laptops, handheld PCs, and this new breed of “AI workstations” AMD’s Strix Halo is aiming for? Well, with three separate processors, maybe. Intel’s confidence in its 18A manufacturing process may be a bit overblown, however, as several portions of Panther Lake are still being manufactured overseas, including the 12Xe version of the new, disaggregated (separate) GPU tile.

In terms of performance, we know some further details. Intel says that Panther Lake’s single-threaded performance should be 10 percent higher than Lunar Lake at the same power. Compared to both Lunar Lake and Meteor Lake, Intel’s Panther Lake offers more than 50 percent better multithreaded performance, Intel says. Intel is also claiming that the total Panther Lake system-on-chip will consume 10 percent less power than Lunar Lake, and demonstrated a trio of laptops running a Core Ultra chip from each generation to back that up. Some of this will be due to the design of the chips; some will come from the process technology, including the 18A manufacturing process that’s become a critical part of Intel’s future.

Intel launched Lunar Lake at Computex 2024, and we had Lunar Lake benchmarking completed by late September. (Intel’s been talking up Panther Lake for over a year, too.) Intel representatives were quite clear that CES 2026 in January will be the launch event for Panther Lake laptops, and they will come to market soon after. I suspect that Panther Lake will be marketed as the Core Ultra Series 3, but that’s just an educated guess.

“Panther Lake literally combines the power efficiency of Lunar Lake and the performance of Arrow Lake in a product family, and we’re using Intel 18A to bring these architectures together,” Jim Johnson, the senior vice president in charge of Intel’s Client Computing Group, said at Intel’s launch event in Phoenix.

Intel is also quite proud of its 18A process, the foundation of Panther Lake, which Intel fellow Tom Petersen called the most expensive die Intel has ever made. “We can confidently say Intel Foundry is in production on the only two nanometer class process that was developed and will be manufactured here in the United States,” Kevin O’Buckley, the senior vice president and general manager of Intel Foundry Services.

Intel’s Panther Lake: what are the features of its three chips?

Intel typically describes the architecture of its upcoming processors during its “Tech Tour” conferences like the one it held to describe Panther Lake near Phoenix, then provides the speeds, features, and prices of the actual processors you’ll be able to buy at launch. This was a combination of the two: some details we know, and some we don’t. Contrast that to Qualcomm’s launch event of the rival Snapdragon X2 Elite: we know how fast each chip is and how many cores are inside them, but not the deep details of their architecture.

Right now, Intel’s “Panther Lake” consists of three chips, primarily consisting of the new “Cougar Cove” P-core and the “Darkmont” E-core and LP E-cores:

• An 8-core chip, with 4 P-cores, 4 low-power (LP) E-cores; 4 Xe3 GPU cores and 4 ray-tracing units; and memory interfaces to either 6800 MT/s LPDDR5x or 6400 MT/s DDR5.
• A 16-core chip, with 4 P-cores, 8 E-cores, and 4 LP E-cores; 4 Xe3 GPU cores and 4 ray-tracing units; and memory interfaces to either 8533 MTs/ LPDDR5x or 7200 MT/s DDR5.
• A 16-core chip, with 4 P-cores, 8 E-cores, and 4 LP E-cores, 12 Xe3 GPU cores and 12 ray-tracing units; and memory interfaces to 9600MT/s LPDDR5x, period.

We don’t know how much power Panther Lake will consume; one executive on Intel’s marketing team said that Panther Lake is “expanding on the segment that we addressed with Lunar Lake,” and that the chip power or TDP will go “up and down” from there. That will obviously affect which products Intel’s Panther Lake chip will fit into. Gameplay demos of an updated Painkiller game running on the chip used both a 45-watt reference platform as well as a 30-watt laptop, and Intel’s Petersen referred to a “max” of 44 watts.

In total, Panther Lake can address up to 96 GB of LPDDR5x memory or 128GB of DDR5 memory.

