In October 2024, Intel introduced Arrow Lake desktop processors (Core Ultra 200S series), which were supposed to be a significant step forward in both architecture and performance. However, in practice, everything did not go as the company expected. Immediately after the release, Arrow Lake came under a barrage of criticism, primarily for the unexpectedly low results of the new processors in games. They turned out to be lower not only than competitors from the AMD camp, but also than their immediate predecessors – the Raptor Lake series processors.
Intel was quick to acknowledge the problems, stating that the launch “didn’t go as planned” and promised to quickly fix the shortcomings with BIOS updates and improved interaction with the Windows scheduler. However, in reality, the effect of all the subsequent fixes was minimal. The achieved performance gains were usually within the margin of error, and in some cases, Arrow Lake’s gaming performance even worsened. Everything clearly indicated that the root of the problems should be sought deeper – not at the software level, but in the internal structure of the new processors.
The key feature of Arrow Lake, which fundamentally distinguishes them from their predecessors, is the disaggregated architecture. This means that the processor now consists of several semiconductor crystals (tiles) joined into a single whole, across which the functional units of the final CPU are distributed. This approach does reduce production costs, but, as practice has shown, it strikes at important consumer qualities. In the context of Arrow Lake, the problem was that the computing cores and the processor SoC, together with the DDR5 controller, ended up in different tiles, which led to a significant increase in overhead costs when accessing memory. Due to the lengthening and complexity of the path between the cores and the memory controller, the practical latency when working with regular dual-channel DDR5 increased by more than a third: if the monolithic Raptor Lake with DDR5-6400 and default settings shows a practical delay of 60-65 ns, then in the disaggregated Arrow Lake this figure reaches 85-90 ns. It is not surprising that such a deterioration of one of the main characteristics of the memory subsystem has dealt a critical blow to gaming performance.
Experimenting with various DDR5 modules in systems built on Raptor Lake processors, we saw how much memory speed affects gaming performance. Overclocking memory and adjusting latencies in the previous-generation platform translates into a significant FPS increase – no joke, installing high-quality DDR5-7600/8000 modules in the system could achieve a frame rate increase of more than ten percent. Something similar would not hurt in the case of Arrow Lake. Unlike BIOS updates, this could really “save” its gaming performance. However, in this case, efforts should obviously be aimed primarily at reducing the dramatically increased latency.
Moreover, Arrow Lake has more than ok memory bandwidth. Compared to previous generations of processors, it has not dropped, and in addition, Intel has added support for DDR5 CUDIMM modules in the new CPUs, thanks to which the memory frequency can now be raised up to DDR5-9600, which entails a proportional increase in bandwidth. But simply overclocking the memory will not solve the problem with gaming performance. As we have already seen in the tests conducted earlier, the gaming FPS responds rather weakly to an increase in the frequency of DDR5 modules, if high latency remains.
In other words, in order to try to normalize the operation of Arrow Lake in games, you need not a straightforward memory overclocking, but more cunning and complex actions. And we know what – in this article we will tell you how to configure the memory in a system based on Core Ultra 9 285K so as to get at least a 15% increase in frame rate in games, and make this processor outperform its predecessor in games. But in this case, we cannot do without good memory. Fortunately, the test lab had a set of Adata XPG Lancer CUDIMM RGB DDR5-8800 48GB modules, and with it the desired effect (and sometimes even stronger) was quite easy to achieve.
⇡#Adata XPG Lancer CUDIMM RGB DDR5-8800 48GB – what kind of memory is this?
Before we move on to discussing how to speed up the Core Ultra 9 285K in games, let’s take a closer look at the Adata XPG Lancer CUDIMM RGB DDR5-8800 48GB module kit, which we will use to conduct all the tests. The main feature of this kit is that we are talking about DDR5 CUDIMM, which means that the modules included in it have an additional element – a CKD clock signal driver. This chip is responsible for regenerating the clock signal coming from the processor inside each module and guarantees stable memory operation at higher frequencies than in the case of regular unbuffered UDIMM modules. A detailed article about how CUDIMM works and what it gives was published on our website a little earlier.
