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.
Based on the results of the previous tests, I selected Cyberpunk with Ultra settings at a resolution of 2560 × 1440 pixels to demonstrate the impact of memory clock on performance. It is one of the tests where performance scaled with memory clock better than average. I started at the default clock, then jumped to a higher stable clock speed, and subsequently increased the clocks in small steps. For each setting, I ran three test loops, with the final values being the average of three measurements. And just to be sure—to prevent temperatures (both GPU and memory) from causing variability—I increased the fan speed to 70% for the tests, which corresponds to approximately 3300 RPM.
One more minor point I must highlight: To make the impact of the clock speeds on performance visible, I had to crop the bottom part of the graph. Therefore, the Y-axis does not start at zero but shows values between 120 and 140 fps. Even though the Samsung memory can be overclocked slightly higher, the overall performance was better with the Hynix chips.
On the chart for the Hynix-equipped card, the tail of the curve clearly shows that once you push the memory clocks “past the line,” performance starts to drop.
The table below lists the exact measurements for each memory clock. Note that at the higher clocks the card can’t really be called stable; all we can say is that it made it through the Cyberpunk test. And as a reminder, memtest produced black screens at much lower clocks—around 2,750 MHz with Hynix and 2,820 MHz with Samsung.
| GDRR6 clock (HWiNFO) | GDDR6 Hynix | GDDR6 Samsung | difference |
| 2505 MHz | 135,1 fps | 133,1 fps | 1,5 % |
| 2505 MHz | 134,8 fps | 133,4 fps | 1,0 % |
| 2646 MHz | 136,6 fps | 133,8 fps | 2,1 % |
| 2666 MHz | 136,8 fps | 134,1 fps | 2,0 % |
| 2686 MHz | 136,9 fps | 133,4 fps | 2,6 % |
| 2706 MHz | 137,1 fps | 133,7 fps | 2,6 % |
| 2726 MHz | 137,4 fps | 134,1 fps | 2,4 % |
| 2746 MHz | 137,7 fps | 134,2 fps | 2,6 % |
| 2766 MHz | 137,8 fps | 134,4 fps | 2,6 % |
| 2786 MHz | 138,1 fps | 134,5 fps | 2,6 % |
| 2805 MHz | 138,2 fps | 134,1 fps | 3,1 % |
| 2825 MHz | 137,7 fps | 134,3 fps | 2,5 % |
| 2845 MHz | 134,7 fps | 134,3 fps | 0,3 % |
| 2865 MHz | 133,7 fps |
With higher memory clocks, memory power draw and temperatures also rise a bit, but the differences are only a few watts.
To wrap up, remember that memory clock or timings aren’t the only things that affect final performance. The results can also be influenced by chip-to-chip variation (the “silicon lottery”) in both the memory and the GPU. With differences this small, even a few dozen MHz on the GPU core can matter. The gains shown in this chapter are on the optimistic side—overall, the benefit from memory overclocking was rather small, and the performance gap between Samsung and Hynix chips was small as well.
I may also have lucked out with a good batch of Hynix memory, since the frequencies I reached seemed slightly above average.
All in all, unless you’re chasing records or leaderboard positions, the performance differences are so small that it doesn’t really make sense to chase a card with one vendor’s chips over the other.
Despite the higher temperatures, I still lean toward the Hynix-based card, though some people will prefer having more thermal headroom before hitting the limits.
⠀
English translation and edit by Jozef Dudáš
