Manufacturers ship RX 9070 XT cards with GDDR6 memory from either Hynix or Samsung. Clocks are the same, yet performance and thermals can differ slightly—and the same goes for overclocking. Samsung chips tend to reach higher memory clocks, which suggests better performance, but it isn’t that simple. We will compare the practical differences between both variants using the same card: Gigabyte’s Radeon RX 9070 XT Gaming OC.
The Radeon RX 9070 XT uses older, slower GDDR6 chips. When overclocking them, you can reach values the card can no longer handle, resulting in a black screen and a completely frozen computer.
However, a situation can also occur where the card continues to run even with memory clocks at which errors are already happening, thanks to built-in error correction. But if the card is forced to correct errors frequently, the memory bandwidth decreases, and with it, the card’s performance.
In other words, you can keep increasing the clocks—it might look like everything is fine—but in reality, instead of higher performance, higher memory clocks can lead to a decrease in performance.
Guides usually recommend starting overclocking at some safe value, slowly increasing clocks, always testing stability thoroughly, and only then going higher. I didn’t have enough time for that, so I did it differently—I quickly found out what the card could no longer withstand and then set lower values.
A popular test for verifying memory bandwidth and stability is Memtest_vulkan, which you can find on GitHub. The situation with Radeons is a bit more complicated than with GeForce because, in the AMD Software, you can set faster timings for the memory in addition to higher clocks.
However, faster timings come with slightly worse stability. So, quickly, using Memtest and AMD software, I found the ceiling both with default settings and with fast memory timings. During testing, I did not register any significant fluctuations in bandwidth due to error correction, most likely because before that could happen, the overclocking attempt ended directly with a black screen.
Another minor detail I noted during the memtest was the higher temperatures of the Hynix memory chips, which were already around 102 °C. I therefore set the fan speed to 70% to prevent thermal throttling from reducing the clock speeds.
With the default settings, I then quickly increased the memory clock from a “safe” ~2650 MHz in steps of 50, and later just 20 MHz, while the memtest was running. One more minor point, mainly relevant for the subsequent operational tests: the monitoring software reports slightly lower clock speeds than those you set in the AMD Software.
Hynix topped out at 2,760 MHz
After increasing the clock to 2760 MHz, I ended up with a black screen and a completely frozen computer. So, we have our first ceiling.
I then set the fast timings and repeated the whole process, starting from a slightly lower base clock. This process ended around 2680 MHz.
Samsung reached up to 2,820 MHz, but…
The achievable clocks on the cards with Samsung chips were somewhat higher—with default timings, I reached 2820 MHz. During this test, it didn’t result in a black screen, but the screen flickered, the fans briefly went silent, and the drivers evidently reset. So, it seems you can achieve higher clocks with Samsung chips. The problem is that this doesn’t necessarily mean higher performance, as cards with Samsung chips are slightly slower at the same memory clocks.
But let’s not get ahead of ourselves.
But I won’t get ahead of myself. In the next chapter, we’ll look at how memory overclocking boosted performance across a broad set of tests on the card with Hynix chips.
⠀