As for the Adata XPG Lancer CUDIMM RGB DDR5-8800 48GB kit itself, the presence of a CKD chip in the modules makes it possible to operate it in DDR5-8800 mode. With regular UDIMM modules, such a frequency would not be available even with Arrow Lake processors, whose memory controller currently offers the best DDR5 overclocking capabilities. Otherwise, the XPG Lancer CUDIMM DDR5 modules are similar to the well-known XPG Lancer RGB DDR5 strips, including in appearance.
The Adata XPG Lancer CUDIMM RGB DDR5-8800 48GB kit in question (article AX5CU8800C4224G-DCLACRSG) consists of a pair of single-rank modules with a capacity of 24 GB each, designed to operate at a frequency of 8800 MHz with a timing of CL42 and a voltage of 1.45 V. Its full passport specifications are as follows:
- A set of two DDR5 CUDIMM modules, 24GB each;
- Operating mode DDR5-8800;
- Timing 42-54-54-134;
- Voltage 1.45 V;
- Support for Intel XMP 3.0 profiles;
- High heat dissipators made of unpainted aluminum;
- LED RGB backlight;
- Lifetime warranty.
The XPG Lancer CUDIMM memory kit family also includes two more DDR5 CUDIMM variants, targeting 8400 and 9200 MHz. But the DDR5-8800 variant we chose is the best one, since despite its high frequency, it is guaranteed to work with Arrow Lake processors in Gear 2 memory controller mode, i.e. with a 2:1 memory bus and controller frequency ratio.
For ease of use, the Adata XPG Lancer CUDIMM RGB DDR5-8800 48GB kit supports XMP 3.0. The only available profile contains settings similar to the passport ones — with it, you can launch the memory without lengthy manual adjustments. The timings in the profile fully comply with the specification. Plus, it additionally declares the tRFC2 and tWR timings — their values are set to 967 and 132, respectively.
A few words should be said about the external design of the modules. It actually repeats the exterior of the XPG Lancer RGB DDR5 memory: the XPG Lancer CUDIMM RGB DDR5-8800 48GB modules use the same massive 39 mm high heat dissipators with a triangular cutout in the center filled with a plastic insert. This insert is illuminated from the inside by RGB LEDs.
At the same time, the CUDIMM modules in question differ slightly in the design of the heat spreaders. This time, they received a smooth surface without any relief, and Adata also decided to leave the unpainted aluminum without additional coating. This allows the XPG Lancer CUDIMM DDR5 modules to fit organically into both white and black assemblies.
We should also note the ECO mark that appeared on the modules. In this way, Adata decided to mark the eco-friendliness of its memory, which consists in the fact that recycled materials are used in the production of heat dissipators – they are 80% recycled aluminum.
The height of the modules together with the radiators is 43.5 mm – this is slightly above average, and in some configurations there may be memory conflicts with massive coolers.
If you have already read reviews of CUDIMM modules on our website, then the filling of the XPG Lancer CUDIMM DDR5-8800 is unlikely to surprise you. It uses the usual 3-Gbit SK Hynix M-die chips, since there is no other element base for such high-speed memory in nature yet.
There are 8 of them installed in each module, all of them are located on one side of the 10-layer printed circuit board.
Also on the board in the center you can see a Richtek voltage converter and a CKD frequency generator chip by Rambus, typical for CUDIMM modules.
⇡#What kind of performance boost does DDR5-8800 give?
The LGA1851 platform is the only one that supports CUDIMM modules, and only Arrow Lake processors are capable of working with memory whose frequency has gone well beyond 8000 MHz. However, is there any real point in using such high-speed DDR5 modules? The answer to this question is far from obvious. The fact is that increasing the memory frequency leads to an increase in bandwidth, which for the same dual-channel DDR5-8800 reaches a gigantic 140 GB/s, but has a rather weak effect on latencies.
