Tobias “Rauf” Bergström leads us through the process and way of thinking that is required for extreme overclocking when the goal is to set a world record for how far computer hardware that Geforce RTX 5090 can be pressed.
My name is Tobias Bergström but goes under Alias Rauf in my semi -professional hobby to compete in overclocking. In this article, I take you through the process behind setting a world record with Nvidia's latest generation graphics card RTX 5090.
Competing in overclocking, or XOC (Extreme Overclocking), is not exactly one of the biggest sports available but it has a strong tradition that can be traced back to at least the 1980s.But first and foremost, what is really overclocking? In most cases, competing in overclocking is really about competing in computer performance. Overclocking is the process of getting a computer component to work at a frequency higher than it was designed for, but the end goal is of course higher performance.
Like trimming a car
I usually explain the competition in overclocking by comparing cars and rally. You trim your components as far as you can to make them perform better. Just as with an engine, a computer circuit usually reacts well to better cooling. Therefore, we use liquid nitrogen as refrigerants, as it is affordable, easy to handle and relatively safe.
In addition to the engine itself in the comparison, we disconnect all protective functions that could otherwise hamper performance. In the rally, airbags, speed limitation, anti -slip and auto brake disappear. Within XOC, we disconnect protection such as temperature monitoring, voltage and power restrictions and the like. Just like the rally where you study the track, the optimal speeds and curve tags learn, we learn in the overclocking about the best operating systems and settings to find the fastest roads here as well.
Finally, we do the competition by measuring the performance with different Benchmarks, where you get a result that you then charge at the top of the database hwbot.org. There are the leaderboards and we can compare ourselves to people from all over the world.
The hardware
I have the privilege of having sponsor contacts for some of the hardware used after many years of competitors. Often, it happens in the form of hardware being sent home to me before a launch of a new generation of products. Sometimes I go to their R&D or development lab. The hardware is often in the form of “engineering samples” (engineering copies), thus not completely ready for the market and is delivered without coolers and cosmetic accessories. In this case, the galaxy has sent copies of its RTX 5090D Hall of Fame graphics cards.
Before the overclocking can begin, the components need to be protected, and sometimes somewhat modified somewhat. I paint the parts to be cooled to protect against condensation and short circuit due to water released at the low temperatures. In this case, the graphics card's power measurement has also been short -circuited, so that the card should not be limited by the fact that it strikes the power ceiling.
The software
The next step is the most time consuming, and one of the things that takes a long time to get really good at. To optimize operating systems and drivers and get the products to perform as well as possible before even starting to screw up the frequency.
We call it efficiency, although in our opinion it has nothing to do with performance per watt, but rather performance per MHz. In some benchmarks it can be very important and some fine adjustments are very creative (and secret).
First tests with liquid nitrogen
When I feel satisfied with the efficiency, it is time for the first test with liquid nitrogen. Sometimes we get some guidance from a R&D department of what we can expect and what frequencies, voltages and temperatures should work.
Often it is still relatively untested land so it is important to test all limits and not be limited by information about how something “should” work. It can differ a lot between two copies of the same product.
In the case of the RTX 5090, it turns out that despite the fact that in the foundation and the manufacturing process is very similar to the predecessor RTX 4090, there is a huge difference in how they work with liquid nitrogen. The RTX 4090 could be cooled down as far as possible with liquid nitrogen (-196 ° C), and only worked better the colder it was.
For the RTX 5090, the graphics card shuts down if the core is below -40 ° C (we call it “cold bug”). The RTX 5090 also turns out to have a peculiarity where the frequency is locked to a maximum of 3,270 MHz if the temperature of the core is less than 0 ° C at load. It also turns out that if the temperature goes below 0 ° C, efficiency becomes very suffering.
For the RTX 5090, it is all about finding the right balance, the right temperature on the cooler which gives just above 0 ° C internally in the core. The right tension that is sufficient to maintain high frequency, but not so high that it kills the card or so high that it is not possible to cool it fast enough.
When overclocking with liquid nitrogen, we use specially made copper containers as a cooling block, which we mount on the graphics card in this case. Then we pour liquid nitrogen into the container to cool the component.
A major problem with the 5090 series is that liquid nitrogen is not so reactive at such high temperatures. It may sound illogical, but liquid nitrogen is slowly evaporated at relatively high temperatures such as -40 ° C, while it starts to evaporate very quickly when it meets a surface that is -150 ° C or colder.
We want the evaporation to be done as quickly as possible as it gives the best cooling effect. Hence, we use a trick that counteracts the Leidenfrost effect, by using a gas burner, ensuring that an ice surface is formed over the copper in the cooler. This reduces the thermal insulating gas layer that is formed when the nitrogen evaporates and the cooling and the evaporation take place more efficiently.
Despite this, it is still difficult to cool the core fast enough, as the power goes from about 150 W idle to over 1000 W immediately at the start of benchmark. For several Benchmarks there are also several part tests, where GPUN goes in Idle between the tests. Then the problem is reversed, the temperature fades when the load from the first part test runs out and I have to tighten the end of the test perfect to have as little liquid nitrogen as possible in the cooler and then quickly burn on with the burner to counteract that the temperature is too low when the next test starts.
The choice of coolers has a great impact to be able to cool the card in a good way. I get to try it out between and find the appropriate relationship between the cooler's mass, internal structure and contact surface.
One challenge is that the load during a test is not even, it can distinguish 300 – 400 W between different sequences in the test. At best, I can keep the temperature of the cooler even during the test, but when the load and power occasionally fall, it decreases the internal temperature in the core, even if the temperature in the radiator is constant.
Here I have to have a small margin to the magic limit of 0 ° C, the temperature goes below, even if only a second the test is destroyed. In addition, the cooler will want to sink in temperature when the load goes down as it has just been perfectly balanced against a higher power development.
Here it is important to get to know each benchmark perfectly, so that the power peaks and valleys can be parked. The power peaks are equally tricky, as the temperature can quickly increase here and lead to gpun crashing.
RTX 5090 – get everything to flap
When I get to know GPUN and know what tension, frequency and temperature that my copy likes, how to parry the different tests and what operating systems and settings are the best, it is important to get the whole together.
Here it is important to pair GPUN with processor, motherboard and frame memories that provide the best conditions for each benchmark. Some tests such as 3Dmark Time Spy requires very powerful workstation CPUER, while other tests require high frequency and fast gaming processors and frame memories. Then the CPUs also need to be cooled with liquid nitrogen and overclocked to the limit.
In summary, it is of course always fun to test new products and get the chance to push them to the utmost that few others can. Although the RTX 5090 has unfavorable “cold bug” and its other peculiarities in terms of temperatures, it has been a slightly more interesting product to test compared to the latest generations.
Nvidia's graphics card has worked very much since the 1080 series came, so little change is content. Hopefully, some of the clarities can also disappear over time with new BIOS and drivers.
The results for my part have meant frequencies on the graphics core of up to 3,460 MHz and the memories of 2,250 MHz, compared with standard frequencies of about 2,700 MHz and 1 750 MHz respectively. Which is not so bad considering the limited cooling possibility.
During these first tests I have beaten 7 world records in different Benchmarks such as 3Dmark, Superpositions and Gpupi. At the time of writing, I hold the record in 3 Benchmarks, and hope to be able to take back some of the others!