Intel Core Ultra 200S Desktop CPUs Unveiled: Arrow Lake Promises Big Efficiency Gains

Back in late 2021, Intel's 12th generation of desktop CPUs, codenamed Alder Lake, was a revelation. After years of mostly spinning its wheels and ceding performance crowns to rival AMD, Chipzilla was finally back on top. And aside from some modest refinements like faster memory support and expanding the number of low-power E-cores, that's mostly where the company's desktop CPUs stayed, but we think the market is ready to move on. And now, so is Intel. Today, Intel's enter Arrow Lake processor enters the arena. 

This is where we say goodbye to Intel Core CPUs, and wave hello to Intel Core Ultra 200S, the desktop lineup that brings many of the architectural improvements of recent Intel mobile CPUs to desktops around the world. Intel is promising some impressive gains on both CPU performance and power consumption. We're going to break all of that down first here, and then talk about the specific SKUs that will be available for purchase. 

Intel Arrow Lake Core Ultra 200S Tiles And Packaging

Meteor Lake introduced "tiles", Intel's design for breaking a larger chip up into functional units. These are discrete pieces of silicon assembled together using a high-bandwidth interconnect on a package substrate. Those CPUs never shipped in desktop systems, though, so Arrow Lake is where that tile concept makes its way out of notebooks and into high-powered machines. The architecture more closely mirrors Lunar Lake, though, and from both an organizational and a performance standpoint, it makes a lot of sense. 

The tiles in Arrow Lake are broken up by purpose. The Compute Tile contains all of the CPU cores of both the E- (efficient) and P- (performance) varieties. Intel Core Ultra desktop and performance laptop CPUs also have a GPU tile, which is stripped down a bit from Lunar Lake, since these PCs are expected to have a discrete GPU.


The SoC tile, which previously had some low-power E-cores in Meteor Lake, is still present. However, it focuses on being more of a utility player for video processing and other light duty tasks outside of the GPU. This tile holds the display output logic for up to four 4K 60 Hz displays and video encode/decode blocks. Supported video codecs for both encoding and decoding include 8k 10-bit HDR encoding for AVC, HEVC, VP9, and AV1. 


Core Ultra CPUs' tile collections are banded together with that speedy interconnect and built with 3D Foveros packaging, which the company initially employed for Meteor Lake as well. This is a chip-stacking technology that allows different kinds of logic to be more efficiently configured for maximum silicon real estate management and deeper integration. 

Intel Arrow Lake P-Cores: Lion Cove Roars on Desktop

Intel had several goals in mind that drove the design of its Lion Cove architecture. First and foremost, Lion Cove was designed with performance-per-watt and performance-per-area (per square millimeter) for top-of-the-line single-threaded performance. As a result, the architecture was overhauled in search of improved IPC and scalability down the road. 

Just like the company's Lunar Lake mobile CPUs, Intel's Lion Cove architecture powers the P-cores of the Core Ultra 200 series. Pretty much everything from our previous coverage rings true today. Gone is Intel's symmetric multi-threading (SMT) technology known as Hyper Threading. Intel feels it is better to recover the silicon used to implement the feature (which includes scheduling and other support logic on top of SMT itself) and instead focus on the best single-threaded performance it could muster. 


The resource allocation hasn't really changed since Lunar Lake hit notebooks, either. We still have the same 18 execution ports (i.e., functional units), the same "8x wider" branch prediction unit, and a wider scheduler compared to Redwood Cove. That particular architecture was never available in desktops, but was part of Meteor Lake which shipped in laptops with TDP specs that ranged from less than 10 up to 28 Watts. However, it's also fair to say Lion Cove is considerably more parallel than Raptor Cove. That should result in some single-threaded performance improvements. 


Intel has made it a habit in recent years to give CPUs a lot of cache, but Arrow Lake steps that up a bit. First of all, compared to the 14th generation CPUs, the size of the L3 shared Smart Cache remains at up to 36 MB, depending on core counts. That's also true for the 4MB of L2 cache per E-core cluster, too. However, each P-core gets 50% more L2 cache, up to 3 MB from the previous 2 MB, so a CPU with eight P-cores would go from 16 MB of L2 to 24 MB of L2 for its P-core group. 


Overall, Intel says that the Lion Cove P-Core is around 9% faster than Raptor Cove cores found in 14th generation desktop CPUs. That's just a clock-for-clock comparison, but if nominal clock speeds are any higher, that advantage will compound. 

Intel Arrow Lake E-Cores: Skymont Efficiency

In our previous coverage of Lunar Lake, we pointed out that the purpose of the E-core has changed. Intel stripped out Hyper Threading, so the scheduler had no choice but to change, but while Intel was doing that, the design team decided that it would be best to schedule everything on the E-cores initially, since they draw less power and produce less heat. That made the E-cores a heavy primary focus for the scheduler, and the E-cores had to be fast. And in Skymont, that's exactly what they are. 


We already know that Skymont has a "wider" design than Meteor Lake's Crestmont cores, but that's even more true when compared to Gracemont from the Alder Lake and Raptor Lake timeframe. Compared to Gracemont, Skymont has deeper instruction queueing for better parallelism, an increased 4 MB of shared L2 cache (up from 2.5 MB), and double the cache bandwidth. Intel calls Skymont its most efficient performant architecture. Rather than being energy efficient and sacrificing performance, Skymont is strong enough to be the primary processing unit for work that doesn't need a ton of sustain behind it. That keeps the P-cores in an idle state and brings down system power requirements. 


All of that adds up to some big desktop gains over Intel's 14th generation CPUs' E-cores. On balance, the company says we should expect around 32% increased IPC in single-threaded INT performance. That 32% also holds for multi-threaded INT performance, too. That's a whole lot faster per clock, without accounting for the differences in clock speed. 


Floating-point performance of Arrow Lake is improved even further, and it's up to 72% faster than Gracemont in single-threaded workloads. That actually dips a bit in multi-threaded floating point performance to around 55%. Again, these are IPC gains at the same frequency, not the sum total of what to expect. Skymont took E-cores from background players to center-stage performers. 

Next up we'll cover changes to threading and power management, and then the specific CPU models that will be available.