For example, the main CAS Latency timing of the Adata XPG Lancer CUDIMM DDR5-8800 modules under consideration is 42 cycles, which is equivalent to an open-line read latency of 9.5 ns. At the same time, the widely used DDR5-6400 modules with CL32 have a latency of 10 ns in terms of conversion, that is, in fact, they are almost no worse in terms of latency if measured not in cycles, but in nanoseconds. Therefore, to understand whether the performance of systems based on Arrow Lake processors can be improved by increasing memory bandwidth alone, it is necessary to turn to real measurements.
When selecting the XMP profile settings for the XPG Lancer CUDIMM RGB DDR5-8800 48GB module kit in question, the following timing scheme is activated on a platform based on the Core Ultra 9 285K processor.
Note that for some reason the motherboard sets a slightly higher CL timing when activating the XMP profile, and for the sake of the experiment’s purity it must be manually corrected to the passport value of 42. After that, the memory subsystem produces the following performance indicators.
Core Ultra 9 285K, DDR5-8800, XMP
The practical memory bandwidth when reading reaches 120 GB/s, which is 86% of the peak theoretical value. Thus, the absence of bottlenecks on the highway between Arrow Lake and memory is confirmed, including with a significant increase in the frequency of DDR5 modules above the nominal values. The reading and copying indicators are slightly lower, although they are also not at all disappointing. But the latency at the level of 85 ns does not look too encouraging.
For illustration purposes, the obtained values can be compared with the results of the Aida64 Cache & Memory Benchmark test in a similar system when installing a typical DDR5-6400 kit with timings of 32-39-39-102.
Core Ultra 9 285K, DDR5-6400, XMP
It is this DDR5-6400 memory that we will consider the starting point for this study. And in comparison to it, overclocker DDR5-8800 modules using default settings do not seem to be a very interesting alternative. Yes, the practical bandwidth increases by more than 20% when increasing the memory frequency from 6400 to 8800 MHz, but the latency changes quite slightly – within 5%.
However, judging real performance based on a synthetic memory subsystem test is not entirely correct, so let’s take a quick look at how performance in games (at 1080p resolution) changes when moving from DDR5-6400 to DDR5-8800. For clarity, we also added the results of the intermediate memory option to this comparison – DDR5-8000 with timings of 40-48-48-128.
Fast memory does provide a frame rate boost, but it looks rather unconvincing. Replacing regular DDR5 memory with overclocker CUDIMM modules results in a 2-4% FPS increase, which is a bit low considering the almost two-fold difference in price. In this case, DDR5-8000 modules seem to be a more profitable purchase, which are not so expensive, but are still capable of making a comparable contribution to increasing the performance of Arrow Lake processors.
⇡#What Happens If You Overclock DDR5-8800 More, and Why It’s Unnecessary
The main advantage of CUDIMM modules is that they have their own clock generator, which ensures stability at high frequencies that go beyond the capabilities of conventional unbuffered memory. Adata XPG Lancer CUDIMM DDR5-8800 modules perfectly confirm this thesis. They not only work great at a frequency of 8800 MHz, which is unattainable for conventional memory, but they can also be easily overclocked even higher. For example, we managed to achieve their problem-free operation in the DDR5-9200 state with a timing scheme of 42-56-56-140.
But there is a nuance: memory overclocking above 8800 MHz in the LGA1851 platform depends on the “silicon lottery” – the quality of the memory controller in a specific processor. If the DDR5-8800 mode will most likely be taken by any Arrow Lake, then their stable operation with DDR5-9000 and further is not guaranteed and may involve the need to make some compromises.
Thus, the Core Ultra 9 285K at our disposal was able to handle DDR5-9000 and DDR5-9200 only when switching the memory controller from the standard Gear 2 mode to Gear 4 mode, which means a twofold reduction in its frequency. In Gear 2 mode, the frequency of the memory and controller is related as 2:1, and in Gear 4 this ratio changes to 4:1. On the one hand, this removes the load from the controller and allows the memory to be clocked at a higher frequency, but on the other hand, it introduces additional delays when working with memory.
Moreover, such delays hit the performance of the entire memory subsystem quite painfully, which is easy to see from the Aida64 Cache & Memory Benchmark results. Below are the results of this test with XPG Lancer CUDIMM DDR5-8800 modules overclocked to DDR5-9200 and the memory controller in Gear 4 mode (i.e. operating at 1150 MHz).
