Intel’s Panther Lake points to a future where integrated graphics doesn’t suck

Intel’s Core Ultra Series 3 platform, Panther Lake, and its champion, the Core Ultra X9 388H microprocessor, offer something unique: powerful, gaming-class 3D performance with battery life that’s almost unheard of in the laptop space.

Intel positioned the Core Ultra Series 3 (Panther Lake) as a chip with the computational power of its Arrow Lake platform, with the low power consumption of the Core Ultra Series 2 (Lunar Lake). The chip maker also predicted that Panther Lake’s gaming performance is roughly equivalent to a laptop with an Nvidia GeForce 4050 laptop chip inside it. As I’ll show you, those are relatively fair claims.

Instead of just a battery of tests, we’ll try to pull out the “story” of Panther Lake, demonstrating its strengths and weaknesses as we go. Let’s just hope you can buy one.

Intel’s Core Ultra Series 3 chip stands alone, for now

In October, Intel originally positioned the Core Ultra Series 3 (Panther Lake) in one of three basic configurations, combining the new “Cougar Cove” P-core and the “Darkmont” E-core and Low Power E-cores. At the high end was what Intel referred to then as the “16 core 12Xe” configuration, with 4 P-cores, 8 E-cores, and 4 LP E-cores, 12 Xe3 GPU cores, and 12 ray-tracing units. When it came time for Intel to announce the Panther Lake chip lineup, that configuration included its formal name, the Core Ultra X9 388H, with the “X9” prefix added to highlight the largest Xe3 configuration.

Intel then let reviewers benchmark the Core Ultra Series 3 chip during CES, but only using games. It was our first indication that Panther Lake could be something special.

Intel prevented reviewers from testing CPU-specific benchmarks, however, probably because the number of cores inside the highest-end Panther Lake chip (16) are less than those inside the rival Qualcomm Snapdragon X2 Elite Extreme chip (18), meaning that Intel would likely lose to the Snapdragon on paper in multi-threaded CPU-specific benchmarks.

The third contender will be the AMD Ryzen AI 400, an upgrade to the excellent Ryzen AI 300, which AMD debuted last year. The Ryzen AI 400 includes just 12 cores, but runs them at a maximum clock speed of 5.2GHz — the fastest speed of all three chips. But laptops with either the Ryzen AI 400 or Snapdragon X2 Elite aren’t yet available.

CPUs don’t suck any more

Both Intel’s Core Ultra Series 1 (Meteor Lake) and Core Ultra 2 (Lunar Lake) were surprisingly average in CPU performance, both in single-core and multi-core tasks. (CPU-specific applications include web browsing, apps like Excel, compiling software, some games, and decompressing files.) AMD’s Ryzen and Qualcomm’s Snapdragon used to hold the advantage. No longer — well, at least compared against its older rivals.

With Panther Lake, Intel has regained its leadership in CPU computations.

Intel provided us an Asus ZenBook Duo (UX8407A) with an Intel Core Ultra X9 388H chip inside as a launch laptop for the Panther Lake platform. It was a slightly odd choice; the ZenBook Duo is a dual-screen laptop, with a gaming-class 99 watt-hour battery, which drastically inflated the battery life.

Intel also offered a prototype Lenovo laptop, which we used as a reality check for the estimated battery life and additional benchmarks. I left the ZenBook Duo in “clamshell” mode, only using one of its 2K screens to render data to produce results I felt confident in comparing to other platforms. I started using the Cinebench 2024 and Geekbench synthetic CPU tests.

To address Intel’s claim that Panther Lake offers the CPU performance of the mobile Arrow Lake chip: yes, that’s true. Intel launched the Core 285H chip last year, and in our review of the Core 285H, I found that the Cinebench 2024 score was 1,012 (multithreaded) and 128 (single-threaded), just a hair under Panther Lake’s performance. In Geekbench (measured below), the older 285H produced a score of 16,755, again slightly less than Panther Lake’s Core Ultra X9 388H.

But if you’re a Windows fanatic like we are, you might be disappointed by the Core Ultra X9 388H’s showing. Referring to the review of the Apple M5 MacBook Pro, our colleagues at Macworld report that the MacBook M4 Pro reported a score of 1,010 in Cinebench 2024 and 14,763 in Geekbench 6. But the MacBook M5 Pro scored 1,126 in Cinebench 2024 and 18,013 in Geekbench 6, besting Intel’s current mainstream laptop chip.

Perhaps an upcoming HX version can do better?