Intel’s Panther Lake is divided into tiles, or what rival AMD calls “chiplets”: a disaggregated design that allows tiles to be placed and swapped in and out. (Intel even provided a knockoff Lego kit of Panther Lake as a tchotchke.) In Panther Lake, there is a compute tile, a GPU tile, a platform controller tile, and a “base tile” that the other tiles are mounted upon. All of the tiles are connected together via a second-gen scalable I/O fabric. The base tile connects to the active tiles via Foveros 2.5D packaging technology, allowing Intel to stack die one on top of the other.

According to Johnson, the disaggregated tiles means more PC segments and more price points. “Panther Lake will be the most broadly adopted, globally available AI PC platform Intel has ever delivered,” he said.

Right now, none of these Panther Lake chips have product names, and Intel isn’t telling us how fast they’ll run either in sustained or in turbo mode. But they will not have hyperthreading; from Meteor Lake on, Intel rearchitected its performance cores to deliver high enough single-thread performance without the need to add an additional thread. Intel chief executive Lip-Bu Tan has called this decision a mistake, but Intel’s plan to eliminate hyperthreading was already baked into the design.

The GPU tile is kept separate. Separating the GPU tile theoretically means that Intel could just “drop in” a replacement, though it’s a bit more complicated than that. Still, it’s natural to see this and the eventual RTX chiplets being talked about as part of Intel’s Nvidia investment as two points on the same line — though Intel carefully declined to confirm this.

Here’s a detail you may find intriguing: the 16-core, 12 Xe3 chip isn’t designed to connect to a discrete GPU, and that chip only contains a total of 12 PCI Express lanes as a consequence. The 16-core, 4 Xe3 version includes 20 PCIe lanes, which certainly implies that this version may appear in gaming laptops. Intel executives said that there’s nothing there preventing 16-core, 4Xe3 chip from connecting to a discrete GPU, so that seems likely.

“The 8-core will service the thin-and-light [market], the second [16-core version] with the small GPU is going to be attached to a discrete GPU, and the big one will have a life of its own, said Fuad Abazovic, principal at ACAnalysis. “It’s whatever you like with Lunar Lake, but with double the cores and much wider, much bigger GPU, but you still keep the battery life.”

The 8-core Panther Lake chip will have 12 PCIe lanes (8 PCIe 4, 4 PCIe 5), while the 16-core chip will have 20 PCIe lanes (8 PCIe 4, and 12 PCIe 5). The 16-core, 12Xe Panther Lake option drops back down to the 12 PCIe lanes and configuration of the 8-core chip.

Also of interest: integrated Thunderbolt 4, but not Thunderbolt 5. That means another year’s worth of Thunderbolt 4 docks, with discrete Thunderbolt 5 controllers probably only attached to pricey gaming PCs.

Enthusiasts shouldn’t really care which process technology Intel will use to make its chips, though Intel has been nearly desperate to win customers for Panther Lake’s 18A manufacturing process and the subsequent 14A process which will follow it. It had to be embarrassing for Intel to make Lunar Lake’s key tiles at rival TSMC. Now, Intel has started to bring its manufacturing in house once again. All of the compute tiles are manufactured on Intel’s 18A process, but all of the platform controller tiles are made at TSMC, as well the 12 Xe3-core GPU tile. The other 4 Xe3 GPU tiles will be manufactured on Intel’s Intel 3 process.

Panther Lake CPUs: Cougar Cove, Darkmont, and better user adjustments

Inside the CPU tile are the P- and E-cores, the NPU 5, plus what’s known as the Image Processing Unit 7.5 as well as the memory interface.

Panther Lake is a system on a chip, so referring to it as a “CPU” isn’t exactly correct. But onboard Panther Lake are two separate CPU architectures: the “Cougar Cove” performance cores, and the “Darkmont” efficiency cores. Stephen Robinson, its chief CPU architect and an Intel fellow, traced the roadmap: Lunar Lake had two types of cores, but it couldn’t scale to a high frequency. Arrow Lake’s Skymont efficiency cores improved performance, but didn’t offer all-day battery. Lunar Lake offered dedicated power delivery for the E cores, and added additional cache memory, for a total of up to 18MB shared level-3 cache.