Core Ultra 9 285K, DDR5-9200, Gear 4
Compared to the results obtained with DDR5-8800 in Gear 2 mode (with a memory controller at 2200 MHz), the performance degradation is very noticeable. Due to the decrease in the controller frequency, the throughput also drops, but the practical latency suffers especially badly, rolling back to a completely indecent 100 ns. The higher operating frequency of the modules does not compensate for this deterioration in performance at all.
How this affects gaming performance can be seen below, although it is obvious that nothing good can be expected from such a mode. Moreover, the real results are even worse than expected: the FPS drop in a system with DDR5-9200 compared to DDR5-8800 is about 5%. This means that DDR5-9200 with a memory controller in Gear 4 mode slows down the Core Ultra 9 285K even below the level provided by regular DDR5-6400.
Thus, using the Gear 4 mode to achieve higher memory frequencies is generally pointless. And this makes DDR5-8800 the best memory option for Arrow Lake processors – overclocking the memory further to increase performance will be pointless in most cases. The Gear 2 mode, which determines the most favorable ratio between the memory and controller frequencies of 2:1, is guaranteed to work in Arrow Lake processors only when overclocking the memory to a frequency of 8800 MHz. To conquer higher frequencies without a drop in performance caused by the need to switch the memory controller mode, you need to have a successful or specially selected processor.
⇡#How and Why to Speed Up Intra-Processor Buses
In our Core Ultra 9 285K review, we detailed how the memory controller in Arrow Lake ended up with such high latencies. In short, the problem is that the data path from the processor cores to the memory has become significantly longer, and this is due to the new disaggregated CPU architecture, due to which the memory controller and the cores ended up on different semiconductor dies.
In previous generations of monolithic processors, the cores had the ability to access the memory controller directly, via a ring bus that unites all the CPU components into a single whole. But in Arrow Lake, the Ring Bus, located in the computing tile, does not have a direct output to the memory controller: all requests must be transferred to the neighboring SoC crystal, for the physical connection between which a completely different bus is responsible – the D2D (Die-to-Die) interconnect. However, this interconnect does not directly connect to the memory controller: it is connected to the SoC’s own internal bus, the NoC (Network-on-Chip), which, in turn, unites all the extra-core elements of Arrow Lake.
Arrow Lake Tile Structure Diagram
Thus, when the processor accesses memory in the case of Arrow Lake, the data has to go through a very long chain “core – Ring – D2D – NoC – controller – DDR5”. And although all sections in the middle of this highway have a gigantic bandwidth exceeding 0.5 TB / s, problems arise at the junctions between them. All intermediate buses – Ring, D2D and NoC – work asynchronously, using their own frequencies, and coordinating data transactions at each stage adds additional delays. This overhead is precisely what manifests itself in the form of increased latency of the Arrow Lake memory subsystem, which is 20-30 ns higher than the memory latency in systems based on Raptor Lake processors.
Unfortunately, this problem cannot be fixed, it is embedded in Arrow Lake at a fundamental level. However, it is possible to try to reduce its negative impact – for this, it is enough to increase the frequencies of all intermediate buses, which will entail some reduction in the coordination delays. Fortunately, in overclocker modifications of Arrow Lake (K series) there is access to multipliers for the frequencies of all necessary buses.
The Arrow Lake ring bus has a nominal frequency of 3.8 GHz. It overclocks rather reluctantly, and, moreover, its frequency should not exceed the maximum frequency of the P- and E-cores. As a result, unless you increase its voltage, typical overclocking will be limited to a frequency of 4.0-4.1 GHz. For example, our Core Ultra 9 285K sample was able to survive only an increase in the ring bus frequency to 4.0 GHz without losing stability.
But overclocking the D2D bus is a much more effective process. Without interfering with the VnnAON voltage, its frequency can be increased more than one and a half times without any negative effects on stability. By default, the D2D frequency is 2.1 GHz, and it is overclocked to values of about 3.5 GHz. Our processor was even able to withstand a D2D frequency of 3.6 GHz – 70% higher than the nominal.
