Gadgetory

this is an exciting milestone for us we've completely overhauled our CPU testing methodology for 2019 and this is what we'll be using for the rest of the year new testing includes more games than before tested at two resolutions alongside workstation benchmarks these are new for us we've added a couple of workstation workloads that we hadn't looked at before like program compile we've added Adobe Premiere Adobe Photoshop compression and decompression v-ray and more today is the unveiling of the workstation half of our new testing methodology with the games getting unveiled separately we're starting with a small list of popular CPUs and we'll add more as we go before that this video is brought to you by the gigabyte ARS ad 27 QT gaming monitor the ad 27 QT is a 27-inch 1440p gaming display with 95% of dcpip 3 color saturation for high color accuracy accompanied by a 1 millisecond response time 10 bit IPS panel and display HDR visa certification additional features include fluid adjustment and slide RGB LEDs for personal Flair and firmware features like cooldown counters at reticles and adaptive and noise reduction learn more at the link below so this is just a preview of something we've been working on for a couple of months now just building out the tests validating them making sure we're happy with how the performance looks and happy with the different aspects of the CPU that the tests represent typically we do a couple of games and then we do maybe one or two workstation benchmarks like blender for example software that we actually use and is used widely for 3d animation 3d rendering and 3d art so that's been our staple workstation benchmark but there's a lot more to workstation performance than blender which is primarily representative of the high thread-count CPUs and doesn't really look at other aspects of the processor this set of charts is not a hundred percent complete yet so it's it's an unveil of the new methodology we'll have more information in the article in the description below if you want to read the full testing methods that will describe them there and what we're going through today is the workstation stuff so that'll be blender the anew compiler compress collection for those of you who've wanted programming or compiling benchmark from us we now have a GCC for that we have v-ray we have Adobe Premiere Adobe Photoshop and then 7-zip for compression and decompression for games going forward not in this video but in a future one coming up soon we have hitman 2 which is new for us we have Assassin's Creed origins one of our staples from previously GTA 5 total war Warhammer to the battle mode total war Warhammer to campaign mode civilization 6 tested just at one resolution because the turn time benchmark Shadow the Tomb Raider and an f1 2018 which leaves us with 15 total test combinations for just games alone because it's 1080p 1440p across all those games and then six is just 1080p so that's our new benchmark suite it is very comprehensive we're pretty happy with the amount of data we end up with but we think that it's still an actually usable amount it's not overwhelming and it doesn't just become a huge mess of charts but what we're going to have to do is kind of split them because otherwise it will be a huge mess of charts so work stations what we're looking at today we don't have the HTTP Intel CPUs benchmarks or revenge marks yet you can still look at our old tests like the 3175 x to see how some of those did but today we're starting with popular cpus and then we also threw and thread refer just because we had one around for other testing so let's get into it we'll look at work station results you can check the article below for more information and then we'll talk conclusions after all the benchmarks we had a lot of requests to add compiler benchmarks to our test suite so we finally did for this we've added the GCC benchmark which involves compiling the ganoush Empire collection version 8 point 2.0 with GCC version 7 point 4.0 we set a flag to allow it to spawn as many threads as possible so the compilation process involves both some single threaded and heavily multi-threaded workloads interestingly the results for this test are the inverse of many of the gaming tests that we'll publish next with all the AMD chips on the table and all the Intel chips at the bottom the 2990 WX with Corp Ryo enabled logged the fastest time and an 11% time reduction versus the stock 2990 WX the AMD chips below this are ordered predictably the 4.2 gigahertz 16 thread 2700 is at top allowing the 2990 wxa time reduction of about half with the r5 2600 at 4.2 gigahertz following the 2700 next this stack shows that frequency still matters as it outpaces the 2700 ex stock CPU marginally the are 727 100 X is 8 point 7 minute compile-time leads the 9900 case stock CPU by 22% a meaningful improvement overall the 9900 K leads the Intel CPUs and again the order is predictable other than the AMD Intel divide the CPS are logically ordered by the highest frequency and core count parts with the stock 8600 K at the very bottom frequency and core count are important factors but they aren't everything and this is one real-world workload that shows it cache for instance can also come into play in this type of benchmark note that other compilers may behave differently and also that Lincoln is a factor worthy of consideration if using linkers that are typically single threaded this can pose a bottleneck we are also doing all of our testing and Windows and so Linux workloads would further exhibit deviation from what we're seeing here 7-zip is next 7-zip includes a built-in benchmark that can generate scores for both compression and decompression