Today we’ll talk about c: find out what XMP is, see what its inclusion gives, and decide whether it makes sense to overclock the memory manually. We will understand the issue using the example of DDR4 ram: despite the fifth generation of DDR memory that has already appeared, it will not be widely used soon.
XMP: pros and cons
So what is XMP? To answer this question, you need to start with the basics. RAM of any standard has a so-called. minimum base frequency. According to the JEDEC specifications, for DDR4 it is 2133 MHz. There is another important figure: the recommended frequency of RAM for a particular processor. For example, for the Intel Core i7-6700K , released in 2015, it was the same 2133 MHz, but subsequently only increased due to the growing appetites of the CPU. So, the Core i9-9900K already required 2666 MHz, and the Core i9-11900K– 3200 MHz. But how to achieve such frequencies if the base is 2133 MHz? With overclocking. You can either overclock the memory yourself, which will require knowledge, time and experience, or using a profile that already contains the necessary parameters for frequencies, voltages and timings (delays). These profiles are created by RAM manufacturers and are called Xtreme Memory Profile – XMP. Thus, XMP is actually factory overclocking, to activate which you only need to press a couple of keys.
Another plus of this overclocking method is that the modern market offers a huge selection of memory with XMP frequencies up to 5100 MHz, but it’s not a fact that all this will work. And that’s why.
Firstly, successful memory overclocking depends not only on the DDR modules themselves: the motherboard with the processor plays an equally important role. If the motherboard is from the budget, then it most likely has few metallization layers and a weak component base, which may well affect both frequencies and timings.
The processor is important because it contains the memory controller that sets the ceiling for its overclocking, and the quality of this controller varies from CPU to CPU, even if it is the same model. Due to the volume of production, the plant is not physically able to test eachthe processor that came off the assembly line manually, and therefore its passport does not indicate the limit of possibilities, but guaranteed modes. Say, for the Intel Core i9-11900K, which we talked about above, the recommended memory frequency is 3200 MHz. Within these values, it is stable, but everything that lies beyond this threshold is pure fortune: before the processor is launched in the finished system, it is impossible to find out how successful it is. Because of this, situations arise when someone buys memory with a frequency of 5100 MHz in XMP, brings it home, and nothing works there, although everything went like clockwork for someone else with the same processor and memory. Such cases are called “losing the silicon lottery.”
XMP has a second pitfall, and it lies in the timings and voltages. You need to understand: an XMP profile is a rather meager set of information stored in the SPD RAM module. Of the really useful there are only five primary timings and the voltage needed to start the memory. Everything else is the concern of the motherboard automation, which will need to select several dozen secondary timings, the voltage for the memory controller, the voltage for the processor system agent … And these are just the basic parameters: there are much more auxiliary ones. It is easy to guess that with so many variables, the chance of an error is very high, so often the motherboard is not able to start the memory at all. They like to attribute this to incompatibility, although in reality these are just incorrect values set automatically in the BIOS.
We took the following set of test hardware. Intel Core i9-11900K processor with 4.3 GHz ring bus, cores overclocked to 4.9 GHz, and 100 MHz AVX shift. ASUS ROG Strix Z590-E Gaming motherboard , Nvidia GeForce RTX 3060 Ti FE graphics card, and Patriot Viper PVB416G400C9K memory with Samsung B-Die chips .
Since the memory in our case is very flexible, with its help we were able to simulate several typical situations. The first will reflect the picture in which the XMP profile in the PC is not activated at all. This often happens when a PC is ordered already assembled, and a person uses it as is, without delving into the BIOS settings. Accordingly, in this scenario, DDR is launched at the base 2133 MHz with the timings specified in the SPD. In our case, this is 15-15-15-36 CR2 at a voltage of 1.2 V.
The following profile we prepared mimics the popular memory kit from Crucial: BL16G32C16U4B . Frequency, timings and voltage are appropriate: 3200 MHz, 16-18-18-36 CR2, 1.35 V.
It is comparable to the higher frequency XMP variant from Patriot: PVS416G373C7K . To simulate it, we copied the following parameters into the BIOS: 3733 MHz, 17-21-21-41 CR2, 1.35 V.
Next, let’s see what manual overclocking will give us. Keeping the same frequency at 3733 MHz, we raise the voltage to 1.55 V and lower the primary timings to 14-14-14-28 CR1. The remaining secondary and tertiary delays will not be left to automation, as in previous cases, but will be optimized independently.
Now the highest frequency preset: 4000 MHz, 15-15-15-30 CR1 and 1.6 V. It is here instead of our native memory (Patriot Viper PVB416G400C9K) XMP profile with a frequency of 4000 MHz, timings 19-21-21-41 CR2 and a voltage of 1.35 V. We decided to make the profile this way because of the 11900K processor, which has a different memory controller from all those that Intel has encountered since the advent of the Skylake architecture .