The NoC bus does not disappoint either. In the BIOS of most motherboards, it is called NGU (Next Generation Uncore), after the name of the main element of the SoC crystal. Its standard frequency is set at 2.6 GHz, but this value can be increased to at least 3.4 GHz (without increasing the VccSA voltage). It is this NoC frequency – 30% higher than the nominal – that we chose for practical testing.
As expected, overclocking the Ring, D2D, and NoC buses does improve the practical performance of the memory subsystem. The increased frequencies, together with the use of 285K DDR5-8800 memory modules in the Core Ultra 9 system, allow us to obtain the following results in Aida64 Cache & Memory Benchmark.
Core Ultra 9 285K, DDR5-8800, XMP, overclocking Ring, D2D and NGU
Overclocking the intermediate buses on the highway from the processor cores to the memory has almost no effect on the practical throughput – everything is fine with it from the start. But the latency really goes down. We managed to reduce it by 7 ns – although not brilliant, but quite a noticeable result. And the most important thing is that it was achieved without increasing any voltages, that is, without risk to the processor silicon and without a significant increase in heat generation and power consumption.
However, real-world gaming tests do not allow us to say that overclocking the internal processor buses can achieve any significant performance improvement. The frame rate increase is about 3-4%.
But relative to the reference point we have chosen in the form of DDR5-6400, fast memory operating at a frequency of 8800 MHz gives a total increase of not 2-4, but 6%. However, that’s not all – the most interesting is yet to come.
⇡#We squeeze the maximum out of timings – there is no way around it
You can fight high memory subsystem latencies not only by increasing various frequencies. There is another way – reducing the timings of the memory itself. Usually, the timing scheme embedded in XMP profiles is far from ideal, and this is exactly the case with the Adata XPG Lancer CUDIMM RGB DDR5-8800 48GB memory kit. Another thing is that setting up optimal latencies is a rather laborious process that requires a certain amount of patience. But in the case of high-speed CUDIMM modules, it plays into their hands that they are all based on the same SK Hynix M-die chips, and therefore, the optimal sets of latencies for different kits with the same frequency are generally similar.
In addition, you can avoid searching for optimal values for all timings and limit yourself to selecting only a few key parameters that affect the performance of the memory subsystem most significantly. These are the base timing CL; primary timings tRCD and tRP; memory refresh time tRFC2 and tRFCsb; and the interval between refreshes tREFI.
However, for the XPG Lancer CUDIMM DDR5-8800 memory modules in question, we carried out a full optimization, as a result of which the timings were reduced to the values shown in the screenshot.
As you can see, the changes look quite significant, and for good reason – due to such fine-tuning, the performance of the memory subsystem really increases noticeably, especially if it is carried out after increasing the frequencies of the intra-processor buses.
Timing optimization affects both throughput and latency. By simply adjusting the memory modules to their rated frequency, we managed to increase the read speed by 13%, the write speed by 30%, and the data copy speed by 26%. The latency also decreased by another 10 ns. Of course, the latency would still be lower in LGA1700 systems, but now the deterioration of this parameter compared to the indicators of previous generations of processors at least does not look catastrophic.
Core Ultra 9 285K, DDR5-8800, customized timings + Ring, D2D and NGU overclocking
But the most interesting thing is that it is the CUDIMM module timings setting that gives the greatest practical effect among all the procedures performed. Even if you do not overclock the intra-processor buses, the increase in gaming performance due to timings alone reaches 9-10%.
Thus, Arrow Lake performance really depends heavily on the memory subsystem parameters. And using high-speed memory, such as the Adata XPG Lancer CUDIMM RGB DDR5-8800 48GB, with these processors is a pretty sound idea. However, relying on default settings and timings from the XMP profile is far from the best option. To fully unleash the potential of high-speed modules, painstaking manual tuning is required.
