New generation of desktop processors with the Intel LGA 1851 platform
Intel unveiled the Core Ultra 200S series CPUs yesterday. Intel’s long-awaited new desktop processor platform is coming with the LGA 1851 socket, replacing LGA 1700 and the problem-plagued Raptor Lake CPUs. And this new platform is one of the bigger changes in the company’s history. In addition to improved connectivity, it features chiplet tiles with 3D packaging for the first time. And for the first time, Intel’s CPU is fabbed at TSMC.
Core Ultra 200S (where S stands for socket, so it is a designation of the desktop series) are processors that you have known mainly under the Arrow Lake designation until now. It’s a sibling processor family that coexists with the recently released Core Ultra 200V “Lunar Lake”. The processors utilize the same Lion Cove (large P-Cores) and Skymont (E-Cores) cores as Lunar Lake, but have a different chiplet (tile) structure that is more or less based on the previous Core Ultra 100 “Meteor Lake”. That procesor generation did not have a desktop version, however.
Arrow Lake: bringing Core Ultra concepts to desktop
Arrow Lake brings up to 8 P-Cores and 16 E-Cores, for a total of 24 threads (as HT is no longer used). These cores are placed in a 3nm CPU chiplet (Compute Tile), which Intel manufactures using the 3nm N3B process node at TSMC, not in its own fabs. This gives Intel an advantage over both AMD and Qualcomm, whose CPUs in this generation still use a 4nm node, but it makes Arrow Lake quite expensive to produce.
The P-Cores and E-Cores are interconnected in the CPU chiplet by a ring bus, as in the existing desktop processors. Each P-Core and each four-core Skymont cluster has its own 3MB L3 cache block, for a total of 36MB of L3 cache. The L2 cache is 3MB for each P-Core (Lion Cove), while the Skymont cores are always combined into a four-core cluster that shares a 4MB L2 cache. Note that Lion Cove in Arrow Lake has a slightly larger L2 cache than the version of this core implemented in the mobile Lunar Lake (which has only 2.5MB of L2 cache).
Both CPU core architectures are new and we have devoted separate articles to their analysis, to which we will refer you here. Intel is now claiming that the P-Core is expected to have 9% higher performance per 1 MHz of clock speed (the so-called IPC) than the P-Core in Raptor Lake processors.
However, the E-Cores are supposed to have a 32% higher IPC compared to previous-generation E-Core (Gracemont). So the biggest performance upgrade has been made to those, which means Arrow Lake is poised to improve in multi-threaded performance, despite the negative impact of removing HT. Things are kind of upside down compared to older expectations. However this multi-thread performance potential of course requires the given program to work well with the hybrid architecture to be realised.
- Read more: Intel’s new P-Core: Lion Cove is the biggest change since Nehalem
- Read more: Skymont architecture analysed: Intel little core outgrows the big?
SoC tile and NPU
The chiplet with the CPU cores is itself connected via a die-to-die interconnect to other chiplets, notably the SoC Tile, which should be a 6nm die that includes some connectivity, chipset functionality, and a DDR5-6400 memory controller (128 bits wide). Unlike Lunar Lake, SLC (system cache) does not seem to be present.
The SoC chip also includes, for the first time in Intel desktop processors, an integrated NPU, i.e. a dedicated unit for AI acceleration. However, this NPU taken from Meteor Lake processors and uses the third generation architecture (with two NCE blocks), not the new architecture from Lunar Lake. Its performance is 13 TOPS, so also similar to Meteor Lake (or Ryzen 8000G “Hawk Point”), which is not enough to officially support Copilot+ PC features. This chiplet should be manufactured by TSMC’s 6nm process node.
In addition to the SoC Tile, part of the connectivity is implemented in a separate chiplet (IO Extender Tile), which is mainly due to the need to place the output contacts around the processor, so it is more of an implementation detail than an actual innovation.
Connectivity from the CPU and Z890 chipset on LGA 1851
Arrow Lake provides 20 PCIe 5.0 lanes from the CPU (×16 for GPU, ×4 for SSDs) and another PCIe 4.0 ×4 interface for a second SSD. The processors are also supposed to include a controller for Thunderbolt 4, which provides up to two 40Gbps ports.