Battery life is massive! But so is the battery

Intel has claimed that Core Ultra 3 laptops will have up to 27 hours of battery life. That’s true — but, as is often the case, it depends. The two screens of the ZenBook Duo suck more power than a single display. However, Asus installed a 99Wh battery inside. That’s a gaming-class battery, and the largest capacity allowed on a plane by FAA rules. In this case, it’s like bolting a self-powered fuel truck to a sedan.

So yes, the battery life was insane: about 22 hours on the ZenBook Duo running on a single screen and 25 to 28 hours (1,704 minutes) on the Lenovo prototype laptop that we used in early tests at CES. Those tests were performed by looping a 4K video until the battery expired. When asked to do a bit more work (simulating office work via the Procyon Office benchmark) battery life dropped to “just” under 14 hours on the ZenBook Duo — still basically the best results we’ve ever seen. We’ll break down the battery life a bit more on the Asus ZenBook Duo in our upcoming, dedicated review.

Again, Intel wants us to believe that the Core Ultra has the performance of its “Arrow Lake” chips with the power draw of its Core Ultra Series 2 (Lunar Lake) chips. We can check that, sort of, by tracking the power consumption of a Lunar Lake and a Panther Lake notebook as they undergo a benchmark. In idle, the Core Ultra 3 chip draws about five watts, but can drop down under a watt. Lunar Lake averages about three watts or less in idle.

It’s not apples to apples, though. Intel used TSMC’s N3 process technology to manufacture the CPU tile in the Series 2 Lunar Lake chip, while Panther Lake uses Intel 18A, with some tiles split between the two companies. In this case, Intel’s older Lunar Lake is a 17W TDP chip, while Panther Lake is 25W — more power to the chip typically means better performance and worse battery life, but the larger battery and Intel’s architecture seem to offset this.

Here’s a power graph showing the two chips in idle, then running a benchmark, then dropping down into idle once again. This graph just measures the power going into the CPU package, not the entire laptop. That power could vary significantly, and is best left to the battery-life comparisons you’ll find in our individual laptop reviews. Still, Panther Lake is throwing a lot more power and performance at the benchmark, and this graph demonstrates that if a Lunar Lake and a Panther Lake laptop contained the same battery capacity, the older Lunar Lake laptop could win.

Still, if Intel convinces laptop makers to add larger batteries to Panther Lake laptops, though, look out. Laptop battery life numbers could explode upwards!

Performance still drops while on battery

One of the interesting things about Qualcomm’s Snapdragon chips is that they run at full power all the time. Intel’s Core Ultra chips do not, clocking down to lower power consumption, extending battery life.

I run all of our benchmarks on wall power, battery power, and at Windows’ maximum allowable settings, just to see how performance varies in different user scenarios. As you’ll notice in our Cinebench 2024 benchmarks, the single-threaded performance usually associated with OS tasks remains unchanged between wall power and battery power, keeping Windows as responsive in both scenarios.

But look here: Intel’s Core Ultra Series 3 chips seem to maintain their performance on battery much better than Intel’s Core Ultra Series 2 or Series 1. We’re using three real-world benchmarks to test this. First, here is Procyon Office, which performs various tasks in Microsoft Office / 365. Performance drops by about 20 percent on battery.

However, on our custom real-world Handbrake test, where the laptop is asked to transcode the open-source Tears of Steel movie, performance dropped by just three percent between wall power and battery.

Here, you can see how our test Panther Lake laptop fared compared to the competition. This is a custom test, different than the one we run in as part of our laptop reviews. I also made sure to download an Arm-specific version of the app, but Qualcomm’s chip fared exceptionally badly here. It usually performs quite well.

Since we’re looking at real-world benchmarks, we can see that Intel’s Core Ultra 300 / Panther Lake fares well in PugetBench’s Photoshop test. The test uses the shipping version of Photoshop. Here, performance dropped just about three percent on my tests when I unplugged the laptop.

Keep in mind that CPU-specific tests are one of Snapdragon’s strengths. And with the Snapdragon X2 Elite generating exceptional CPU performance — preliminary numbers crush Panther Lake, and the Elite X1 still ranks highly — this might be an area where Qualcomm catches up. This race ain’t over.

Unfortunately, Puget Systems’ PugetBench benchmark hadn’t caught up to the version of Adobe Premiere Pro (26.0) that Adobe makes available for download, so I was unable to test that application.

Panther Lake’s GPU performance is incredible

Remember, Intel’s flagship Panther Lake chip is the Core Ultra X9 388H–this is different. The Core 9 is now the Core X9, which means the GPU has 12 Xe3 cores. Essentially, the “X” means that you’re getting the best Intel has to offer in terms of graphics.