⇡#Description of the test system and testing methodology
By now, we have confirmed that there are three methods for increasing the memory subsystem performance in the LGA1851 platform: overclocking the memory modules by frequency (but only up to DDR5-8800), overclocking the Ring, D2D and NoC (NGU) buses, and minimizing timings. Each allows for a small increase in performance, but all of them together seem to be able to take the performance of the rather unsuccessful Core Ultra 9 285K processor to a new level.
To test this assumption, we decided to run a more detailed test and answer the question: how much additional performance can you get from an Arrow Lake-based system if you equip it with a quality overclocker memory kit like the Adata XPG Lancer CUDIMM RGB DDR5-8800 48GB and spend time carefully tuning it.
The full list of components we used in these tests is provided below.
- Processor: Intel Core Ultra 9 285K (Arrow Lake, 8P+16E-core, 3.7-5.7/3.2-4.6 GHz, 36 Mbytes L3).
- CPU cooler: custom liquid cooling system made from EKWB components.
- Motherboard: MSI MEG Z890 Unity-X (LGA1851, Intel Z890).
-
- Adata XPG Lancer CUDIMM DDR5 AX5CU8800C4224G-DCLACRSG (2 x 24GB, DDR5-8800 CUDIMM, CL42-54-54-134);
- G.Skill Trident Z5 F5-6400J3239F24GX2-TZ5RK (2 × 24 Гбайт, DDR5-6400 UDIMM, CL32-39-39-102).
- Video card: Palit GeForce RTX 5090 GameRock (2017/2407 MHz, 28 Gbps, 32 GB).
- Disk subsystem: Intel SSD 760p 2 TB (SSDPEKKW020T8X1).
- Power supply: Deepcool PX1200G (80+ Gold, ATX 12V 3.0, 1200 W).
The testing was carried out in the Microsoft Windows 11 Pro (24H2) Build 26100.2605 operating system, which includes all the necessary updates for the correct operation of the schedulers of modern AMD and Intel processors. To further improve performance, we disabled “Virtualization-based security” in the Windows settings and enabled “Hardware-accelerated graphics processor scheduling”. The system used the latest GeForce 572.83 Driver graphics driver.
The performance study compared five memory subsystem configurations:
- DDR5-6400 32-39-39-102 (timings from XMP profile);
- DDR5-8800 42-54-54-134 (timings from XMP profile);
- DDR5-8800 42-54-54-134 (timings from XMP profile) plus overclocking Ring up to 4.0 GHz, D2D up to 3.6 GHz, NGU up to 3.4 GHz;
- DDR5-8800 40-52-52-64 (timings adjusted manually);
- DDR5-8800 40-52-52-64 (manually adjusted timings) plus overclocking Ring to 4.0 GHz, D2D to 3.6 GHz, NGU to 3.4 GHz.
Description of tools used to measure computing performance:
Synthetic benchmarks:
- AIDA64 Engineer 7.20.6800 – Cache and Memory Benchmark memory subsystem test.
- Geekbench 6.3.0 measures single-threaded and multi-threaded CPU performance in common user scenarios, from reading email to image processing.
Tests in applications:
- 7-zip 24.08 – testing compression and decompression speed. A built-in benchmark with a dictionary size of up to 64 MB is used.
- Adobe Photoshop 2024 11.25.0 – testing performance when processing graphic images. The PugetBench for Photoshop 1.0.1 test script is used, simulating basic operations and working with the Camera Raw Filter, Lens Correction, Reduce Noise, Smart Sharpen, Field Blur, Tilt-Shift Blur, Iris Blur, Adaptive Wide Angle, Liquify filters.
- Adobe Premiere Pro 2024 24.5.0 – testing video editing performance. The PugetBench for Premiere Pro 1.1.0 test script is used, which simulates editing 4K videos in different formats, applying various effects to them, and the final rendering for YouTube.
- Blender 4.2.0 – testing the speed of final rendering on the CPU. The standard Blender Benchmark is used.
- Cinebench 2024 is a standard benchmark for evaluating CPU rendering speed in Redshift, the engine used by Maxon’s Cinema 4D suite.
- FastSD CPU-measurement of the speed of quick II generation of images in Stable Diffusion 1.5 in LCM-Lora mode to CPU. The image is created by a resolution of 1024 × 1024 in five iterations.