measured in millions of instructions per second interestingly decompression appears to be more thread dependent than compression the 9900 K leads in compression by a good margin with 21.4 present more instructions per second than the 9700 K when both are overclocked this shows that threads that still do affect the score even though this chart lines up differently than the decompression chart we'll look at next the worst performer in this category is the stock I 580 600 k with only 6 cores and 6 threats but the 32 core 64 thread 2990 WX only lands in the middle of the charge we seem to be constrained to some extent by single threaded performance in decompression however threader parterres ahead of everything else was 115 percent more instructions per second than the overclocked 9900 k the stock AMD r7 2,700 x and overclocked r7 2700 actually managed to beat out these stock 9900 K although overclock into 5.2 gigahertz does allow it to regain the lead at the expense of power the 2700 out performs at the overclock 290 700 K at 5.1 gigahertz from the 2700 is stock illustrating a limitation in the i7 thread count as the r7 gains a 5% lead even the r5 2600 at 4.2 gigahertz it nearly keeps up with least at 9700 K we're next using Puget systems at Photoshop benchmark it recommends 32 gigabytes of system memory which is the main reason we now use 4 8 gigabyte sticks of RAM for all of our tests rather than just 4 CPUs that can run memory in quad-channel we run the extended version of this benchmark which produces seven scores as well as an overall score to summarize them thread count isn't much help here and the 5.2 gigahertz at 9900 K it tops the chart once again the chart is roughly ordered by CPU frequency with the 2990 WX scoring slightly worse than the stock r5 2600 even with game mode or Corp Ryo the ordering up by frequency is further illustrated by the 80 600 K at 5 gigahertz outperforming a stock 8700 K or the 9700 K at 5.1 gigahertz outperforming the stock at 9900 K Photoshop appears to be frequency bound in these tests which includes the application of various filters transforms resizes photo merges and more and if you're curious about the intel h EDT cpus we first started using this benchmark back in the 7900 era including most recently with the 3175 ax blunder benchmarking has also changed with this round of GM's updated test methodology we're keeping our in-house made benchmark scenes using realistic workloads with things like ray tracing and movie ready render settings also realistic effect but we've eliminated blender 2.78 2.78 has grown long in the tooth so we've struck out that monkey had rendered test and the temperamental 2.79 splash render as well as been removed leaving only the blender 2.79 monkey head render and the GN logo renders the monkey heads produce a varied work load on the CPU exclusively by using different types of textures transparencies and elements within the scene while the GM logo how is the CPU more explicitly with ray tracing and very very high sample count for the final render it's typically the test that causes unstable overclocks to reveal themselves as unstable blender is an important test to us because we actually use it and it directly benefits us to know which CPUs handle it best let's start with the monkeyhead render the 2990 WX takes this test handily with its 64 threads and the rest of the stack lines up in a similar fashion to the other thread bound workloads with the 9700 K being the only CPU breaking the more cores higher score pattern the 2990 WX finished testing in ten point nine minutes stock roughly the same with Corp Rio and demonstrates clear value to professionals who work in tile based rendering applications like blender and cycles the biggest value add is when high system memory requirements exist as this can rapidly exit the confines of GPU memory allotment thus limiting usefulness of CUDA rendering we still need to retest our h EDT Intel CPUs so thread Ripper remains relatively isolated for now the 9900 K at stock completes its render in twenty point six minutes a time reduction of 11% from the stock r7 2,700 ax is 23 minute render time the 2,700 X comes close given the price difference and that's because blender tends to favor thread count since it spawns one tile per thread for the G M logo render the TR 2990 W X obviously still leads and we'll wait until further updates to the chart before it has any company the 9900 k stock CPU roughly equates an overclocked 2700 at 4.2 gigahertz although outperforms the stock 2700 X with a 6% render time reduction the stock 2700 and stock 9700 k also end up roughly tied with the 8700 k approaching both of these CPUs and performance premier is a test that we've done in the past but we haven't kept up with it faithfully due to the amount of manual setup required this is something that we've resolved going forward our last Adobe premier test was with our old test methodology where we showed the Intel 28 court 31 75 X proven genuinely good when overclocked it drew a hell of a lot of power but it also managed the fastest render times even outperforming the 99 ATX II and the high speed 9900 K we haven't retested with our new test suite just yet but we do have more mainstream CPUs test for today we've streamlined at this process and now we're rendering three videos one at 1080p60 it's a convention show floor report one at 4k 60 with a roll and b-roll and one 4k video consisting entirely of charts h.264 is our codec for these for now with a high profile and 35 megabit per second output starting with the 1080p 60 a roll and b-roll footage from a convention where we use clips entirely off of our Panasonic UX 180 and recorded through our zoom audio