Essentially, Panther Lake takes the low-power islands of Lunar Lake, combines them with the additional cores previous generations offered, and upgrades the performance cores for improved single-thread performance. (The latter aspect is usually what gets attributed to how quick and responsive an OS is, at least in the MacOS world.)

Intel’s new 18A process also offers what’s called “backside power delivery,” or PowerVia, which routes power away from the signal logic. That improves frequency while reducing idle power loss — both exactly what Intel wanted to achieve.

According to Robinson, Cougar Cove is “optimized” for 18A, with better branch prediction and memory disambiguation. In Darkmont, Intel achieved better prefetching. It also implemented what it calls “nanocode,” where Intel has taken some microcode and added the ability to decode in each of its parallel front-end clusters, Robinson said.

It all works out to 40 percent lower power for the entire Panther Lake chip at the same performance of Arrow Lake, or 10 percent more performance at the same power, Intel executives said.

Intel also exposes how tasks will be routed through each type of core through a technology called Thread Director. In Panther Lake, Thread Director should feel familiar: a workload will first land on the low-power E-cores, then move to the full-power E-cores if it proves too much for the low-power cores. From there, it will be pushed to the performance cores, Rajshree Chabukswar, the Intel fellow in charge of the technology, said.

Intel’s technology is also smart enough to auto-assign some threads based on the application, so Microsoft Teams will always begin on the LP E-cores and likely remain there. A benchmark like Cinebench, which asks for all cores and threads, will get them across all three different types of cores. In games, however, Panther Lake can do something interesting: assign the game to E-Cores, then to P-cores…but then use some of the leftover power and give it to the GPU.

“Using OS containment zones and some of our graphics driver hints and our power management, we are able to deliver 10 percent [more] frame rate because we are making power headroom available to graphics,” Chabukswar said.

Normally, users can adjust the Windows 11 performance by adjusting the Windows power slider. With Panther Lake, Intel is going a bit further with what it’s calling the Intelligent Experience Optimizer. Intel already has the Dynamic Tuning Utility, which uses AI to dynamically optimize the system for performance, battery life, or thermals — like the “AI Mode” in an MSI laptop’s system settings. Most gaming laptops already have “turbo” or “silent” modes, but the impression that Intel gives is that users will have more input into how the laptop performs.

“It’s purely load-driven. When I see that I need more performance, it will move there,” Chabukswar said. “When I see that I need more efficiency, it will move there automatically.”

In two examples, UL Procyon’s Office Productivity benchmark and the single-threaded Cinebench 2024 benchmark, turning on Intelligent Experience Optimizer boosted performance by 19 percent on both tests.

Panther Lake GPU: 50 percent better performance than Lunar Lake

Intel’s integrated graphics path continues: Meteor Lake included Xe, Lunar Lake added a Xe2 GPU, and Panther Lake moves to Xe3. The killer selling point? Multi-frame generation, which injects AI-generated frames between actual rendered frames.

However, there’s an issue with branding. Intel’s “Battlemage” discrete GPU was considered part of the Xe2 generation, while Intel’s Panther Lake contained “Xe3” cores. Yet both Battlemage and Panther Lake are also both part of the Intel Arc B-series of products, which doesn’t make much sense. Intel did show a roadmap that signals a “Xe3P” variant is coming, however.

Panther Lake, from what Intel showed, will have both 4Xe and 12Xe variants. Not everything scales linearly; while the 32 XMX engines on the 4Xe variant triples to 96 in the 12 Xe version, the amount of level-2 cache quadruples from 4MB to 16MB.

Who is the 12Xe version of Panther Lake aimed at? “The cheesy answer is ‘everybody,” said Daniel Rogers, vice president and general manager of PC products at Intel, in an interview. “I think you’ll see it show up in a few ways, One, maybe the most obvious is for gamers. In some gaming designs and some handhelds as well — that’s a good fit.”

