New CPU Test Methodology: Code Compile, Gaming, Transcoding, & More (2020)
video description
Date: 2020-05-06
Related videos
Comments and reviews: 10
Royameadow
As somebody who has only ever used Entry to Mid Tier Desktop hardware all the way up to now, I do feel that there are some extra tests to consider for future workloads, which can help determine which is best for more unique types of setups: Gaming (Emulation): BSNES HD, RPCS 03. Video Compression: XMedia Recode. Video Editing and Rendering: MAGIX Vegas Pro (Version 0I6 (Build 0352) and any relevant recent build of Version 0I7. BNES HD is a fork of BSNES and its successor Higan that focuses on High Definition rendering for Mode 07 and a Widescreen Hack of up to 2I: 09 for Super Famicom titles; it is CPU and GPU intensive as you go up in resolution and the scaling of HD Mode 07 and enhancements such as 0I0x Supersampling, which is noticeable when testing F Zero, Star Fox, Super Mario: Yossy Island (Yoshi's Island, and Star Fox II on an Intel 8400/AMD RX Lexa (02 GB) combo. RPCS 03, the PlayStation 03 emulator, is notorious for framerate variances and stutters that are based on your hardware combo; NBA Elite 'II, the rarest game on the console, is the best title to test on RPCS 03 due to its high demand for Core Frequency and VRAM to have a genuinely stable experience, as the emulator will crash on lower end parts with this game in particular and thus serves as one of the best titles to experiment with on the lower and high end alike to see what makes it tick. XMedia Recode has been used by my team for improved file sizes for multiple programs before we upload or broadcast them and also for faster data transfer to assure that they can be sent to somebody online or copied to a Flash Drive at a reasonable pace; XMedia Recode benefits best from a decent CPU that has the Core Count and Frequency while also likely relying in specific codecs to be able to convert the footage faster in Realtime Priority, architectural optimizations may also play a factor and it is something that I would love to see tested with Ryzen CPUs, due to only having used an Intel 4160 and 8400 with this workload up to now. Vegas Pro is already a part of Gamers Nexus' testing methodology, but specifically only for the annual Best Workstation GPU for (Insert Year Here) program done by Rob Williams of Techgage: Vegas is an interesting beast, as like Adobe Premiere, it can do CPU Only Renders up to 032 physical cores (not sure on the thread count, though 048 is the projected limit in the Preferences Menu) and CPU+GPU Renders cap out at 0I6 cores (which is where the 9980XE, 10980XE, and 3950X reign supreme over the Zen II Threadrippers); the other big thing about Vegas Pro is how well it favours AMD's cards over that of Nvidia due to Vega to the 064th Power and Radeon VII's OpenCL capabilities superseding even the Titan RTX in this workload, which is mostly due to Nvidia's OpenCL capabilities being significantly worse and not built for a Compute Heavy workload such as this in particular, which I don't know if it will carry over to their next generation of products, not to mention the Navi cards' lack of official support in Vegas up to now (though we will see fixes to this on both the RDNA II and CDNA lineup, once MAGIX can handle the matter; I feel bad for anybody who got a 5700 (XT) at launch, only to find out that it didn't work in Vegas at the time, it's the perfect testing ground for bizarre results and honestly scales interestingly from Standard HD (0720p) to 08K Resolution (4320p.