What does this mean? For years, integrated graphics has been able to play games: older, 2D sprite-based games, and some older 3D games at lower settings. They ran. And that was fine. With Panther Lake, we’re navigating a transition into integrated graphics performing almost as well as gaming-class discrete graphics — and when you add AI upscaling and frame generation to the mix, recent top-tier titles are near your grasp.

Some gamers refer to those as “fake” frames, which is why it’s helpful to look first at both traditional, non-accelerated tests. Here, we use UL’s 3DMark, specifically the Time Spy and Steel Nomad Lite benchmarks.

A terrific increase in gaming performance

This was one of the big stories of CES 2026: Intel’s claims that Panther Lake offered the power of a gaming laptop with a discrete Nvidia GeForce 4050 GPU, but inside an integrated package.

This, for me, was the eye-opening moment. A year or so ago, I was testing Intel’s Core Ultra 1 (Meteor Lake) and Core Ultra 2 (Lunar Lake) with custom runs of games like Cyberpunk: 2077 at Low settings, which we show below.

But those tests prompted me to “graduate” Panther Lake into our gaming benchmarks, too, with the settings that traditionally more powerful laptops now use. Even using our aggressive gaming settings, a game like Shadow of the Tomb Raider reaches playable frame rates. (Skip down to find these results.) Yes, it absolutely is an older game, dating from 2018. Yet Shadow was a top-tier AAA title, and integrated graphics has caught up. And that’s just pure, unadulterated, farm-to-table frames, too.

Don’t get too excited, though. Metro: Exodus was released in 2019, but its 4A Engine remains out of reach for Panther Lake. On our test laptop, the game averages 24 frames per second when run at 1080p on the Highest setting — 35 fps if Windows’ performance settings are cranked to their maximum.

AI frames make an enormous difference, if they’re supported

It feels very strange to test Intel’s Core Ultra Series 2 chip using dialed-down benchmarks centered around 1080p gaming at Low settings — often a hint for a PC gamer that it’s time for a new machine or card. But the Core Ultra Series 3 hit the 60 fps threshold that signaled a “playable” game even with just rendered frames. Panther Lake’s GPU also includes two different methods of artificially increasing frame rate — 2X upscaling, or rendering a frame using a lower resolution and then increasing the resolution to the desired level — and XeSS 3, which can interpolate three additional frames using AI. Naysayers call these “fake frames,” but Panther Lake allows for purists and more aggressive gamers alike to find what they want.

Our test laptop shipped with Intel Graphics Software, a custom Intel app that allows you to control various aspects of your display and graphics — including forcing on XeSS frame generation, or AI-generated frames that can inject up to four interpolated frames for every frame the GPU renders. That’s big — or is it?

What I discovered is that, yes, turning on frame generation can make an enormous difference. Simply turning on upscaling and XeSS 3 increased the framerate to a whopping 140 frames per second! Dialing up the Windows power slider tacked on a few additional frames. Both are included in the “Max” result at the top of the chart.

The effects seem to differ depending upon the image quality, though. When running Cyberpunk on our traditional 1080p Ultra settings, frame rates jumped from 52 to 92. Pushing the Windows slider to maximum performance gave me frame rates of 143 fps.

The difference, though, is that Cyberpunk specifically supports XeSS modes. Metro: Exodus does not — and “forcing” XeSS on using the Intel Graphics Software app didn’t work. Modern games seem more forgiving of older hardware, and support for AI frame generation certainly makes those games playable by modern laptops. Still, I wonder if there will be a tier of AAA games like Metro: powerful enough that Panther Lake laptops won’t be able to run them, but old enough that they won’t be able to support the frame generation that would otherwise bridge the gap.

I tried a handful of other games. Total War: Warhammer 3 crashed when running the “battle” benchmark, but its campaign map benchmark played back at 44 frames per second at 1080p High settings. The 2014 Thief remake produced a even 60 fps when played at 1080p at the Highest settings. Neither supported XeSS or any frame generation. Forza Horizon 6 generated 62 fps on 1080p Ultra settings with frame generation forced on, but without explicit support for it.

Can Panther Lake compete with a 4050 laptop?

This was the most provocative claim that Intel made about Panther Lake at CES, right before we had a chance to test out the chip on a prototype Lenovo laptop. Using purely rendered frames, it falls a bit short. When frame generation is included, it keeps up.