It uses the same dividers as AMD ‘s. With a ratio of 1:1 – or, as Intel calls it, in Gear1 mode (the frequency of the memory controller is equal to the frequency of the RAM) – the lowest latency (latency) of the memory subsystem is achieved, but its frequency potential is limited to 3733-3800 MHz. In a ratio of 2:1 (Gear2), values from 4000 MHz become available. However, this is how the controller and DDR modules work in asynchronous mode, which slows down the entire memory subsystem, and dividends from its high frequency are nullified. To demonstrate this, we accelerated our XMP profile by clamping the primary and secondary timings to the limit possible at the selected voltage: it’s more visual.
Performance was measured primarily in games. Settings: Full HD, ultra-graphics presets, RT effects turned off and anti-aliasing turned off. From the working programs, we chose Adobe Premiere Pro (visualization of a five-minute video in 4K resolution with the H.264 codec), Sony Vegas Pro (the same task, but using the x264 codec) and Blender (built-in benchmark). From pure synthetics, we took 7-Zip , Corona Benchmark and Aida64 Extreme.
Note that the frequencies of the processor cores and its ring bus have always remained unchanged. So we eliminated their impact on the final performance.
First, let’s compare our two “fake” XMP profiles with stock memory. And one thing can be said right away: if you ordered a ready-made PC assembly, never heard of factory overclocking profiles, or reset BIOS settings, check if you have XMP activated or not. Without it, the performance in both games and working software is sad: the difference compared to XMP is on average 18% (the minimum difference was recorded in Assassin’s Creed Valhalla and amounted to 2%, the maximum – in Dying Light 2 : 60%) . Therefore, if it turned out that the XMP of your memory is somewhere around 2666 MHz with CL timing around 19, which is not uncommon, then try to overclock it or at least reduce the delays.
The next thing we want to pay attention to is the difference between the two XMPs. In most tests, it is either practically absent, or it adds up in favor of a preset with a lower frequency. This is explained by too high timings for the Patriot PVS416G373C7K we are imitating. They neutralize all the advantages of 3733 MHz, from which a simple conclusion follows: megahertz itself does not give anything.
Now XMP vs manually configured modules (3733 MHz 14-14-14-28 CR1). In the case of software, the increase is small, so if your home PC is a workstation, there is not much point in fooling your head.
The exception is archivers: they have a significant performance increase. As for games, there is no clear conclusion here: in most cases, the increase in fps is tiny, and a similar picture can be traced almost everywhere where the main load falls on the video card. On the other hand, in eSports titles, where the graphics are simpler and most of the work falls on the processor, everything is different. In CS: GO performance increased by 5-9%, in Quake Champions by 7.6-9.5%.
But besides the average number of fps, there is one more nuance: friezes. In the case of XMP, they periodically got out here and there. More specifically, the problems were in Call of Duty: Warzone , where 0.1% dropped to 8 fps, in Assassin’s Creed Valhalla (12 fps), in Cyberpunk 2077 (14 fps) and in F1 2021 (37 fps). These values were not included in the graphs only because in such cases we restarted the game and brought the final results to the average value. A PC with manually tuned memory was much more stable and always showed the same result, although the conditions were the same: three test runs with an OS reboot after each.
As for Gear2, the conclusions are simple: you don’t need to turn it on. Despite the impressive 4000 MHz and CL15, it surpassed only XMP profiles: 3733 MHz in Gear1 mode is still faster. Given that the preset required an undesirable 1.6 V voltage, this is a dubious pleasure, for the sake of which we did not torture the modules for a long time. We decided to limit ourselves to four games that most readily respond to changes in the memory subsystem.
Let’s summarize. As you can see from the tests, keeping the memory at frequencies in the region of 2133-2600 MHz is at least stupid. Therefore, if possible, be sure to enable XMP, and if it is not registered, find a good guide on basic overclocking (at least the first timings) on the Web and get the most out of your kit. Whether it is worth going further, abandoning XMP and getting into secondary timings is your choice. In most cases, the profit will not be worth the effort, nor the risk of ruining the memory with increased voltage over a long distance. However, in some cases the system will still become more stable, small friezes will disappear, and some games will add 5-6% fps.
Motherboard: ASUS ROG Strix Z590-E Gaming (Intel Z590, LGA 1200, 14nm, 4x DDR4
Processor: Intel Core i9-11900K (LGA 1200, Rocket Lake-S, 16MB, 4.9GHz )
RAM: 2x 8GB Patriot Viper PVB416G400C9K (4000MHz CL 19, 1.35V)
Storage: Crucial MX 500 SSD (250GB, SATA Rev.3), Patriot P210 (1TB, SATA Rev.3), Patriot P300 (1TB, PCIe Gen3 x4)
Cooling system: be quiet! Silent Loop 2
Power supply: 850 W be quiet! Dark Power 12
Case: be quiet! Silent Base 802 Window