Earlier this month, Intel formally revealed its new Panther Lake and Clearwater Forest architecture for consumer and datacenter CPUs, and while we didn’t get the chance to attend the Tech Tour event in Arizona, United States this time around, there are a lot of information to take in from this event. In this article, we’ll focus on the consumer side of things – Panther Lake – and see what you can expect from the next-gen laptop processors by Team Blue.
Meet Intel Panther Lake
Panther Lake will be powering all the next-generation laptop silicon – presumably under the name of Core Ultra (X) 300 despite Intel yet to explicitly confirm this – and above is the overview of the new architecture. Most aspects of the chip have received new upgrades, including CPU, GPU, NPU, IPU, memory, PCIe, I/O, and networking.
All told, these are the numbers you can expect with Panther Lake chips when compared to Core Ultra 200 series processors, which includes Arrow Lake for high-performance laptops and Lunar Lake for premium lightweight laptops.
Panther Lake will be split into three major configurations: the smallest 8-core model featuring 4P+4LPE cores with 4 Xe3-cores; the first 16-core variant with 4P+8E+4LPE cores and the same GPU configuration; and the second, full-fat 16-core variant with 4P+8E+4LPE cores, along with full 12 Xe3-cores on the GPU. Each ‘Tile’ is manufactured from different processes, with the CPU tile all produced in-house using Intel’s latest 18A process, while the 4-core GPU is made using Intel 3 node. All platform controller tiles in this lineup, along with the 12-core Xe3 GPU, are built on external fabs (TSMC).
The two 16-core variants aims to serve laptops of different classes, each with some tradeoffs. The 12-core GPU has lesser PCIe lanes compared to the 4-core GPU variant (since it’s used up by onboard graphics), though it’ll feature the fastest LPDDR5X memory speed at 9600MT/s, so graphical-intensive workloads will be less likely to be bottlenecked by memory subsystem. As for the other 16-core variant, having 20 PCIe lanes should free up some room for additional PCIe 5.0 M.2 SSD support, and it will support standard DDR5 (i.e. SO-DIMM/CSO-DIMM) and LPDDR5X, both of which will be rated at slightly slower speeds compared to the former.
CPU: Cougar Cove & Darkmont
The CPU cores are now updated with new architectures, including Cougar Cove P-cores and Darkmont E-cores. They replace Lion Cove and Skymont/Crestmont respectively from Arrow Lake and Lunar Lake architectures, and this marks the return of sharing identical architectures across E-cores and LPE-cores since the original tile-based design, Meteor Lake (Core Ultra 100).
Intel claims Panther Lake offers over 10% more single-threaded performance at similar power limit over both Lunar Lake and Arrow Lake (both featuring Lion Cove P-cores), and conversely, Panther Lake requires 40% less power to maintain the same performance as both architectures. In terms of multi-threaded performance, Panther Lake is outright faster than Lunar Lake, while the new design also requires 30% less power to match Arrow Lake’s peak performance. While Intel didn’t explicitly mention the uplift at max power, the chart above suggest somewhere in the 10-20% range.
Intel’s Thread Director technology is the chipmaker’s hardware-based scheduler which allocates resources to the best-fit core clusters to maximize performance and efficiency, and it has gone through several generations at this point since its introduction back in Alder Lake (a.k.a. Intel 12th Gen Core). In the Lunar Lake iteration, the logic is pretty simple – start the workload in LPE-core cluster first, and if the performance demands exceeds what this cluster can offer, move them to P-core cluster.
Similar logic applies to Panther Lake as well, except there’s a new E-core cluster acting as the middle ground too. As seen in the scheduling charts above, low-demand workloads like Microsoft Teams puts the workload almost entirely in LPE-core cluster, while demanding ones like Cinebench puts all cores into work. Conversely, in gaming, Thread Director skips LPE-core cluster altogether to avoid performance bottlenecks in slower cores.
