Question Why overclocking 7900XTX is almost meanless?

[FAhentai]

Hi guys, I am a Chinese- Australian. Recently I had watched a review of all brands' 7900XTX in China for one vloger. If you do understand Chinese, here is the link
https://space.bilibili.com/434570236/channel/collectiondetail?sid=1727031
However, there is one test result that makes me feel so strange. According to his test, there is one card from asrock, which is AMD Radeon™ RX 7900 XTX AQUA 24GB OC, that has 550W max power consumption, with stock split water cooling cooler. When it comes to 3D mark test, this very card hits 33000 on 3D Mark time spy, which hits RTX4090 level and pales the rest of 7900XTX. Later, he did the test again, by install Asrock AMD Radeon™ RX 7900 XTX AQUA 24GB OC's bios into other brands' 7900XTX and put his Aio cooler (his primary job is selling customized AIO cooler for graphic cards), manually over clocked card, the result is almost the same. Yet, when it comes to actual gaming, the result is totally different, it seems like the performance difference between those 7900XTX with only 2X8pin power connectors, which bascially are FE or some cheap ones, and 7900XTX with 550W power consumption is very small. At the end, he gave a conclusion, over clocking 7900XTX for better gaming performance is totally meanless yet he did not know the reason. This conclusion also confused me, I mean, I do understand that both AMD and Nv, even Intel would like to "optimize" their driver to get better performance in 3D Mark, yet the problem is those 7900XTXs that are overclocked to 550W power consumption is also crashed those normal 7900XTX in 3D mark time spy, if there is driver "optimizing", then both of cards should be "optimized" at the same level. How can the performance difference in gaming become so small? Can any one give me answer?

 

[CountMike]

Hi guys, I am a Chinese- Australian. Recently I had watched a review of all brands' 7900XTX in China for one vloger. If you do understand Chinese, here is the link
https://space.bilibili.com/434570236/channel/collectiondetail?sid=1727031
However, there is one test result that makes me feel so strange. According to his test, there is one card from asrock, which is AMD Radeon™ RX 7900 XTX AQUA 24GB OC, that has 550W max power consumption, with stock split water cooling cooler. When it comes to 3D mark test, this very card hits 33000 on 3D Mark time spy, which hits RTX4090 level and pales the rest of 7900XTX. Later, he did the test again, by install Asrock AMD Radeon™ RX 7900 XTX AQUA 24GB OC's bios into other brands' 7900XTX and put his Aio cooler (his primary job is selling customized AIO cooler for graphic cards), manually over clocked card, the result is almost the same. Yet, when it comes to actual gaming, the result is totally different, it seems like the performance difference between those 7900XTX with only 2X8pin power connectors, which bascially are FE or some cheap ones, and 7900XTX with 550W power consumption is very small. At the end, he gave a conclusion, over clocking 7900XTX for better gaming performance is totally meanless yet he did not know the reason. This conclusion also confused me, I mean, I do understand that both AMD and Nv, even Intel would like to "optimize" their driver to get better performance in 3D Mark, yet the problem is those 7900XTXs that are overclocked to 550W power consumption is also crashed those normal 7900XTX in 3D mark time spy, if there is driver "optimizing", then both of cards should be "optimized" at the same level. How can the performance difference in gaming become so small? Can any one give me answer?

XTX denotes it's already set for peak performance with best binned parts, pushing over it just invites troubles with little to no gain.

 

[FAhentai]

XTX denotes it's already set for peak performance with best binned parts, pushing over it just invites troubles with little to no gain.

I do understand that 7900XTXs are "set for peak performance with best binned parts". Yet, the only thing I can not understand is that by 3D Mark Time Spy, or even some thoery performance test, "normal" 7900XTX VS OCed one can still make such huge difference, while it can not reflect to gaming.

 

[CountMike]

I do understand that 7900XTXs are "set for peak performance with best binned parts". Yet, the only thing I can not understand is that by 3D Mark Time Spy, or even some thoery performance test, "normal" 7900XTX VS OCed one can still make such huge difference, while it can not reflect to gaming.