Rogers said the 12-Xe Panther Lake will also be the “flagship solution for commercial notebooks on the AI side,” Rogers said.

Is it Intel’s answer to AMD’s Strix Halo? “Everybody’s chasing high-performance local AI, and we will certainly do the same,” Rogers said.

The Xe3 engine also improves the ray tracing, eliminating backups in the pipeline through a new thread sorting unit. It also doubles the anisotropic filtering on die, Petersen said.

The Xe3 GPU was re-designed for scalability, Petersen said. In the GPU’s render slice, Intel increased the number of Xe cores from four to six, increased the level 1 cache from 192 kilobytes to 256KB, and upgraded the level-2 cache from 8 megabytes to 16MB, reducing the need to access the local memory and increasing performance. Intel also moved to a variable allocation strategy when assigning threads, with “dramatic” effects on performance, Petersen said.

While Intel declined to show actual benchmarks, Petersen did show how traffic to the SOC’s memory fabric dropped by 17 percent to 36 percent on key games. He also demonstrated some “micro benchmarks,” or internal tests that Intel itself uses to determine generation-over-generation performance in specific tasks. The X3 architecture also improves the way in which a frame is rendered, including DirectX calls and pre-rendering, but Petersen said real gains were granted by the larger level-1 cache. All told, Intel was able to cut the time to execute one frame to about 22.84ms on the 12Xe chip, versus 45.44ms for Lunar Lake.

Essentially, Intel can generate about 50 percent higher GPU performance than Lunar Lake, or about 40 percent more performance per watt than Lunar Lake, Petersen said.

Multiframe AI rendering hits Panther Lake’s integrated graphics

The Xe3 GPU also supports cooperative vectors, which Intel showed off with Microsoft this June. Cooperative vectors use AI as a way to replace the overwhelming requirement for shaders. Microsoft has also proposed storing them in tbe cloud. Cooperative vectors, in Petersen’s words, essentially “replace that render pipeline, that raster process, with a 3D model.”

“So effectively, we can completely eliminate the render pipeline and replace it with a per pixel AI,” Petersen added.

“You can see this might be kind of a look forward into a future where graphics is really substantially different, and it is primarily AI,” Petersen added.

That future is already here…sort of. XeSS 2, launched alongside Intel’s “Battlemage” GPUs, launched XeSS Super Resolution, XeSS Frame Generation, and Xe Low Latency, all designed to speed up the graphics powering video games. XeSS Frame Generation injected an AI-generated frame between two “real,” rendered frames, and used the low-latency technology to offset the delay that engendered.

“Hybrid rendering is kind of where we are for most titles today,” Petersen said. “Things have indeed moved forward, and we are now into the world of hybrid rendering, plus AI. And in this world, not all pixels are rastered. In fact, most pixels are not rastered. They are generated.”

Now, Intel has launched XeSS-MFG, or XeSS Multiframe Generation, which can inject as many as three additional interpolated frames, in conjunction with the existing upscaling and multiframe generation technologies. You’ll be able to control it as paert of Intel’s existing Intel Graphics Software package, where you’ll have the option of setting the additional frames, or letting the application itself make that decision, Petersen said. That software will also offer options such as specifying how much system memory is shared with the GPU, a feature Intel announced earlier and that AMD has exploited for better AI.

The idea is simple: more frames means a smoother gaming experience, and happier gamers with PCs which can exceed playable frame rates. Even games designed to benefit from XeSS 2 will support XeSS 3, Petersen said.

Yes, those frames will introduce lag. Each generated frame will add about 1.5 milliseconds of latency per frame, with about 1 ms of setup, Petersen said, or about 6 ms per three generated frames. That’s bad when the frame rate is always high, he said, but when frame rates are at 25 to 30 fps, the latency isn’t as bad versus the value of the added frames.