reply
As somebody who has only ever used Entry to Mid Tier Desktop hardware all the way up to now, I do feel that there are some extra tests to consider for future workloads, which can help determine which is best for more unique types of setups: Gaming (Emulation): BSNES HD, RPCS 03. Video Compression: XMedia Recode. Video Editing and Rendering: MAGIX Vegas Pro (Version 0I6 (Build 0352) and any relevant recent build of Version 0I7. BNES HD is a fork of BSNES and its successor Higan that focuses on High Definition rendering for Mode 07 and a Widescreen Hack of up to 2I: 09 for Super Famicom titles; it is CPU and GPU intensive as you go up in resolution and the scaling of HD Mode 07 and enhancements such as 0I0x Supersampling, which is noticeable when testing F Zero, Star Fox, Super Mario: Yossy Island (Yoshi's Island, and Star Fox II on an Intel 8400/AMD RX Lexa (02 GB) combo. RPCS 03, the PlayStation 03 emulator, is notorious for framerate variances and stutters that are based on your hardware combo; NBA Elite 'II, the rarest game on the console, is the best title to test on RPCS 03 due to its high demand for Core Frequency and VRAM to have a genuinely stable experience, as the emulator will crash on lower end parts with this game in particular and thus serves as one of the best titles to experiment with on the lower and high end alike to see what makes it tick. XMedia Recode has been used by my team for improved file sizes for multiple programs before we upload or broadcast them and also for faster data transfer to assure that they can be sent to somebody online or copied to a Flash Drive at a reasonable pace; XMedia Recode benefits best from a decent CPU that has the Core Count and Frequency while also likely relying in specific codecs to be able to convert the footage faster in Realtime Priority, architectural optimizations may also play a factor and it is something that I would love to see tested with Ryzen CPUs, due to only having used an Intel 4160 and 8400 with this workload up to now. Vegas Pro is already a part of Gamers Nexus' testing methodology, but specifically only for the annual Best Workstation GPU for (Insert Year Here) program done by Rob Williams of Techgage: Vegas is an interesting beast, as like Adobe Premiere, it can do CPU Only Renders up to 032 physical cores (not sure on the thread count, though 048 is the projected limit in the Preferences Menu) and CPU+GPU Renders cap out at 0I6 cores (which is where the 9980XE, 10980XE, and 3950X reign supreme over the Zen II Threadrippers); the other big thing about Vegas Pro is how well it favours AMD's cards over that of Nvidia due to Vega to the 064th Power and Radeon VII's OpenCL capabilities superseding even the Titan RTX in this workload, which is mostly due to Nvidia's OpenCL capabilities being significantly worse and not built for a Compute Heavy workload such as this in particular, which I don't know if it will carry over to their next generation of products, not to mention the Navi cards' lack of official support in Vegas up to now (though we will see fixes to this on both the RDNA II and CDNA lineup, once MAGIX can handle the matter; I feel bad for anybody who got a 5700 (XT) at launch, only to find out that it didn't work in Vegas at the time, it's the perfect testing ground for bizarre results and honestly scales interestingly from Standard HD (0720p) to 08K Resolution (4320p.
reply
Bastard
I find the power testing VERY interesting and I'm looking forward to GN going down that path. Especially idle power draw, which is something i find that enthusiast gear is particular bad at for no good reason. Done a fair bit of fiddling with this myself over the years, and have noticed that on modern systems you can usually retain close to 99% performance while only using something like 80% of the power. This is just by tweaking voltages and finding out what various settings actually do when it comes to power vs performance, things like XMP usually does far more then just tweaking memory. And even more interesting is that idle power varies enormously with some seemingly unimportant settings, and can really make a difference to your room temperature and noise levels (less cooling required. I remember from back when THG was relevant that the enthusiast community had a hard time understanding that a CPU which takes more power can still be more efficient if it's also higher performing, in short a CPU that draws twice the power can get the job done in 1/3 of the time and thus be more efficient (mind-blown. gif.
reply
I find the power testing VERY interesting and I'm looking forward to GN going down that path. Especially idle power draw, which is something i find that enthusiast gear is particular bad at for no good reason. Done a fair bit of fiddling with this myself over the years, and have noticed that on modern systems you can usually retain close to 99% performance while only using something like 80% of the power. This is just by tweaking voltages and finding out what various settings actually do when it comes to power vs performance, things like XMP usually does far more then just tweaking memory. And even more interesting is that idle power varies enormously with some seemingly unimportant settings, and can really make a difference to your room temperature and noise levels (less cooling required. I remember from back when THG was relevant that the enthusiast community had a hard time understanding that a CPU which takes more power can still be more efficient if it's also higher performing, in short a CPU that draws twice the power can get the job done in 1/3 of the time and thus be more efficient (mind-blown. gif.
reply
Earthishome7569
if you are overclocking or setting overclocks for intel then why dont you tune in ryzen for best configs. the 3900x will single core pbo to 4. 7 pretty easy and using these settings on a 3950x will boost clock up to at least 4. 9 or higher. with a few simple settings adjusted. also not using the extra memory bandwidth on Ryzen at 3800 cl 14 with a good pbo clock set for gaming on your test bench would really show a smaller gap in gaming benchmarks and any single core benchmarks compared to the 9900k try these settings on your test suite with the 3900x and 3950x with testing on them I would love to see realistic results with AMD cpu's memory 3800 cl 14 14 14 14 pbo settings PPT-300 TDC-300 EDC-10 PBO scalar- 4x llc x3 offset +. 0250 -. 0500 then possible lower this to a negative offset but have to test the boost clocks to tune the system fully for the voltage offset. if gaming testing is reliant on highest cpu single core boost clocks these settings on your testing suite would yield way higher single core boost clocks. over the Auto settings with in ryzen.
