{"id":265535,"date":"2025-11-17T23:44:12","date_gmt":"2025-11-17T22:44:12","guid":{"rendered":"https:\/\/www.hwcooling.net\/?p=265535\/"},"modified":"2025-11-20T01:08:58","modified_gmt":"2025-11-20T00:08:58","slug":"hynix-vs-samsung-on-rx-9070-xt-in-depth-gddr6-thermal-analysis","status":"publish","type":"post","link":"https:\/\/www.hwcooling.net\/en\/hynix-vs-samsung-on-rx-9070-xt-in-depth-gddr6-thermal-analysis\/","title":{"rendered":"Hynix vs Samsung on RX 9070 XT: in-depth GDDR6 thermal analysis"},"content":{"rendered":"

Radeon RX 9070 XT cards use GDDR6 from multiple vendors. The biggest differences between Samsung and Hynix appear in temperature and power\u2014Samsung runs notably cooler. Earlier tests with automatic fan control confirmed this. Today I\u2019m adding a temperature comparison at equal fan speeds and, alongside internal sensor data, surface temperatures measured on the chips. That gap is much smaller than expected.<\/p>\n

Here on HWCooling, I’ve already compared the Radeon RX 9070 XT with Samsung and Hynix memory in two articles. The first focused on a detailed comparison of clocks, thermals and power consumption<\/a>, while the second concentrated on performance<\/a>.<\/p>\n

When comparing cards with Samsung and Hynix memory, the most discussed aspects are the differing chip temperatures and their power consumption. I focused on this primarily in the introductory article, where I compared the operating characteristics of the cards on both BIOS versions with default settings and automatic regulation.<\/p>\n

However, such a comparison has a few shortcomings. Regarding the memory temperatures themselves, I relied solely on what the card’s monitoring and internal sensors near the chips reported. The second shortcoming is that the average GPU temperature under load differed slightly between the two cards, and more importantly, the hotspot temperatures\u2014the hottest sensor on the chip\u2014differed. Because of this, the fan speeds on the cooler diverged for both cards after they warmed up.<\/p>\n

If we are primarily interested in how GDDR6 from Samsung differs from Hynix chips under similar conditions, we need to constrain the automatic regulation. If the fans on one card are spinning faster, the memory chips will necessarily be cooled better and have lower temperatures.<\/p>\n

Therefore, in today’s article, I will supplement the previous tests with temperature measurements using external sensors at default settings, and add tests of the cards’ operating characteristics with the fans set to the same speed.<\/p>\n

The card with Samsung chips is the same as in the previous tests. The Radeon with Hynix chips in this comparison is already the third sample\u2014the white variant, the RX 9070 XT Gaming OC Ice 16G from last week’s test<\/a>.<\/p>\n

\"\"<\/a><\/p>\n

With some help from a Raspberry Pi Pico 2<\/h2>\n

Most graphics cards today are covered by a metal backplate or plastic shroud that a thermal camera can\u2019t see through. It shows where the heat sources or thermal pads are, but it won\u2019t give you the actual memory temperatures. Under load, the temperatures on the plate\u2019s surface were much lower than the sensor readings.<\/p>\n

\"\"<\/a><\/p>\n

In addition to the usual measurements from the card’s monitoring recorded via HWiNFO, we will also add temperatures measured using three external sensors. We will measure temperatures using thin-film thermistors (type NTC 3435 10K) on a polyimide foil, connected to a Raspberry Pi Pico 2 via voltage dividers with an RC filter. The software processing the measured values was “written” by ChatGPT. It’s not exactly trivial. Most of the pitfalls of this solution\u2014such as deviations caused by component heating, noise, or thermistor characteristics\u2014are compensated for in software. The NTC circuit utilizes duty cycle power supply with a settling delay before measurement, software oversampling (multiple sampling + averaging), and linear five-point segmented calibration that corrects for non-linearity and unit-to-unit probe variations.<\/p>\n

The thermometer is calibrated on a five-point curve against a UNI-T UT325 thermometer and a Fluke Ti125 thermal camera at temperatures of 26 \u00b0C, 40 \u00b0C, 60 \u00b0C, 80 \u00b0C, and 100 \u00b0C. For calibration, I used the heated bed of a Bambu Lab A1 (which has another temperature sensor on it).<\/p>\n

Each thermometer used during calibration has different measurement accuracy tolerances. The key point is that the differences between the thermometers in the 26\u2013100 \u00b0C temperature range were within \u00b1 1 \u00b0C. After calibration, the temperature differences between the trio of sensors connected to the Raspberry should be under 0.2\u20130.3 \u00b0C. This is more due to inaccuracies from manual calibration of individual sensors rather than inaccuracy of the device itself.<\/p>\n

However, in the graphs in the following chapters, the temperatures from the Raspberry sensors will differ by several degrees Celsius. These differences are not due to measurement inaccuracy, but rather because those specific locations have different temperatures under load. For illustration, let’s look at how the temperature profiles change on the sensors of a card that is not under load. Sensor S0 is blue; it is inserted between the chip and the heatsink, near the chips between the GPU and the bracket. Sensor S1 is orange; it’s on the opposite side of the GPU, between the heatsink and the memory on the rear part of the PCB. The third sensor, S2, is green. It is below sensor S1, but on the opposite side of the PCB, between it and the backplate.<\/p>\n

\"\"<\/a><\/p>\n

The first hundred seconds show only the desktop with fans set to 3500 RPM. Between 100\u2013500 seconds, I opened the AMD software, causing the memory to heat up a bit more. At the 500-second mark, I switched the regulation to automatic, so the card turned off the fans. And after the 700-second mark, I closed the AMD Software and left only the desktop displayed again.<\/p>\n

\"\"<\/a><\/p>\n

You can see that when the fans are spinning, the temperature on the GDDR6 chip between the GPU and the bracket (S0) is slightly higher. I explain this by the fact that the heatsink draws heat away faster from the chips in the rear part of the card, where a larger portion of the heatsink is cooled by two fans, making those chips cooler. This is also evident from the temperatures equalizing across the sensors after the fans are turned off.<\/p>\n

Compared to the values we usually see in monitoring software, the temperature curves are substantially smoother. This isn’t because I’m smoothing the measured data. Out of interest, you can look at the second image to see the typical measurement error and how the temperatures differ on individual sensors after calibration when the card has been in a stable state for a longer period. In the smaller window showing the temperature of each sensor, next to the post-calibration values in brackets, are the raw temperature values obtained by calculation based purely on the parameters of the components used.<\/p>\n

\"\"<\/a><\/p>\n

Even though I was surprised by how accurately the sensor data matches, please still consider it as indicative. I’m not in the business of building precision measuring instruments; this is a “thermometer” assembled from cheap components.<\/p>\n