- Microsoft Visual Studio 2022 (17.13.3)-measuring the compilation time of a large MSVC project —blender version 4.2.0.
Games:
- Assassin’s Creed Mirage. Graphics settings: Graphics Quality = Very High.
- Baldur’s Gate 3. Graphics settings: Vulcan, Overall Preset = Ultra.
- Cyberpunk 2077 2.01. Graphics settings: Quick Preset = RayTracing: Medium.
- Horizon Zero Dawn Remastered. Graphics settings: Preset = Very High, Anti-Aliasing = TAA, Upscale Method = Off.
- Kingdom Come: Deliverance II. Graphics settings: Overall Image Quality = Ultra, Horizontal FOV = 100.
- Marvel’s Spider-Man Remastered. Настройки графики: Preset = Very High, Ray-Traced reflection = On, Reflection Resolution = Very High, Geometry Detail = Very High, Object Range = 10, Anti-Aliasing = TAA.
- Shadow of the Tomb Raider. Настройки графики: DirectX12, Preset = Highest, Anti-Aliasing = TAA, Ray Traced Shadow Quality = Ultra.
- Starfield. Graphics settings: Graphics Preset = Ultra, Upscaling = Off.
In all game tests, the average number of frames per second, as well as 0.01-quantile (first percentile) for FPS values are given as results. The use of 0.01-quantile instead of the minimum FPS is due to the desire to clear the results from random bursts of performance that were provoked by reasons not directly related to the operation of the main components of the platform.
⇡#Application Performance
Until now, we have been looking at the frame rate in games to assess the impact of the memory subsystem speed on overall performance. However, memory can also affect the performance of the Core Ultra 9 285K in resource-intensive applications. Therefore, detailed testing begins with them.
Judging by the Geekbench 6 benchmark, which uses real common algorithms to evaluate performance, the fast DDR5-8800 can provide a 4% increase in multi-threaded performance even without any special settings, simply by increasing the bandwidth. But this result can only be considered a starting point. If you add manual timing adjustment and overclocking of intermediate intra-processor buses to the transition to advanced CUDIMM modules, the total advantage reaches 14% in multi-threaded and 4% in single-threaded load.
This increase is a good illustration of the importance of RAM in Arrow Lake-based systems. Of course, not all applications can achieve such a significant performance boost by overclocking memory, but real-world tests show that adding an additional 8-10% to the average performance of the Core Ultra 9 285K is an achievable goal.
Rendering:
Photo processing:
Work with video:
Compilation:
Archiving:
Neural networks:
It is worth noting one important detail: simply buying and installing overclocking DDR5 modules, like the Adata XPG Lancer CUDIMM DDR5-8800 we use, is not really enough. This is just the beginning of the optimization path. Such modules in the default configuration provide almost no positive effect, and in order to achieve a noticeable increase in performance, you will need to put in some additional effort in setting up. But the final result can exceed the wildest expectations. In some cases, such as when archiving in 7-zip, the performance increase compared to the “base” DDR5-6400 can reach a fantastic 28%. Although there are also opposite examples: for example, during final rendering, the impact of memory on the operating speed is quite limited and does not go beyond 5%.
⇡#Gaming performance
The situation with games is more interesting. The fact is that modern games operate with large amounts of data, and a fast memory subsystem is one of the key factors for achieving good gaming performance. In fact, it is the increased delays when accessing memory that cause the weak results of the Core Ultra 9 285K in games, which is why this processor is criticized by enthusiasts.
For this reason, any actions aimed at improving Arrow Lake’s memory performance have a positive effect on FPS. At the same time, the amount of gain that can be achieved with fast and properly configured DDR5 makes you seriously think about the fact that high-speed modules like the Adata XPG Lancer CUDIMM DDR5-8800 are a very desirable addition to the Core Ultra 9 285K. Indeed, with their help, we were able to increase gaming performance (compared to the “regular” DDR5-6400) by an average of 16%, and in the most memory-sensitive games, the average FPS increase reaches 20%.