reply
if you are overclocking or setting overclocks for intel then why dont you tune in ryzen for best configs. the 3900x will single core pbo to 4. 7 pretty easy and using these settings on a 3950x will boost clock up to at least 4. 9 or higher. with a few simple settings adjusted. also not using the extra memory bandwidth on Ryzen at 3800 cl 14 with a good pbo clock set for gaming on your test bench would really show a smaller gap in gaming benchmarks and any single core benchmarks compared to the 9900k try these settings on your test suite with the 3900x and 3950x with testing on them I would love to see realistic results with AMD cpu's memory 3800 cl 14 14 14 14 pbo settings PPT-300 TDC-300 EDC-10 PBO scalar- 4x llc x3 offset +. 0250 -. 0500 then possible lower this to a negative offset but have to test the boost clocks to tune the system fully for the voltage offset. if gaming testing is reliant on highest cpu single core boost clocks these settings on your testing suite would yield way higher single core boost clocks. over the Auto settings with in ryzen.
reply
Matt
I know you've mentioned this before, but given this is a CPU suite of benchmarks, I really wish you would add some Linux benchmarks, maybe linux kernel compile time, or any number of the CPU Phoronix benchmaeka. Especially if you're going to be testing all the way up to Workstation/HEDT workloads, where Linux is a rather big deal, and since Linux both outperforms Windows in CPU tasks generally AND with scaling up to very high core counts. There's not much point testing a 3990X on Windows when it can't hardly use all the cores, especially relative to Linux which actually scales remarkably well. And now that Linux gaming is actually becoming a thing, and Linux is already a huge Workstation/HEDT force and dominant in servers, it'd be nice to see some Linux benchmarking included, since you guys actually do a really good job. Hardware Unboxed did some Linux benchmarks which were very interesting.
reply
I know you've mentioned this before, but given this is a CPU suite of benchmarks, I really wish you would add some Linux benchmarks, maybe linux kernel compile time, or any number of the CPU Phoronix benchmaeka. Especially if you're going to be testing all the way up to Workstation/HEDT workloads, where Linux is a rather big deal, and since Linux both outperforms Windows in CPU tasks generally AND with scaling up to very high core counts. There's not much point testing a 3990X on Windows when it can't hardly use all the cores, especially relative to Linux which actually scales remarkably well. And now that Linux gaming is actually becoming a thing, and Linux is already a huge Workstation/HEDT force and dominant in servers, it'd be nice to see some Linux benchmarking included, since you guys actually do a really good job. Hardware Unboxed did some Linux benchmarks which were very interesting.
reply
Diego
The compile test is nice, but not really representative of how programmers spend their time. Briefly, the source code of a large project like chromium is composed of thousands of source files, each of which can be compiled separately. Compilers are mostly single threaded programs, so you only get to use multiple cores if you have many source files to compile. However, when a programmer is working on something they generally are only modifying a few source files at a time, and therefore they only need to compile a few source files to test their changes. In which case, the majority of the time probably is not even spent compiling, but doing tasks like checking dependencies and linking. The point of all of this is that it would be very useful if you also tested the time to compile and link a when only a small number of sources need to be recompiled.
reply
The compile test is nice, but not really representative of how programmers spend their time. Briefly, the source code of a large project like chromium is composed of thousands of source files, each of which can be compiled separately. Compilers are mostly single threaded programs, so you only get to use multiple cores if you have many source files to compile. However, when a programmer is working on something they generally are only modifying a few source files at a time, and therefore they only need to compile a few source files to test their changes. In which case, the majority of the time probably is not even spent compiling, but doing tasks like checking dependencies and linking. The point of all of this is that it would be very useful if you also tested the time to compile and link a when only a small number of sources need to be recompiled.