There’s also a new featured called Intel Intelligent Experience Optimizer, which essentially automatically adjusts the system’s power profile based on system load. Since most users never touch this power slider in the battery settings (unless you own a gaming laptop and knows the ins and outs of it), this feature takes system load into account and provides the best performance or efficiency as needed. This means performance doesn’t get held back by specific power profiles, hence more performance in the benchmarks shown above.
GPU: Xe3
The GPU within Panther Lake is also getting a major upgrade, now switching to a brand-new Xe3 architecture with upgrades across the board. There’s more cache, improved ray tracing, better power efficiency, broader codec support, multi-frame gen support, and most importantly, more performance outright. Considering Lunar Lake’s Arc 140V graphics is already quite powerful (albeit being held back by drivers as Intel still has some ways to go), expect Panther Lake laptops to take on gaming with some decent framerates to boot.
As a side note, Intel also revealed that the next generation beyond Xe3 will be designated as Xe3P, which will likely power upcoming Celestial dGPUs. Nothing much is said beyond this quite roadmap overview though, since the main focus is still on the laptop chip. With that, let’s have a look at GPU configurations.
In the case Panther Lake, there are two GPU variants available: one with 4 Xe3-cores powering both the 8-core model, and the GPU-light 16-core model. The second 16-core model features the full 12 Xe3-cores GPU, with a claimed 50% faster performance over Arc 140V found in Lunar Lake (Core Ultra 200V) processors. Efficiency-wise, it’s also 40% better at perf-per-watt when compared to Arrow Lake-H models (which uses the original Xe graphics), although this comparison does skew in favor of the newer architecture, so take it with a bit of skepticism.
Gaming
Joining NVIDIA, Intel is also introducing its take on multi-frame generation, dubbed XeSS-MFG. The implementation are very similar to how DLSS MFG works, taking into account in-game motion and depth information to generate the next immediate frame, then using optical flow reprojection to generate subsequent frames until the next native frame is produced. Up to 3 intermediate frames can be generated (4x effective FPS), and Intel says the override option will be available via Intel Graphics Software for games that supports frame generation feature.
On that subject, while some players may prefer enabling MFG for smoother visuals, that number does not necessarily represent the true smoothness of the game, which can manifest as input lag if the input framerate is particularly low. So, for those looking to monitor in-game performance metrics, Intel’s PresentMon tool is getting a new upgrade that displays both the effective framerate (with MFG), as well as the native framerate for reference. Besides that, there’s also the frame type differentiation that tags generated frames in logs, along with the new Animation Error metric that further monitors low-percentile framerate behavior.
When you first launched a brand-new game you just downloaded, chances are you had to wait for several minutes to let the game compile shaders, which ensures that the graphics are functioning properly and avoids stuttering issues when these are compiled in real-time manner. For gamers excited for the latest game release, this can be annoying – so Intel has introduced something called Precompiled Shader Distribution, which functions in similar ways as Microsoft’s own Advanced Shader Delivery.
Essentially, the shader compilation process is moved to GPU vendor’s servers, then stored in a cloud – when a user downloads a game, said user can simply download the compiled shader package that matches their GPU and system configuration, which cuts down on launch time and improves performance. Of course, driver updates will sometimes invalidates that data, which can once again trigger the compilation process, so PSD takes that into account and provides the package accordingly when that happens.
To maximize the GPU power headroom, Intel also brings in Intelligent Bias Control v3, adding on top of the previous version (released earlier this year, now available on Lunar Lake) that made drivers aware of workloads like gaming, which puts the CPU on a more stable clock speed to avoid stuttering as a result of aggressive CPU boosting that ends up throttling the GPU at the same time. The new version goes further and puts these workloads on a low-power E-core cluster, freeing up more power headroom for the Xe3 GPU to push out more frames, as seen on the graph on the right (in purple color).
NPU 5
Next is NPU, now on its fifth generation dubbed NPU 5. The key upgrade this time around is not necessarily the raw compute power itself, which is slightly improved over Lunar Lake’s NPU 4 (48 TOPS vs 50 TOPS); rather, the primary design change lies on the significantly smaller die area thanks to the newly-redesigned compute block. This way, more space is freed up to include other functions of the processor, as NPU is among the most area-expensive block on a modern silicon right now given its relative lack of utilization in software today.
