able to complete multiple runs of 3DMark Timespy at 2075MHz . Pulling the temperature down to the 5 ’ C or even 0 ’ C didn ’ t help much other than allow for the card to start forming water at the back ( at this point I hadn ’ t insulated properly because I literally was testing the GPU silicon limits ).
The memory ’ s behavior didn ’ t change at all with the cooling of the GPU . Adding + 650MHz or even + 700MHz to the memory clock was possible , however it turns out that if you overclock the memory too high , you will lose performance in the benchmarks . You won ’ t spot any artifacts in the benchmark run usually , but the numbers will be much lower and as such you ’ re better off with the lower frequencies . I ’ ve read and heard that there ’ s a possible way around this GDDR5X behavior , but I ’ ve yet to find a concrete solution to it . If you ’ re wondering if it happens on other regular GTX 1080 Ti graphics cards ,
yes it does , the difference is that you might get screen artifacts even at + 600MH on the memory clock . Eventually I settled at + 550MHz for 1539MHz SDR ( 6156MHz QDR ) for around 541GB / s of memory bandwidth . Sounds alright , but when you push the GPU core farther than even 2,250MHz you will run into this memory bandwidth limit .
That aside , it is something that I should be able to work around in the second round of LN2 testing , but my primary focus was the core .
It turns out that even besides having all the right software and hardware for LN2 overclocking and somewhat basic knowledge of what should be done . Much like with some 980 Ti cards , the 1080 Ti requires a patient approach to how you achieve your clock speeds . I found that going straight for a high clock frequency would often lead to crashes . That is after you boot into the operating system , shooting straight for 2,400 MHz on the core for example would lead to the driver crashing . A smarter way is to slowly raise the clock speeds , so let ’ s for example assume that you set 1.3V for the GPU core . Obviously the real voltage will be different , so it ’ s important to have a DMM handy for reading the actual voltage ( more on this later ). This voltage will work up to a certain point which much be matched with the right temperature . At -60 ’ C you ’ ll not get the expected clock frequencies , but at the same time -140 ’ C may be too cold for the GPU . The trick is figuring out the right balance for that particular sample you ’ re dealing with . There are general rules to guide you , but the specifics will vary with each card , much like it does with CPUs .
Some GPUs you can literally go “ full pot ” on and still get high GPU clock frequencies ( although this may have detrimental effects on your memory overclocking ), but for the most part around -125 ’ C is where you ’ re going to be .
Issue 42 | 2017 The OverClocker 29