Benchmarks are not definite proof of performance, just a faint baseline made in some controlled environment. GPUs are just one part of equation and other parts are also important but also variable. Just raw numbers are not necessarily transferable to SW and games.

 

[DSzymborski]

I do understand that 7900XTXs are "set for peak performance with best binned parts". Yet, the only thing I can not understand is that by 3D Mark Time Spy, or even some thoery performance test, "normal" 7900XTX VS OCed one can still make such huge difference, while it can not reflect to gaming.

Because synthetic benchmarks and practice are very different things. Benchmarks are *designed* to scale as perfectly as possible, but that's a secondary thing for games.

 

[boju]

Because synthetic benchmarks and practice are very different things. Benchmarks are *designed* to scale as perfectly as possible, but that's a secondary thing for games.

Exactly. There's more to games than just slapping a few scenes together. That's simplifying things but close enough. It's a huge task for developers to optimise games and shouldn't be compared to synthetic benchmark like performance. There's also no user input, mouse keyboard or network for system to consider either.

Benchmarks are a reliable way to compare products in a perfect world kind of sense, even playing field, that's about it.

 

[FAhentai]

Exactly. There's more to games than just slapping a few scenes together. That's simplifying things but close enough. It's a huge task for developers to optimise games and shouldn't be compared to synthetic benchmark like performance. There's also no user input, mouse keyboard or network for system to consider either.

Benchmarks are a reliable way to compare products in a perfect world kind of sense, even playing field, that's about it.

In another words, can we understand it in this way: 7900XTX in fact still has large potential to dig. Its driver still not able to full release its power. However, if we are going to see that day or not, we will never know.

Btw, till now I still do not understand the true reason behind it, and I also fail to understand why over clock needs software, especially driver to support it, it is not like unlock the locked cuda core/cu, that you need driver to let your system knows that this card, comparing with normal cards, has more Cuda/Cu, please use them. Here is very ancient example cuz last time I did over clocking is almost 10 years before, back to old days, there were some GTX560TI, which had 384 cuda core, and 1Ghz core speed (FE and stock speed 800). At the same time, there was GTX570, which has 480 cuda cores and stock speed at 732MHZ. Of course, those 560TI is a little bit faster, since 384 Cuda work at 1G is faster than 480 cuda work at 732MHZ already. It also reflects to benchmarks as well as normal gaming.

Now the only possible reason for 7900XTX goes like this that I can think of is its vram speed, or shall we say, its bandwidth, is not enough to support overclocked GPU, eventually the bottleneck of whole system is the Vram. Back to the old days, it is very often, some time even Nvdia/ATI(It is AMD now) makes this kind of bottleneck on purpose so that they can position different cards to different levels, Eg, NV TNT2 has Ultra, PRO/Normal version, M64 and Vanta, Gf256 has Sdram verion Vs DDR SDRAM version, while ATI, has Radeon 9500 Pro with the same GPU specs as Radeon 9700, while it uses 128bit SDRAM comparing with Radeon 9700's 256 bit. Thus, back to old days, overclock Vram is even more effective comparing with overclock GPU. However, in recent years, Vram iterates way too fast, especially after DDR2, GDDR and DDR became totally different things. The Ram bandwidth is no longer a problem and over clocking Vram became meanless. In fact, in recent years there is only one card I can tell that its Vram become its performance bottleneck, which is RTX 1660TI cuz there is RTX 1660 SUPER, and its GPU specs is even lower than 1660TI yet it has higher Vram bandwidth, in the end, they both performe same.

However, I don't have 7900XTX, if any one can do a test based on this, I will be very appriciate.

 

[alceryes]

Without excellent watercooling, overclocking the 7900XTX won't get you more than a 5-10% boost in the graphics score in Time Spy. As others have said though, this graphics score increase does NOT translate to 5-10% higher FPS. Game operation is much more nuanced than GPU benchmarks. Also, remember that you would only see an improvement in FPS if your GPU is the actual bottleneck. Your actual bottleneck could be CPU subsystem speed, memory speed, drive speed, operating system issues, a combination of these, etc.

Even with excellent watercooling you'll probably max out at a ~15% higher score AND have a card whose error-free life is probably cut in half.