devices we get the chart on the screen now we haven't yet put the h EDT cpus through this one so the 99 100 K is 3.9 minute render is at the top render times reduced by nearly 8 percent with an overclocked to 5.1 gigahertz the 9700 K at 5.1 gigahertz shows that despite being at frequency locked with our chart-topping 9900 K the extra threads are actually beginning to help more as Adobe Premiere Pro CC 2019 gets updates but I knew I had heard K a 5.1 gigahertz holds a lead of 15 percent over the 9700 k advantaged by its double thread count and these are 7 2700 at 4.2 gigahertz and the 2700 X finished the rendering about 4.4 to 4.5 five minutes and in roughly equal to the stock 9700 K despite increasing thread utilization premier still likes higher frequencies this much is proven by comparing the 2700 stock result of five point two minutes to the OC result where we see a reduction of 13 percent versus baseline the r5 2600 at 4.2 gigahertz does exceptionally well when considering its more budget oriented positioning although professionals who use premiere every single day would still want to consider a higher-end option toward the top of the chart or something we haven't yet tested our 4k 68 roll and b-roll render was more intensive clearly as it's dealing with a higher quality output and more pixels the 9900 KOC finishes this rendered in 10.8 minutes which is a render time reduction of 10% from baseline this stack is almost exactly the same for these tests except B 2700 stock and the 8700 K reshuffle but they're within error margins of each other in both tests the difference between the 2700 X and a 900 K both stock is that the 9900 K finished as a trend in 13% last time the chart render in particular seems to be a lighter workload on the CPU and we'll be scrapping it moving forward this is an illustration of a workload that isn't as heavily reliant upon the CPU how do we see CUDA kicking in to help more frequently with our chart renders that be what you're seeing on the screen now it's it's a chart render because that's well that's the idea of what we're doing for the 4k 60 chart render test it's a lot of this but anyway because of how scattered these results are especially with how unreliable the differences are and how some of these just straight up don't make a lot of sense we have to scrap it going forward because the test data is not reliable and because we're exiting cpu-bound scenarios and entering scenarios where other parts of the system are taking more control chaos groups v-ray benchmark includes a GPU and a CPU test but we use the CPU test only for this benchmark it takes between one and two minutes to complete on most of our CPUs and the results align closely with something like Cinebench or blender for rendering benchmarks with the higher thread count CPUs like the 2990 WX and the 9900 K heavily advantaged leaving the stock 8600 K and that lasts until we retest our h EDT Intel CPUs the 9900 K at five point two gigahertz is Intel's highest results on our chart this one completes the workload in 0.95 minutes allowing the 2990 WX a time reduction of about 52 percent demonstrating that the array actually leverages the threads available to it although the 9900 K does well to keep up with an h EDT part extra threads do went out in this test ultimately the 2700 at 4.2 gigahertz outperforms the overclocked at 9700 K demonstrating again that there's a thread advantage but it's barely beaten by the 9900 case stock CPU price is a factor granted and the 2700 comes in about 265 dollars cheaper the r5 2600 just for reference at 4.2 gigahertz lands at 1.5 three minutes time completion putting it within range of the 8700 K a stock CPU so that's it for our new workstation tasks we technically also have Cinebench just for internal validation we we don't really publish those numbers because we have blender for an actual workload for that but we still take it just to make sure the CPUs are where they should be for internal validation of performance we have thai-inspired that we didn't show in this video perhaps in the future though and then the really interesting stuff is GCC v-ray stuff like that if you have specific requests of things that you would still like to see leave them below there's no guarantee we'll add them into this test suite but we do actively take suggestions these tests were advanced because of suggestions from our audience so if you want to be considered even if it's in six months from now for additional testing leave your idea below and maybe a brief reason of why you would like to see that even if it's just look I'm a professional in this space and we use this application we'll certainly consider it so there's obviously a lot more that can be done but this is the most comprehensive suite we've ever had for CPU testing and we are greatly looking forward to continuing to use this suite advancing it and figuring out what new things it teaches us about processor performance so for now you get a recap of the 99 hundred K of the 2700 series X and non X the overclocked performance for some of those and try to represent a couple of those charts as well and then we'll we'll continue advancing as rise and 3000 series comes out and up until then so let us know what you think of the suite it's fairly comprehensive but again open to suggestions subscribe for more go to store documents nexus dotnet to support us directly for this type of multi month effort and you can pick up something like for example our GN teardown cubes which have a 3d laser engraved logo GM logo with things like MOSFETs capacitors inductors and a fake V RMS probably can't turn that into a real one but you could try subscribe for more I'll see you all next time