According to Petersen, the latency or lag that bothers gamers isn’t really the “click to photon” latency where there’s a delay in the player’s action. The more noticeable lag is what he called “motion to photon,” where the latency between a user’s input and perceived motion creates nausea in VR and wobble when mousing. Eventually, that could be solved by AI prediction of mouse movements or a variable rate of rasterized to AI- rendered frames.

“There’s a whole bunch of technologies that have not been announced, that are not done, that could make that [latency] better, but right now you can see some,” Petersen added. “And if you’re more latency sensitive, depending on the game type, I would say, turn off frames and just run regular frames.”

In addition to working toward cooperative vectors, Intel is also launching precompiled shader distribution, where your PC doesn’t have to wait to compile shaders, it just downloads precompiled shaders from the cloud. It is also developing Intelligent Bias Control version 3, first launched this summer, in which the CPU and GPU talk to one another and route power to the appropriate logic. It’s the complement to Thread Director, where GPU performance could be increased by 10 percent by that technology alone.

Finally, Intel’s PresentMon software will differentiate between rasterized and rendered/generated frames, and provide an indication whether a given frame was real. It will also demonstrate animation stutter, if frames are out of sync.

According to James Sanders, a senior analyst at TechInsights, said that there’s potential for creating a market for thin-and-light gaming laptops with the 16-core, 12Xe3 Panther Lake part.

“Games that are like Marvel Rivals I would imagine would be a really good fit for that type of hardware,” Sanders said. “If you’re doing esports, that’s something that would be a good fit. But if you’re playing Final Fantasy, where you’re worried more about the graphics performance, you’ll want to go to a discrete GPU. But it is clear that that communicating what that market is going to be something that Intel and their OEMs need to work on, and that’s going to take time.”

Panther Lake’s NPU: more effective, but not much changes

The story of Panther Lake’s NPU: it’s about the same as Lunar Lake, just more efficient. Lunar Lake’s NPU 4 supported 48 TOPS. The NPU 5 inside Panther Lake supports 50 TOPS in a much smaller package.

An NPU basically performs a ton of multiply-accumulate operations, performed by what’s known as a MAC array. Lunar Lake and Panther Lake both have the same number of MACs — 12,000 — but Lunar Lake spread those out into six neural compute engines, while Panther Lake has only three. Panther Lake’s NPU 5 also has 4.5MB of scratchpad RAM, 256KB of level-2 cache, and 6 SHAVE DSPs. Panther Lake’s NPU can perform 4,096 MACs/cycle at int8 (8-bit integer), 4.096 MACs/cycle at FP8 (8-bit floating-point), and 2,048 MACs/cycle at FP16. Those numbers become more significant and more familiar when you’re performing generative AI functions, where you’re specifying the complexity of a particular task.

“What we really wanted was more MACs and less of the other stuff,” Petersen said.

Some of the improvements are just Intel trying to be smarter about how it handles data. Running Stable Diffusion’s AI art generator, the algorithm doesn’t really need the more complex FP16 granularity. Instead, FP8 can be used; in Petersen’s example, energy usage dropped 35 percent from 108 joules to 70 joules.

Panther Lake also includes what’s known as the Image Processing Unit 7.5, responsible for taking the image your PC’s webcam sees and making it look its best.

“That means handling the full spectrum of lighting from bright outdoor sunlight streaming in through a window to dim corners of a warehouse at night,” said Tomer Rider, the IPU product marketing manager. “It means capturing and preserving fine details and high resolutions without introducing the noise, so what you see is crisp and accurate. It also means creating a life like natural image, vivid colors, true to life, skin tones and high frame rate that makes everything feel real wherever you are and whatever the line is, the IPU makes sure that the image looks its best.”

For Panther Lake, the IPU does three things: it adds enhanced HDR for wider dynamic range, and includes AI-based noise reduction and AI-based tone mapping. The “staggered HDR” takes both a short exposure and a long exposure and merges them together. All of these work with both integrated laptop webcams as well as standalone devices. Up to 5 Mpixel sensors are supported, which translates to 2560 x 1920 pixels in the real world, all whie shaving off 1.5 watts or so of power.

