Gadgetory


All Cool Mind-blowing Gadgets You Love in One Place

Is the GTX 1080 bandwidth starved?

2016-12-27
hey guys welcome back to harbor unboxed for today's video we'll be looking at the performance of Nvidia's GeForce GTX 1080 graphics card or more specifically how memory frequency impacts performance actually the card will be using for testing will be gigabytes really heavy Extreme Edition card now when these cards were first announced or this GPU rather was first announced many were concerned that the GPU would be limited by the use of gddr5 X memory rather than HP m memory also compounding the issue is the fact that we were only seeing a sort of mid-range if you will 256-bit memory bus being used but once the cards landed they pretty much delivered mind-blowing performance in the latest games and yet many have still been suggesting that their memory bandwidth limited the concern seemed reasonable enough after all the previous generation gtx 980ti sported a 384 bit wide memory bus and when coupled with 7 gigabits per second gddr5 memory that resulted in a peak bandwidth of 336 gigabytes per second the GTX 1080 on the other hand is limited to a bandwidth of 320 gigabytes per second despite using faster gddr5 X memory and the reason for this being that it uses a narrower 256-bit wide memory bus given that the gtx 1080 boasts vastly superior shader power to that of the 980ti you have to wonder how a reduction in memory bandwidth might impact performance and this is why many were suggesting a memory bottleneck could be an issue those figures alone would suggest that it very much is however there are a few tricks in videos latest Pascal architecture has up its sleeve being the more refined architecture Pascal boasts a number of improvements over Maxwell that helps with memory bandwidth for starters each SM unit features half as many cuda cores and this is important because each cluster still has the same 64 kilobytes of shared memory essentially this means each core has 50% more memory at its disposal you'll also find a larger level 1 cache along with a much larger level 2 cache then as I'm sure many of you know as we did touch on it in our launch date coverage Pascal features a more advanced ansed color compression algorithm which helps reduce memory bandwidth demands even further then getting back to the gtx 980ti that particular GPU wasn't memory bandwidth limited at all in fact down clocking the memory frequency showed that it has balance to spare and of course after all nvidia isn't going to waste money developing a complex GPU only to hinder its performance with a memory bandwidth bottleneck so you can safely assume that the GTX 1080 isn't memory bandwidth starved but we thought why not take a look anyway it's also worth keeping in mind that although nvidia spec sees the 1080p with gddr5 X memory capable of 10 gigabits per second it has been proven time and time again that this same memory has no issue providing a throughput of 11 gigabits per second and some cards even reach 12 gigabits per second so it's fair to say that the memory has been clocked a very conservatively had the 1080 required more bandwidth you can bet Nvidia would have pushed the memory harder increasing the gddr5 X memory transfer speed to 11 gigabits per second increases the 10 eighties bandwidth by 10% to 352 gigabytes per second and that's a decent boost anyway for a fun little test I've taken the gigabyte GT x 1080 extra motion graphics card which comes with the core overclocked by almost 10% out of the box the memory has also been factory overclocked slightly above the Nvidia reference specification but that's not an issue for this test as we've manually set the memory frequency for each test for testing our overclocked core i7 6700k test machine has been used and as always the latest display drivers available at time of testing have been used I figured it best to check out performance and not just for okay but also the 1440p resolution in total just three games have been used for testing but they are newly released in very demanding titles so they paint a pretty clear picture with that let's jump to the benchmarks first let's see how the gtx 1080 memory scales in Gears of War 4 out of the box the gigabyte extremely Asian card was good for an average of 99 FPS with a minimum of 82 FPS at 1440p reducing the memory throughput by 8% drop the average frame rate by just 2% and the minimum by even less overclocking the memory boosted the average frame rate by 2% and the minimum by almost 4% that's not bad that we found exceeding 11 point 2 gigabits per second didn't improve further increasing the resolution of 4k we average 49 fps at the stock memory speed with a minimum of 41 fps reducing the memory bandwidth by 8% so a 4% reduction in performance or 5% if we looked at a minimum frame rate increasing the bandwidth by 12% to 11 point 2 gigabits per second didn't improve the average frame rate and an early boost of the minimum by a single frame Deus Ex mankind divided has some slightly different results for us out of the box the gigabyte Extreme Edition card was good for 62 FPS on average with a 52 FPS minimum overclocking the memory fair had absolutely no impact on anything a bit like a fanboy arguing in the comment section however when we take a bit of wind out of the 1080 sales by reducing the memory throughput to 9.2 gigabits per second the performance drops by 8% for the average and 10% for the minimum this does suggest that the 1080 is right on the limit in terms of bandwidth even at 4k we see no extra performance gained when overclocking the memory the same 32 FPS on average is rendered however reducing the memory throughput by 8% impacted performance quite heavily here the average frame rate dropped by 6% and the minimum by 11% interestingly the performance he was even greater than the reduction in bandwidth as it seems the memory bottleneck really hurts efficiency like mankind divided we found that there is absolutely no performance gains to be had when I book locking the gddr5 X memory further at least without also overclocking the CUDA cores the same 107 FPS average can be seen when running the memory at ten eleven point two and eleven point eight gigabits per second the minimum frame rate did jump up by a single frame but that's almost not worth talking about what is worth mentioning is the 5% reduction in performance seeing when under clocking the memory granted it isn't that significant but we do find a tipping point where the GTX 1080 becomes bandwidth limited even at 4k we see no more than a single extra frame being rendered with the memory of a clock under clocking the memory reduces performance by 5% on average and this isn't a significant drop-off but it does prove that even a factory overclocked GT X 1080 graphics card such as the Extreme Edition version we used is pretty much getting the most out of the GPU with the stock gddr5 X memory well there you have it I did originally intend to show you guys some side-by-side gameplay footage but with just a few frames the difference it didn't really make for an exciting compare for those of you wondering why I didn't wander the gddr5 X memory speed down even lower the reason for that being this was the lowest frequency that MSI's afterburner software would allow it may have been possible to create a custom BIOS to go over the lower or do something along those lines but for now I think the 9.2 gigabits per second data confirms what we set out to find even when running a heavily factory overclocked GTX 1080 such as the extreme gaming model that I used the standard gddr5 X memory specification is ample for the most part under clocking the memory by 8% resulted in a performance loss for all three games tested for the most part around 5% of the original performance was dropped though at times we did see up to an 11 percent drop in short it looks as though the GTX 1080 is right on the edge when it comes to memory bandwidth lowering the memory frequency just a bit you do start to see the performance fall away going the other way though unless you've overclocked the cuda cores increasing the bandwidth or rather the memory frequency won't result in any other performance this makes the upcoming GTX 10 atti rumor that the card might not even feature gddr5 X memory quite interesting personally I believe it will have to even with the 384 bit wide memory bus of the Titan XP the 10 ATT I would only have 20% more bandwidth than the GTX 1080 if it were to use gddr5 at gigabits per second memory given that the TI model is meant to offer 30% more criticals it's going to need a little extra bandwidth anyway enough speculation I think I'll just wait for the 1080 Ti to come knocking on my door before I read too much further into that one and hopefully it does so shortly after Vega hits our test pad anyway I hope you guys found the testing interesting I know it wasn't super in-depth but I think the test results were fairly conclusive anyway let me know what you guys think of the testing and if you have any ideas for similar kinds of tests let me know anyway I'm your host Steve thanks for watching you
We are a participant in the Amazon Services LLC Associates Program, an affiliate advertising program designed to provide a means for us to earn fees by linking to Amazon.com and affiliated sites.