However, speaking about such a positive effect from the transition from DDR5-6400 to DDR5-8800 CUDIMM, we cannot help but mention again that the obtained increase is the result of several factors being added together. The Adata XPG Lancer CUDIMM DDR5-8800 modules themselves, when operating in the default mode, do not raise the frame rate in games very convincingly – the average FPS increase is only about 3%. The main advantages of fast and high-quality DDR5 modules are revealed with additional fine-tuning of the system. By manually compressing the timings, you can add another 9-10% increase to these 3%. And overclocking the Ring, D2D and NoC buses adds the final touch in the form of an additional 4% increase in the results of gaming tests.
All this means that in Arrow Lake-based systems, great attention should be paid to timing settings. High memory subsystem latencies are the main drawback of the new Intel architecture, and it is timing settings that allow you to most effectively neutralize its negative impact. Good and high-quality overclocker memory like Adata XPG Lancer CUDIMM DDR5-8800 will serve as an excellent source material in this case – for such memory, manufacturers choose the highest quality and most flexible DDR5 SDRAM chips, and thanks to this, they are not only able to operate at high frequencies, but also provide ample opportunities for reducing latencies. Plus, the high frequency of such memory will be an additional factor playing into the hands of Arrow Lake.
⇡#Conclusions
Any study of the LGA1851 platform inevitably comes to the conclusion that the Arrow Lake memory controller has serious problems. This will have to be repeated this time: the sharply increased delays caused by the implementation of a disaggregated architecture in this generation of processors made the Core Ultra 9 285K and its brothers quite controversial products, especially when it comes to gaming applications. Compared to Raptor Lake, Arrow Lake’s gaming performance literally took a step back, as a result of which Intel not only received a fair amount of negativity from users, but also failed miserably in the current round of competition in the consumer CPU market.
Fortunately, as we found out today, the situation can be fixed. With the help of fast memory and simple settings, the average frame rate that the Core Ultra 9 285K delivers in games can be increased by a very noticeable 16%. This, of course, does not give the new Intel processor the ability to challenge the gaming leadership of the Ryzen 7 9800X3D, but at least puts it a step above the flagships of the Raptor Lake generation.
In other words, the Core Ultra 9 285K is not as hopeless as it is commonly thought. Arrow Lake processors have a pretty strong trump card up their sleeves – their memory controller is adapted to work with high-speed DDR5 modules, the frequency of which can reach 9-10 GHz. This is achieved both by supporting the CUDIMM standard and by adding a special Gear 4 mode to the memory controller for high-frequency DDR5 modules.
However, it is not that simple. As testing has shown, increasing the DDR5 frequency by itself does not improve the performance of the LGA1851 platform as much as we would like. But it turns out that there are a couple of additional techniques that reveal the potential of fast memory much more effectively. The first is increasing the frequency of internal buses connecting individual semiconductor crystals inside Arrow Lake, due to which some reduction in memory access delays is achieved. And the second is optimizing the timings of the DDR5 modules themselves, which affects the integrated performance of the memory subsystem even more.
By applying this entire set of tools, Arrow Lake can achieve a radical reduction in the total memory access latency from the original 90 ns to values of about 65-70 ns, which no longer look catastrophically high. Moreover, thanks to overclocking and additional tuning, the practical throughput also improves, which can ultimately be brought to a very impressive 140 GB/s. Thus, the main weak point of Arrow Lake can be partially neutralized.
However, to fix the gaming performance of the Core Ultra 9 285K, desire and skill alone are not enough. To do everything we did in this article and get an additional 16% of gaming performance, you also need high-quality memory, preferably CUDIMM modules that can take high-frequency modes. The Adata XPG Lancer CUDIMM RGB DDR5-8800 48GB kit used in this test is just an example of such memory that is well suited for “fixing” the Core Ultra 9 285K. The modules included in it not only perfectly overclock in frequency, revealing the full frequency potential of the Alder Lake memory controller, but can also operate at significantly reduced timings, which in total makes the Core Ultra 9 285K, although not the best, but quite an acceptable solution for gaming PCs.