Would a gamer with a desktop PC running a GeForce RTX 5090 turn off AI frame generation? Possibly. I think that most enthusiasts, already feeling the pinch of skyrocketing RAM, SSD, and GPU prices, will turn on frame generation without much thought. Again, here’s the Core Ultra X9 388H running Shadow of the Tomb Raider, without frame generation, facing off against a number of existing, but older gaming laptops.

And here is the Core Ultra X9 388H running Cyberpunk 2077 with frame generation enabled. This feels like a scene from an 1980s TV show, where Voltron finally pulls out his blazing sword or K.I.T.T. goes into turbo mode. The episode would be a lot simpler if both had happened from the get-go.

Basically, setting aside the scorn some have for AI and “fake frames,” AI frame generation is the “win” button here.

AMD’s Ryzen AI Max is another option

AMD tried to work the refs (us) harder than Indiana football coach Curt Cignetti complaining about personal fouls during a halftime interview. The company claims that we should be comparing Intel’s Panther Lake to AMD’s Ryzen AI 400 chips as well as its Ryzen AI Max processor instead.

To that, we say, ship us one! We’re happy to review the Ryzen AI 400 when laptops finally are available. As for the Ryzen AI Max, well — we’ve reviewed it inside a (Framework) Desktop, and we’ve seen it in an HP ZBook Ultra G1a laptop, too. As our review benchmarks show, the Ryzen AI Max outperforms the Core Ultra X9 388H handily, though we’d probably put it in a tier that Intel’s eventual HX gaming processors will eventually compete against, rather than a power-sipping laptop chip.

AI is less important than before

But the Ryzen AI Max does have a point, so to speak. If people do want to run private LLMs locally, the Ryzen AI Max (Strix Halo) does provide gobs of VRAM necessary for such LLMs to run. An AMD driver allowed the Framework Desktop to assign 96GB for running LLMs. Our Asus ZenBook Duo review unit, which has an Ultra X9 388H and 32GB of RAM, supplied 18GB of VRAM for games and AI applications. That includes an NPU that can provide 50 TOPS, or 122 total TOPS with the GPU roped in.

And let’s face it — AI has struggled on the PC, leaving us wondering a bit if the early emphasis on the NPU was worthwhile. What we do know is that the graphics chip is the most powerful AI processor. UL provides several benchmarks; I’ve ditched the abstract “Vision” benchmark in favor of the Procyon image-generation (AI art) benchmark. (The test is a work in progress, excluding Arm and providing an odd implementation for AMD’s Ryzen processor.) But UL’s test can generally run on either the NPU or the GPU, with some exceptions.

Basically, this test reflects the score UL assigns to the process. In the real world, it shows that the ZenBook Duo with a Core Ultra 3 chip inside creates a 512×512 image once every 4.5 seconds using the GPU, while our test laptop with Intel’s Core Ultra 2 chip inside creates the same image once every nine seconds. But the Ultra X9 388H’s NPU performance suffers, and the AMD’s Ryzen AI NPU outperforms it, too.

The same goes for running UL’s LLM benchmark. Originally, this test was one of the few that evaluated the NPU, and that was useful. But as Procyon begins adding support for the GPU, it does make you wonder why we’re using an NPU when a more powerful alternative is right there.

Procyon’s test loads and runs several models, then provides a series of prompts and generates a score. Some tests simply don’t run on some processors (Arm, again) and only run on a couple of the others. Some tests will only run on the NPU. This test is really best to compare the three generations of Intel Core Ultra processors.

Again, the test doesn’t do a great job in describing real-world results. In this case, the plugged-in Core Ultra 3 system running the LLama 3.1 (8 billion parameters) on the NPU generated about 20 tokens per second, which would appear on your screen at about four characters per token per second–a comfortable reading speed for me. Using the GPU under the Windows balanced settings, the token output was about 25 tokens per second for the same model. Running Llama 2 (13 billion parameters), the token output was between 13 to 15 tokens per second, which might be a little slow.

I considered noodling around with Intel’s AI Playground, but the app stalled out when preparing the llamaCPP-GGUF backend, so I abandoned the project.

2026 will be an interesting year

I honestly thought that 2025’s crop of laptop processors were the best ever — you could buy a laptop whose processors were made by either AMD, Intel, or Qualcomm and go away happy. But 2026 looks like it could be even better.

Remember, though, that Intel is first out of the gate with this new generation of chips. AMD will eventually answer with the Ryzen AI 400, and Qualcomm’s Snapdragon X2 Elite Extreme is waiting in the wings. Given that Intel traditionally commands about 80 percent of the notebook PC processor market, an early jump could be a powerful advantage, especially with only older chips to compare it to. But we’re not done yet!