Panther Lake sets stage for Intel's 2 nm comeback, but many details still TBD

Notebook chip promises 8 to 16 cores and up to 180 TOPS of total AI performance when it hits shelves in January

Intel has begun clawing back production from TSMC with the introduction of its Panther Lake processors, the company's first chip based on its long-awaited 18A process tech.

The x86 giant began volume production of the processor's compute tiles at Fab 52 located at its Ocotillo campus in Arizona this summer, with the first notebooks and computers based on the SoC slated for release around the time of CES early next year.

If Intel is to be believed, the process and architectural improvements introduced with Panther Lake will be well worth the wait. Chipzilla claims the process will deliver up to 10 percent single-threaded and 50 percent in multithreaded uplift in performance per watt over last gen's Lunar and Arrow Lake processors.

Process shrink meets architectural overhaul

Much of Panther Lake's efficiency and performance gains can be attributed to the chip's new process tech, which, in addition to a node shrink from Intel 3 to a 2 nm-class process it calls 18A, shifts power delivery to the back of the wafer.

According to Intel, 18A's denser RibbonFet transistors account for more than a 15 percent increase in performance per watt, while backside power boosts transistor density by up to 30 percent.

However, Intel hasn't quite quit TSMC just yet. Only Panther Lake's compute die, which contains the CPU, NPU, and image processors, is actually taking advantage of Intel's 18A.

Panther Lake isn't one chip but rather three interconnected by a base-die - Click to enlarge

Much like Meteor Lake, Intel's latest client CPUs use a heterogeneous, multi-die chip architecture, which combines silicon from various wafer fabs and process nodes into a single package. For Panther Lake, the chip comprises three key components:

1. A compute tile fabbed on 18A
2. A platform controller tile built by TSMC
3. A GPU tile built on either Intel 3 or TSMC depending on the SKU

These three chips are packaged alongside a few filler tiles on a base die that facilitates power and communications between them using Intel's Foveros advanced packaging tech.

Unpacking the Panther Lake line-up

Speaking of SKUs, Intel will offer Panther Lake in at least three configurations: a base model with eight cores and two higher-end options with 16 CPU cores and the option of either additional PCIe connectivity for dedicated graphics or a beefier iGPU.

Here's a quick overview of Intel's entry-level Panther Lake SKU - Click to enlarge

Working up the stack, the base model features a compute tile with eight CPU cores. Half of those are Intel's new Cougar Cove performance (P) cores while the four remaining cores are arranged in a low-power efficiency (LP-E) core cluster based around its Darkmont microarchitecture.

If Darkmont sounds familiar, that's because it's the same underlying architecture that underpins Intel's datacenter-centric Clearwater Forest Xeons we looked at during Hot Chips earlier this year.

Located alongside the P and LP-E core clusters is a 50 TOPS neural processing unit and an upgraded image processing stack, which are shared across the entire line-up.

I/O functionality is handled by a Platform Controller Tile that offers four Thunderbolt 4 ports and two USB 3.2 ports, Wi-Fi 7, Bluetooth 6.0, four lanes of PCIe 5.0 for storage, and eight lanes of PCIe 4.0 for additional connectivity and peripherals.

The eight-core chip comes equipped with a quad-core GPU tile fabbed on the chipmaker's older Intel 3 process tech. Xe3 is new for the Panther Lake generation and represents a modest but not insignificant upgrade over the Xe2 cores found on last gen's Lunar Lake processors.

The 16-core version of Panther Lake can be had with up to 20 lanes of PCIe connectivity, opening the door to additional storage, or dedicated graphics - Click to enlarge

Alternatively, Intel will also offer a version of the 16-core chip with a beefier integrated graphics tile - Click to enlarge

Step up to the 16-core part, and things get a bit more interesting. In addition to the P and LP-E cores, the part picks up another eight E-cores which are connected directly to the chip's L3 cache.

