Overclocking the Radeon 7900 XTX

In this article I look at the overclocking potential of the 7900 XTX. This is a follow-up to my review of the 7900 XTX; I’m using the same tests and settings, with new raw data added to the same shared document.

Overclocking rules and settings

I am not into competitive overclocking, and don’t care about benchmark scores; I only wanted to find the best performance for my regular usage of playing games. With that in mind, I have some rules:

• No modding. Closed case. Default fan curve.
• 100% stable system. Zero tolerance for any crash or failure.

The last item can be hard to ensure. But my final tuning passed every test (twice as many as the ones reported here) plus a few days of normal usage mixing games, productivity apps, web browsing, etc.

Here are the settings that worked for me:

• GPU: Max Frequency = 3,000 MHz, Voltage = 1110 mV
• Memory: Timing = Fast timing, Max frequency = 2,700 MHz
• Power limit: +12%

With this config my stress tests hit 390W of board power: 15W over the specced limit of the 2x8-pin PCIe cables & socket. I also hit a maximum hotspot temperature (Tjunction) of 92ºC, with the GPU fans spinning at 1,900rpm to hold that temperature: pretty good for overclocking. AMD’s reference cooler works well, at least when it’s not a defective unit.

GPU versus Memory

You will be tempted to max out every tunable setting, but that doesn’t work. Making both the GPU and VRAM run at very high frequencies results in too much power draw and/or heat, so this graphics card will not reach both limits simultaneously. The GPU’s SMU and/or the VRM will choose some balance, or you can force a choice by reserving headroom for it. I chose to leave the GPU frequency almost at stock to favor memory frequency.

I experimented with a higher GPU clock — max frequency 3,200MHz with all other settings identical — but that resulted in tiny differences, within the margin of error, in every test. Also not 100% stable, I had one test failure. I did not experiment to combine a high GPU OC with a lower memory OC, I don’t want that option for another reason: overclocking the memory is almost a “free lunch” in hotspot temperatures so I always want to max out memory OC before any GPU OC, which will cost me in Celsius. Topics like these are more detailed in Brent Justice’s OC analysis of the same card.

I owe my good results with VRAM OC in part to the “Fast timing” option. You want both that and the higher frequency. This combination may not be stable for everyone; this is part of the silicon lottery. If it doesn’t work for you, maxing out the GPU clock may be the better strategy.

My setting of 2,700MHz translates to 21,6Gbps, or a total bandwidth of 1.036TB/s in the 7900 XTX’s 384-bit VRAM bus. I imagine that the emphasis on memory OC will work even better for the 7900 XT’s 320-bit bus.

Overclocking results

Starting with the synthetic benchmarks, I get around +6% for all the main 3Dmark tests that better approximate games. The tests at 4K resolution (TimeSpy Extreme, FireStrike Ultra) gain more than their 1440p versions, perhaps showing a large contribution from the memory OC. The single outlier is the Blender rendering benchmark, barely above +1%.

My real-game benchmarks have similar results, ranging from +5% to +7.5% extra FPS. That’s the same average of +6% found in the 3DMark tests. But we have a couple of interesting outliers:

• Red Dead Redemption 2 hits almost +10% over stock, just amazing.
• Callisto Protocol gets a disappointing +1%, but that’s because the game has severe CPU/engine limitations (see again my full review).

Overclocking x Upscaling

I also tested upscaling (mostly FSR 2). Like any rendering option, this can change the balance of workload across the GPU’s subsystems: shaders, accelerators, caches, memory, etc. These changes result in varied OC impacts for each upscaling level in some games.

The most surprising result is Callisto Protocol: only +1.21% native, but FSR 2 gets +6,03% for Quality and +7,48% for Performance. If you have read my full review, Callisto had zero improvement with FSR 2 because of the game’s engine bottlenecks. Other reviewers had similar findings although you can push the game farther with faster CPUs / DRAM. I find it curious that only the combination of OC and upscaling unlocked extra FPS for me.

In Dying Light 2 I got a boost of +6,62% native but FSR 2 improved that to +10,39% for Quality, +8,26% for Performance. It’s interesting that the sweet spot is the more limited Performance setting. For Quake II RTX, the game’s TAA scales more linearly also with OC so the biggest improvement is for the Quality setting (50% scale) with a gain of +9,70%. Finally, RDR2 starts well in native rendering +6,93% with OC, but FSR 2 hits a CPU bottleneck with Performance making only +0,65% over stock (i.e. nothing); this game is just limited to 150fps in my testing platform and configuration.

Conclusions

I’m happy with the gains of a minimum +5% FPS, average +6%, even +10% in a few cases. These are not “free upgrade to a more expensive GPU” gains, but noticeable, and better than my expectations for AMD’s reference 7900 XTX whose cooler and TDP look modest next to this generation’s 3.5-slot behemoths. The reference card is more limiting if you care about fan noise and temperatures, but this is true even at stock; see Gamers Nexus’s review.

In Dying Light 2 — now speaking in absolute numbers — the game starts at 52fps at stock, goes to 55fps with OC only or 82fps with FSR 2 Quality only, and 90fps with both OC & FSR 2 Quality. Results like this are pretty good, justifying the overclock for regular use, at least with headphones so I don’t care if the GPU fans spin a little harder.