AI noise reduction actually uses AI to filter out noise, especially in low-light environments, Rider said. And tone mapping divides up an image or video into different regions, then uses AI to improve the visuals.

In this case, the IPU may be actually asking the NPU to perform these functions. “Either the NPU or the GPU” will handle the AI, Rider said. “We’re working with all the NPU and GPU teams in order to make sure that the system makes the right choice when selecting whether it’s the GPU or the NPU.”

Panther Lake’s wireless features are surprisingly cool

Finally, Panther Lake also includes a wireless module, which places the Bluetooth and Wi-Fi MAC onto the chip and separates the remainder into a separate die inside the package, known as Whale Peak 2. Panther Lake supports both Wi-Fi 7 and Bluetooth Core 6.0, but with some neat twists.

The most interesting wireless features Panther Lake adds are part of Bluetooth. Panther Lake now supports Auracast, a wireless technology launched in 2024 that supports simultaneous playback across multiple devices, so that you don’t have to share an earbud with a friend. Intel’s chip also supports platform sounding, which uses distance- and phase-change modeling to track the distance between two devices, to better locate them quickly. Bluetooth also uses both wireless antennas, and not just one, which should lengthen Bluetooth connectivity. Intel fellow and wireless CTO Carlos Cordeiro used 52 meters as an example distance.

Panther Lake actually supports what’s known as Wi-Fi 7 Release 2, which haven’t been officially certified. Basically, the R2 features facilitate better communication between the access point and your PC, resulting in fewer dropped connections and improved communication speed and latency, Multilink reconfiguration allows the router to shift your PC from one channel to the other, while restricted TWT allows the two to figure out which applications (and devices) deserve the most priority. Single-link eMLSR allows one of the two antennas on your laptop to “sniff” out a different channel, and quickly shift if it opens up. It also allows the laptop to signal of it’s using a channel for peer-to-peer communication.

You probably don’t use Intel’s Connectivity Performance Suite, a small Windows app which allows you to see the various access points available to your PC and allows it to prioritize voice and video calls or streaming apps. Intel is adding AI to this list of options, so if yor PC is having a prolonged session with a cloud service like ChatGPT, you can ensure that that traffic is given priority.

What’s next for Panther Lake?

Intel’s deep dive in the desert certainly opened the doors to reveal a number of features about Intel’s next-gen mobile chip. What we still don’t know is how many versions of the chip itself will eventually ship, and what they’ll be called. Reports from Asia now indicate that the 12Xe version will be branded as the Core Ultra X, to differentiate them from the “vanilla” version of the Core Ultra chip itself.

Naturally, we’ll have to wait even longer for the first performance testing of the chip itself, as well as the subsequent announcements of laptops using the chip. Those should happen at or near the time of the CES 2026 show in January, we’re told. It’s there we should expect customer announcements, as well.

It was a little surprising not to see any future Intel CPU roadmaps, at all, after Panther Lake has been talked about for over a year. “Nova Lake” should be the next step.

“Meteor Lake served its purpose,” said Mario Morales, the group vice president and general manager of semiconductors at IDC. “You needed to have a dog in the race. That can be easily relegated into the lower end. Lunar Lake came in and hit the sweet spot, and it did emphasize a lot more performance efficiency, which I think was another important message for them as they introduced new P-cores and E-cores into the…PC space. Panther Lake is important for them, as it’s the next step. Nova Lake is what gets them to par — with AMD for sure, but also maybe with Apple.”

Still, we now know details of Qualcomm’s Snapdragon X2 Elite as well as Intel’s Panther Lake. Waiting in the wings: AMD. Is Gorgon Point the next name you’ll need to know about in laptop processors for 2026?

Disclosure: Intel held its press briefings in Phoenix, and would not pre-brief reporters in other locations or over video meetings. The company paid for my room, boarding, and travel expenses, but did not ask for or exert any editorial control over this story or other PCWorld content.