Intel has also upgraded Thread Director, an application designed to optimize workload provisioning across the various core types to function a bit like a waterfall. As the LP-E cores reach capacity limits, they're pushed into the more powerful E-cores and then onto the P-cores as needed.

The idea here is that, for power constrained environments – a thin and light notebook, for example – it may be preferable to keep processes on the more efficient cores and only fire up the beefier P-cores when absolutely necessary.

Moving on to graphics, there are two paths. The first is a larger, more powerful GPU tile from TSMC that triples the number of Xe graphics cores to 12. According to Intel, the GPU is up to 50 percent faster than the iGPUs found on its last-gen Arrow Lake and Lunar Lake parts.

Alternatively, you can keep the quad-core GPU tile and opt instead for an upgraded platform control tile with eight additional lanes of PCIe 5.0 connectivity. This additional I/O should open the door to some interesting system designs dedicated to Arc, Nvidia, or AMD graphics, or even external GPU docks.

Here's a high-level overview of the various package configuration for Panther Lake - Click to enlarge

Can't quite TOPS the competition 

Despite the improved CPU and GPU processing, one area that Panther Lake falls short of the competition is in AI PC performance.

Panther Lake's all-new neural processing unit (NPU) manages to achieve 50 TOPS of AI compute – that's exactly two more than Lunar Lake, and far fewer than the competition. AMD's Strix Halo, which is already shipping, boasts 60 TOPS, while Qualcomm's upcoming X2 Elite will feature a 75 TOPS NPU.

With that said, all three exceed Microsoft's 40 TOPS Copilot+ PC requirement, and when it comes to AI/ML workloads, software can have a bigger impact on performance than what the spec sheet might suggest.

Speaking of TOPS, between the CPU, GPU, and NPU, Intel boasts its latest chip will deliver 180 platform TOPS. As great as that might sound, you can't harness all of that performance at the same time, at least not yet.

Arguably just as important to AI workloads – and graphics-heavy applications like gaming – are memory capacity and bandwidth. Panther Lake supports up to 128 GB of DDR5 memory at up to 7,200 MT/s or 96 GB of LPDDR5 at 9,600 MT/s, which works out to between 112.5 and 150 GB/s.

Memory bandwidth has major implications for graphics-heavy workloads like gaming as well as local AI inference on large language models, like Meta's Llama 3.1 8B. This is one of the reasons why we see GDDR or HBM used on dedicated GPUs rather than slower DRAM.

While the memory specs represent a decent jump, particularly in capacity, over Lunar and Arrow Lake, they make it one of the slowest upcoming mobile processors in terms of bandwidth. Qualcomm's top-specced X2 Elite Extreme parts manage 228 GB/s, while AMD's Strix Halo processors manage 256 GB/s. Meanwhile, Apple's notebook processors can be specced for up to 546 GB/s of memory bandwidth.

Unanswered questions 

Speaking of the competition, Intel isn't ready to compare its latest AI PC parts against the best AMD, Apple, and Qualcomm have to offer just yet. As such, it's hard to say just how big a difference maker Intel's new process tech and CPU/GPU architectures really are.

In fact, Chipzilla hasn't shared full details on things like pricing, clock speeds, or power consumption either. We're told to expect parts ranging from 15 to 45 watts, which suggests we may see a wider range of SKUs than the three main configs disclosed during the Intel Tech Tour.

With so much riding on the chip, we can understand why Intel might be hesitant to share more until we get closer to CES. Intel hasn't had a great track record when it comes to executing on its promises, a bad habit that newly minted CEO Lip Bu Tan has made it his mission to squash.

Getting Panther Lake right is especially important because, in addition to being its flagship mobile processor, it's a showcase for what Intel's Foundry division is capable of. Promise too much too soon and Intel could end up scaring away potential Foundry customers.

We look forward to learning more about Panther Lake, and how it stacks up to the competition, as we get closer to CES. ®