When installed in a board with the Z890 chipset, the chipset (PCH) delivers an additional 24 PCIe 4.0 lanes, up to 10 USB 10Gbps (USB 3.2 Gen 2) ports or optionally five USB 20Gbps (3.2 Gen 2×2) ports, and up to 14× USB 2.0 ports. The boards can provide up to eight SATA ports from the chipset. The chipset connects to the processor in the same way as Z790 and Z690, via DMI4 ×8 (which has bandwidth equivalent to PCIe 4.0 ×8).
The chipset can provide the digital part of the WiFi 6E / BT 5.3 wireless adapter (you need to add the module with the analog part) and gigabit Ethernet. If there is need for WiFi 7 / BT 5.4 and 2.5Gb/s Ethernet (and also for Thunderbolt 5), dedicated additional adapters have to be added.
(Not)Arc integrated graphics
Arrow Lake processors include an integrated GPU with up to 512 shaders (4 Xe Cores), which is also architecturally adopted from Meteor Lake processors and thus uses the Xe LPG architecture. The architecture is derived from the standalone Arc Alchemist graphics cards, it just doesn’t have the XMX AI units. However, there is full ray tracing support (four RTUs) and the GPU also has generous 4MB L2 cache. The GPU is also on a dedicated chiplet, which is expected to be 5nm (again, manufactured at TSMC).
The GPU in desktop Arrow Lake is roughly half the configuration of the standalone ACM-G11 GPU in graphics cards like the Arc A380, except for the use of shared memory. The GPU only gets the generic “Intel Graphics” designation, as “Arc” is apparently reserved for more powerful models. Output from the integrated GPU is supported over HDMI 2.1 and DisplayPort 2.1 UHBR 20 (i.e. the version with maximum bandwidth, see this article). But you can also pull video output out of the processor in the form of Thunderbolt 4 (via USB-C, up to two ports).
Models
In this first wave of Core Ultra 200S (Arrow Lake) processors, there will only be 125W enthusiast models that have unlocked multipliers for overclocking. There are five of these models – two Core Ultra 5s and two Core Ultra 7s, since there is both a base K version and a KF version SKU without integrated GPUs available for these configurations. However, the top-of-the-line Core Ultra 9 model no longer has the cheaper KF variant available (this is a change from previous generations).
The most powerful Arrow Lake is the Core Ultra 9 285K. The processor with 24 threads and 36 MB L3 cache has eight P-Cores clocked at 3.7 GHz base. The maximum normal boost is 5.5 GHz, with Turbo Boost Max 3.0 (on preferred cores only) you can go up to 5.6 GHz and with active Thermal Velocity Boost you can then go up to 5.7 GHz (as long as the temperature is low enough). This 5.7 GHz is currently the highest clock speed available with the Lion Cove architecture and Arrow Lake processors.
The processor has 16 (Skymont) E-Cores with a base clock speed of 3.2 GHz and a maximum boost of 4.6 GHz. The GPU has 512 shaders and a clock speed of 300–2000 MHz. The base power consumption (TDP) is 125 W, but the maximum turbo power, which tells what the real power consumption will be in multi-threaded applications, is 250 W.
Intel has set a price of 589 USD for this model. In our part of the world it now comes out to about 652 EUR with VAT.
The Core Ultra 265K / Core Ultra 265KF models have 20 threads, consisting of eight P-Cores and 12 E-Cores, i.e. as in the Core i7-14700K. Their L3 cache is reduced to 30 MB (unless it is a mistake and it should be 33 MB, which would be more appropriate). The P-Cores no longer have Thermal Velocity Boost, the base clock speed is 3.9 GHz and the normal boost is 5.4 GHz, with Turbo Boost Max 3.0 the full maximum is 5.5 GHz.
The twelve E-Cores have a clock speed of 3.3–4.6 GHz. The integrated GPU also retains all 512 shaders and the clock speed is also 300–2000 MHz on the 265K. The 265KF model is without the GPU, but it is unknown whether the GPU chiplet is physically absent or just deactivated. The TDP is again 125 W and the maximum turbo power consumption is also 250 W.
The prices are 394 USD for the 265K and 379 USD for the 265KF. For us, this works out to around 436 EUR and 420 EUR VAT-inclusive prices.
