Despite the huge number of troubles that fell on the shoulders of people in 2020, last year was also full of wonderful events, such as the release of the long-awaited Zen 3 microarchitecture and the no less promising RDNA2.
If there was no doubt about the processors that AMD did succeed again, then the fate of the Navi 21 video cards was very vague. Ray tracing, a significant increase in operating frequencies, a new architecture, Variable Rate Shading, AMD Smart Access Memory, L3 Infinity Cache and a suspiciously low memory bus width constantly muddied the waters, but, fortunately, everything worked out.
The beautifully veiled L3 Infinity Cache turned out to be nothing more than a horseman in gleaming armor, whose task was to increase memory bandwidth and he succeeded. 128 MB with a 1024-bit interface reduces the number of external memory accesses, which ultimately allows a 256-bit interface to be more energy efficient than 512 and have a similar throughput.
Complementing this picture are 5120 stream processors and an average frequency of 2200 MHz, which looks quite impressive compared to the capabilities of Navi 10.
If we talk about specific numbers, then the increase in energy efficiency was more than 50%, and with the same power, RDNA2 can operate at frequencies 30% higher. Nevertheless, knowing AMD and its predilection for overvoltage, I decided to create this material.
In addition to a short digression into the RDNA2 overclocking manual, you will also be waiting for comparative testing of the drain, drain + SAM and water + SAM.
AMD Smart Access Memory
I would like to pay special attention to the unique AMD Smart Access Memory technology (in the common people Resizable BAR). The technology appeared for the first time with Navi 21 video cards — it allows the central processor to use the entire video memory array of the video card at once, and not only access its part up to 256 MB.
Thanks to this, in some games you can achieve a performance increase of up to 16%. Fortunately, now, in addition to processors based on the Zen 3 microarchitecture, with the arrival of AGESA 184.108.40.206, processors with the Zen 2 microarchitecture are also supported, but only for motherboards based on the B550 and X570 chipset.
This miracle is activated quite simply (by default, this technology is disabled), in the UEFI menu we look for the PCI Subsystem Settings menu and activate Above 4G Decoding, otherwise Re-Size BAR will not be available.
Overclocking AMD Radeon RX 6800 XT
The fact is that the Navi 21 boost has an identical behavior with Zen desktop processors, which means that we are dealing with a limited PPT (burst power), TDC (current), and also temperature.
Since the video card is the same piece of silicon as the processor, this means that each sample is individual, and also that the process technology used has a “sweet spot” where the temperature does not clamp down on Boost, while it is possible to free up additional TDC for the same Boost. Why am I talking about this, because overclocking is associated in most cases with twisting the sliders to the maximum?
Everything is simple. With the RDNA2 generation, AMD decided to make a «gift» for enthusiasts, completely depriving them of the opportunity to modify the limits of the video card to achieve beyond the results. This means that everyone’s favorite MorePowerTool for RDNA2 by Igor Wallosek won’t help at the moment (I think many of you used it on the 5700XT, which could have run at more than 2100 MHz). The video card driver now knows how to check the limits and in which case it simply resets them to the drain. All that is available to the user is + 15% to PPT and TDC. Not much.
Such «gifts» are often made for products with a limited life cycle, so that the current product does not turn out to be as productive as the «refresh». Perhaps some of you will think: “this is dangerous for the video card!”. Alas, no, all video cards based on the Navi 21 chip have a very serious VRM, for which a peak current of 400–500 A is not something special or dangerous.
For example, AMD Radeon RX 6800 XT reference solutions are equipped with 70 A MOSFETs from Infineon.
Their number is 15, 12 of which are for the GPU power system and three for VRAM.
The GPU is powered by a 16-phase Infineon XDPE132G5D PWM controller, and the VRAM is IR35217 (aka IR35201).
I want to note that non-reference video cards in most cases have a similar power system. This is due to the same limitation of the video driver.
The heat flow of Navi 21 has significantly decreased relative to Navi 10, 0.57 W / mm2 versus 0.89 W / mm2 due to the increased chip area, that is, cooling the chip has become much easier than the younger solution of the previous generation.
Summing up all of the above, the only thing that will help us achieve the desired results in this situation is a lower temperature and lower operating voltage. Reducing the voltage will in turn do three positive things: actually, lowering the temperature allows the silicon to operate at a higher frequency, lowering the temperature, which means lower leakage currents, and lowering the voltage will affect the TDC, it will also decrease.
Since temperature is the only negative factor that we can deal with, I will use water cooling (more on that in the next chapter) and lowering the supply voltage.
Before proceeding to the step-by-step instructions, I want to note one important nuance — the video card can “swallow” even an overestimated VRAM frequency or Min Frequency GPU. As a result of this, the performance of the video card may decrease relative to the stock mode. This mistake was made in a huge number of reviews, because twisting the VRAM frequency slider to 2150 MHz gave almost no result, and in some games it was even negative. The same goes for Min Frequency GPU.
First of all, I set the voltage to 1050 mV (stock value 1150 mV) and Power Limit 15%. Then the limits for the conditionally minimum and conditionally maximum GPU frequencies were set. Offhand, this is 2400-2450 / 2650-2700 MHz. Then, step by step, using 3D Mark Time Spy, the GPU Min Frequency was selected so that the final result of the test benchmark did not result in a decrease in performance. The standard step is 25 MHz.
After the optimal minimum frequency was found, the GPU Max Frequency was selected in a similar way. We must pay tribute that it is easier to select the maximum frequency, since the presence of instability immediately leads to a crash and reboot of the driver, and not to stretching (frequency stretching), which is observed when entering an incorrect value for the GPU Min Frequency.
Overclocking VRAM is similar to finding the GPU Min Frequency. The optimal value was 2125 MHz, while an increase to 2150 MHz did not demonstrate instability, but significantly reduced the performance of the video card. Therefore, I repeat — the era of «everything to the maximum» (even with water) has officially died. Now, either with the mind, or better than nothing.
As a result, a result of 19198 points was obtained, which is even more than the stock AMD Radeon RX 6900 XT boasts.
I also suggest that you download my profile to speed up the time to find the perfect value.