New for NPU 5 is support for FP8 compute, which allows AI models to process faster over typical FP16-based compute. Based on the Stable Diffusion test on the graph shown above, Intel claims the processing can be done faster while saving power at the same time, achieving a 50% increase in effective performance-per-watt despite technically not having any increased compute power over NPU 4 design.
Intel quotes combined AI TOPS figure with all three components each contributing a specific part of AI compute, and in total the system can achieve up to 180 TOPS with CPU, GPU, and NPU all combined. The majority of that increase over Lunar Lake (120 combined TOPS) is credited to the significantly more powerful GPU contributing 120 TOPS on its own, while the CPU and NPU provides 10 and 50 additional TOPS on top of that.
IPU 7.5
Teleconferencing has been quite important especially since the Covid pandemic happened several years ago, which was perhaps a key point in time where laptop makers finally started to take webcams seriously. Since then, the webcams has gradually saw upgrades that improves their ability to capture the subject in various environments, and the IPU 7.5 – a half-generation jump over the IPU 7 found in Lunar Lake – focuses on improved HDR and AI-based quality enhancements.
The new IPU 7.5 features three new blocks – Staggered HDR blending, AI-based noise reduction, and Advanced Tone Mapping. Staggered HDR functions by merging overexposed and underexposed images into one output that retains details on both highlights and shadows. This function is hardware-accelerated and supports 4K resolution, and there’s also power savings to be had as well, which is always a plus in laptops.
The next two new features are straightforward: utilize AI to best optimize the raw sensor image for best quality. Noise reduction process happens on the NPU or GPU through a trained neural network, and IPU takes over the processed image to apply YUV (brightness & color), forming the final output. AI-powered tone mapping makes color adjustments locally based on image content, instead of the one-size-fits-all traditional tone-mapping used in existing implementation.
Connectivity: Wi-Fi 7 R2, Bluetooth 6.0
Finally, connectivity – Panther Lake will integrate both Wi-Fi 7 R2 and dual-Bluetooth 6.0 into the Platform Controller Tile, so laptops with these chips will no longer require a discrete module to attain latest connectivity standards, and that comes with a minor improvement in freeing up additional space on motherboards for things like battery. Intel also updated its proprietary CNVio interface to CNVio3, which increased maximum throughput to 11Gbps (compared to 5Gbps max on CNVio2).
That 11Gbps max speed is one of the few features that differentiates Wi-Fi 7 R2 from the original “R1” standard, which also includes multi-link reconfiguration (reassign bands on devices to faster, underutilized bands and shuts down the slower ones to save power), restricted target wake time (allows devices to “book” time slots on specific channel to avoid dropouts), single-link enhanced Multi-Link Single Radio (quickly switches antennas to available uncongested bands), and peer-to-peer channel coordination (separate radio channels for local/P2P file transfers).
Panther Lake now supports one new feature for Bluetooth LE (Low Energy) Audio called Auracast – this essentially allows Bluetooth audio devices to hear nearby audio sources, essentially like a public radio. This means you don’t have to take off your earbuds to hear something physically close to you as long as the audio source supports Auracast.
Besides that, Panther Lake’s Bluetooth 6.0 supports Channel Sounding allows devices to measure range with high accuracy, making it useful for things like presence-based security and two-factor authentication. Finally, Panther Lake packs two Bluetooth antennas, effectively giving it almost twice the range over predecessor, along with better connection stability (since the signal will be stronger – Intel claims +3 to +5dB – at the same distance).
For Intel Connectivity Performance Suite 5.0, new this time includes advanced Bluetooth monitoring, bidirectional QoS management, and finally, AI-aware QoS. The last one is particularly interesting: Intel will specifically prioritize network traffic that involves AI, so your experience over at websites like ChatGPT will perform slightly better than, say, watching a YouTube video. Hard to say how exactly this will reflect in real-life experience – we suppose the key takeaway here is that companies will make doubly sure that your AI experience will be as smooth as possible.