The cheapest option on the LGA 1851 platform for now will be the Core Ultra 5 245K and the Core Ultra 5 245KF without a GPU. These models have 14 threads and 24MB of L3 cache – six P-Cores and eight E-Cores. The clock speeds of the P-Cores are 4.2 GHz base and 5.2 GHz in maximum boost, and for the E-Cores 3.6 base, 4.6 GHz in boost.
The GPU in the 245K model still has 512 shaders, but with the clock speed reduced to 300-1900 MHz. The KF model is again without a GPU. Intel maintains a 125W TDP for this model, but the maximum turbo power consumption is a lower 159W.
The prices are 309 USD for the Core Ultra 5 245K and 294 USD for the 245KF. For us it comes out to about 342 EUR and 326 EUR (including 21% VAT).
The prices of Arrow Lake are virtually the same as in the previous generation, but the difference between the K and KF models has narrowed, making the 245K and 265K models some 10–15 USD cheaper than the 14600K and 14700K prior models. However, the highest-end and cheapest models have the same suggested prices as before.
The new proposition is power efficiency instead of performance
As the leaked slides a few days beforehand indicated, Intel has failed to increase gaming performance on Arrow Lake / Core Ultra 200S above the level offered by the previous Raptor Lake (Core 14th) generation. According to unofficial information, this could be not only due to the regression of P-Core clock speeds, but also apparently due to the negative impact of the SoC chiplet and uncore taken from Meteor Lake processors, which is characterized by high RAM latency (caused by the chiplet concept, but also significantly lower ring bus clock speed is allegedly also part of the problem). High RAM latency is something that hurts gaming the most.
Because of this, the company is presenting Arrow Lake as a CPU that supposedly has the same performance as Raptor Lake, providing significantly reduced power consumption, supposedly by nearly a half. When running single-threaded Cinebench 2024, the power consumption is reportedly 40% lower according to the package power sensor (this is from comparison between Core Ultra 9 285k and Core i9-14900K).
Elsewhere, Intel talks about a 42-58% reduction in power consumption, but this may be the case when tasks are running on the little core (likely the case of the the -58% result, which is measured during a Zoom video call). Thread Director Technology in Windows 11 should automatically run tasks on E-Cores (Skymont) by default and only move them to the big cores when it detects they need maximum performance, on Arrow Lake.
This reduction in power consumption is also expected to result in lower temperatures (reportedly 10°C lower in games). However, this is something where it will be better to wait for independent test results. The alleged power consumption reduction is supposed to be true for games and in light loads, as the power limits are still high under full load (250 W), basically the same as in Raptor Lake.
For multi-threaded workloads, Intel promises a performance boost, unlike what is said about gaming. Arrow Lake is supposed to have up to 15% better multithreaded performance vs Arrow Lake (in SPECint 2017), elsewhere in the presentation Intel mentions increases of up to 20%. It’s questionable whether that’s a good result given how it took a significantly upgraded architecture, expensive chiplet design, as well as a transition from 7nm technology to a state-of-the-art 3nm process node (a jump of two full new generations) to get the 20 %.
The reduction of the currently high power consumption in games is welcome in any case (it should be said that the same is already offered by AMD processors with 3D V-Cache, which will also offer better performance, although their idle power consumption may be higher). Power consumption was something that was almost completely disregarded or downplayed in previous generations, and Intel, on the contrary, chased the highest performance, even at the cost of ignoring its own power limit specifications.
So it’s likely that the emphasis on power efficiency here in the Arrow Lake generation is a bit of a Plan B Intel opted for solely because the company can’t advertise higher gaming and single-threaded performance results, and the turn to power efficiency is probably not an entirely voluntary change of heart. If Intel has the ability to increase performance in future generations at the cost of degrading power consumptionagain, we may still see a return to the previous mode of operation (though we hope this doesn’t happen).
Release in two weeks
This unveiling is only formal for now. But the real launch with physical availability of the processors in stores is also not far away, it will take place in two weeks on October 24 (also on Thursday). All five revealed models and the Z890 chipset boards for them will be released on this date. The time the reviews will be published is not quite clear yet, it is possible that the embargo on them will be lifted on the same day and time.
English translation and edit by Jozef Dudáš
⠀