Today we have prepared one of these "for the sake of science" reviews that we like to put together from time to time. The Radeon RX 6700 XT is the latest GPU from AMD that was recently tested and offers RTX 2080 Ti-like performance. While this is an impressive GPU, it's not great in terms of value, but that's the theme for 2021.
Compared to the previous model, the 5700 XT, we expect an average of 30% more performance with a price increase of 20%. So this isn't a great value for a generational leap, but still a solid output. When you consider that the 5700 XT and 6700 XT have the exact same core configuration using the exact TSMC 7nm manufacturing process, a 30% increase in performance is nothing short of astonishing.
Typically, you can expect such big gains when you switch to a newer and superior manufacturing process like we've seen for Nvidia's Pascal architecture, going from the 28nm process used by Maxwell to the 16nm process was converted by TSMC. This enabled higher clock speeds, which dramatically improved performance and efficiency. However, AMD managed to do something similar in the same process. But how do you get there 30% faster?
There are a number of architectural differences between RDNA1 and RDNA2. The 6700 XT cores have a slight IPC advantage, but the geometry performance when discarding basic elements is worse, since the 6700 XT is configured with only half of the primitive fixed functional units compared to the 5700 XT.
At least for now, the decisive performance advantage of the 6700 XT over the 5700 XT seems to be the clock speed. On paper, the 5700 XT is said to work with a boost frequency of 1905 MHz, while the 6700 XT boost clock is 2581 MHz, which corresponds to a significant increase of 35%. In our tests, the difference is closer to 42% than when comparing the AMD reference models.
Nevertheless, the Radeon RX 6700 XT should be at least 35% higher and we were able to determine an average increase in performance of 30%. So it seems that a lot of this gain is simply due to the architectural design that allows RDNA2 to clock higher. We're going to look into that today. If the 5700 XT and 6700 XT are clocked at the same frequency and most of the increase in performance is simply due to the improved clock speeds, this test will easily highlight that and we can verify that our guesses prove to be true.
We use AMD reference models for testing. Both were clocked at 1.8 GHz as we were able to maintain that frequency by maximizing the performance goal. At 1.9 GHz, the 5700 XT's frequency began to fluctuate quite a bit, often dropping to 1.8 GHz. However, at 1.8 GHz, the clock speeds were very consistent for both models, so that's about as fast in terms of operating frequency as you can expect from apples to apples.
During the test, the clock speed was constantly monitored to ensure that both GPUs were operating at the target frequency.
As for memory, we left that inventory as it is the best way to run this test. You might think that this gives the 5700 XT an advantage as the memory bandwidth is 17% higher thanks to the wider memory bus, but it doesn't. If anything, the benefit is carried over to the 6700 XT as it uses the memory much more efficiently with better delta color compression and most importantly the infinity cache. Essentially, RDNA2 can achieve higher frame rates for a given bandwidth.
The infinity cache plays a key role, the 96 MB on-chip cache works similarly to an L3 cache on a CPU. This local cache buffers reads and writes to main memory and is much faster than working out from VRAM, increasing the memory bandwidth of the 6700 XT compared to the 5700 XT.
With that said, now is the time to test these GPUs cycle by cycle. We'll do this on our new Ryzen 9 5950X test system with 32GB of DDR4-3200 CL14 memory. Let's get into that.
That's pretty interesting, if not entirely unexpected. In Watch Dogs Legion we see very similar performance when they are matched at the same clock speed.
The average frame rates are identical for all three resolutions tested, although the 1% lower performance for the 6700 XT is consistently 4-9% better, depending on the resolution.
The Assassin's Creed Valhalla scores are also very similar, although this time the 5700 XT was slightly faster at all three resolutions tested, delivering an additional 2-3 fps.
Again, both GPUs maintained an operating frequency of 1.8 GHz, so the differences are due to the changes in the architectural design.
The results of the F1 2020 were too tight to be named, as we are only looking at a deviation of the data of up to 3%. The 5700 XT was a few percent faster at 1080p and 1440p, while the opposite was the case at 4K.
The results of Rainbow Six Siege are a little more interesting. At 1080p and 1440p, the 6700 XT was ~ 3% slower than the 5700 XT, which is a negligible margin. However, at 4K we see that the 6700 XT boosts performance by 8% over the 5700 XT. I suspect that this improved 4K performance is due to the Infinity cache.
We see similar behavior in Shadow of the Tomb Raider. At 1080p, the 6700 XT offers no advantage over the 5700 XT if it is adjusted clock by clock.
However, at 1440p the 6700 XT is up to 6% faster and then 10% faster at 4K. Again, I believe this is due to superior memory management.
Horizon Zero Dawn's performance is pretty even across the board. The 6700 XT was a few frames slower at 1080p, similar at 1440p, and then a few frames faster at 4K, although overall it can be said that the data is within margin of error despite the three-run average.
Tests with Death Stranding show identical performance between these two GPUs when clocked at 1.8 GHz. We see exactly the same frames per second in all three resolutions tested.
The frame rates in Hitman 2 are also similar, though the 6700 XT starts out with increasing resolution, if only a tiny 2-3 fps.
Finally, we have Cyberpunk 2077 and the performance trends are again similar. We see identical numbers at 1080p, with the 6700 XT just managing to push the higher resolutions forward.
What we learned
So there you have it, a quick and easy yardstick for this one. Basically, it is confirmed in today's games that the majority of the performance increase for the Radeon RX 6700 XT is due to the increased operating frequency of the GPU cores. It's a remarkable feat that AMD made that leap on the same process node, despite clearly having experience working with TSMC's 7nm TSM beyond GPUs.
AMD anticipated its goal with RDNA2 to achieve a 50% per watt performance jump over RDNA1, and that this would be achieved entirely with architectural improvements, not process improvements. We knew it was coming, but I have to admit I was skeptical.
A lofty goal, but as far as we can tell, they made it with the 6700 XT. They deliver 30% more power than the 5700 XT and at the same time reduce power consumption by around 15%. With the next generation of games, we expect the margin to widen between 5700 XT and 6700 XT, so this will be an interesting situation for years to come.
In terms of performance, the 6700 XT is a necessary step up from the 5700 XT to keep up with stiff competition from Nvidia and the equally aggressive RTX lineup of this generation. Unfortunately, they are all being undermined across the board by market conditions and price increases.
AMD caught up with Nvidia on grid performance, and with RDNA3 expected to take another step forward, this is great news for gamers. Hopefully things will have returned to normal by then, which should lead to improved availability and a price war between Nvidia, AMD, and maybe even Intel, though that's probably a little too optimistic at this point.
- AMD Radeon RX 6700 XT on Amazon
- AMD Radeon RX 5700 XT on Amazon
- Nvidia GeForce RTX 3070 on Amazon
- Nvidia GeForce RTX 3060 Ti on Amazon
- Nvidia GeForce RTX 3060 on Amazon
- AMD Radeon RX 6800 on Amazon
- Nvidia GeForce RTX 3080 on Amazon