reply
Max
Remark your comment the low thread utilization of Handbrake for more than 16 cores, which are actually limitations from x264 and x265 encoders. For future CPUs with more and more cores, you could/should consider the SVT encoder series from Intel, which are open sources and integrated in FFmpeg. Their SVT-AV1, SVT-HEVC, SVT-VP9 encoders are very threading & AVX256 optimized, and also the first two are very production ready/in used. Phronix. com uses those encoders for Video-encoding testing for a pretty while already and the results are always amazing for higher core counts like Threadripper Gen 3. (BTW. If you already considered this option, it would be great for me to know why you discard this possibility: )
reply
Remark your comment the low thread utilization of Handbrake for more than 16 cores, which are actually limitations from x264 and x265 encoders. For future CPUs with more and more cores, you could/should consider the SVT encoder series from Intel, which are open sources and integrated in FFmpeg. Their SVT-AV1, SVT-HEVC, SVT-VP9 encoders are very threading & AVX256 optimized, and also the first two are very production ready/in used. Phronix. com uses those encoders for Video-encoding testing for a pretty while already and the results are always amazing for higher core counts like Threadripper Gen 3. (BTW. If you already considered this option, it would be great for me to know why you discard this possibility: )
reply
Narek
Very nice, but I don't see any Blackmagic DaVinci Resolve benchmarks. It's a shame since a lot of people have and are switching to it, including me. Also I would seriously advice you to use Blender 2. 83. It's in beta right now but will release very soon. And there's a special button in Blender 2. 83 called Adaptive Sampling, which provides MASSIVE performance gains in rendering. On my machine with CPU + GPU rendering of the BMW benchmark I went from 68 seconds to 40 seconds! More than 40% performance gains! The gains are also big with CPU only, but I don't remember right now by how much. I have a Ryzen 2700X OC 4. 1GHz, a GTX 1660 Ti and RAM 32GB 3GHz C16 Tile size in Blender is 32x32px.
reply
Very nice, but I don't see any Blackmagic DaVinci Resolve benchmarks. It's a shame since a lot of people have and are switching to it, including me. Also I would seriously advice you to use Blender 2. 83. It's in beta right now but will release very soon. And there's a special button in Blender 2. 83 called Adaptive Sampling, which provides MASSIVE performance gains in rendering. On my machine with CPU + GPU rendering of the BMW benchmark I went from 68 seconds to 40 seconds! More than 40% performance gains! The gains are also big with CPU only, but I don't remember right now by how much. I have a Ryzen 2700X OC 4. 1GHz, a GTX 1660 Ti and RAM 32GB 3GHz C16 Tile size in Blender is 32x32px.
reply
Jesse
For encoding tests I think it makes sense to start doing a multiple instance tests, especially for HEDT. Even for desktop it makes sense, Intel is moving to 10c20t in the mainstream, and I wouldn't be surprised if AMD moves beyond 16c32 thread sooner rather than later. 2 x instances for desktop 4 x instances for HEDT Doesn't seem out if the question. As the core wars continue, I think it's going to become more and more important to try and capture the ability to megatask. Adobe premiere type work plus video encoding seems like an organic situation to test.
reply
For encoding tests I think it makes sense to start doing a multiple instance tests, especially for HEDT. Even for desktop it makes sense, Intel is moving to 10c20t in the mainstream, and I wouldn't be surprised if AMD moves beyond 16c32 thread sooner rather than later. 2 x instances for desktop 4 x instances for HEDT Doesn't seem out if the question. As the core wars continue, I think it's going to become more and more important to try and capture the ability to megatask. Adobe premiere type work plus video encoding seems like an organic situation to test.
reply
pokegeaks2
9: 00 GamersNexus you say that the 9700K performs better than. AMD's processors with high core counts, but on that list the only AMD processors it beats are those with less cores, with the 3600 coming close because of its 12 threads. But the 9700K has 8 cores. Where do you start using the term High core count because I dont see 6c as being particularly high these days, given the commonality of 8c and the rise of 10, 12, 16 cores. How is this phrasing you've used kept accurate in this case?
reply
9: 00 GamersNexus you say that the 9700K performs better than. AMD's processors with high core counts, but on that list the only AMD processors it beats are those with less cores, with the 3600 coming close because of its 12 threads. But the 9700K has 8 cores. Where do you start using the term High core count because I dont see 6c as being particularly high these days, given the commonality of 8c and the rise of 10, 12, 16 cores. How is this phrasing you've used kept accurate in this case?
reply
Sergio
question. Does the R7 2700x still maintain (near) parity with the 3600 with the new methodology? I've only seen 7zip results for that. anything else or you'll be releasing more metrics later on? edit: ok did a quick look on the written article. you used the the 2700 non-X so I can assume that the X might see some increased performance? seems that with new methodology it might have better results vs the 3600 edit2: also, totally forgot. many thanks for this massive overhaul on methodology.
reply
question. Does the R7 2700x still maintain (near) parity with the 3600 with the new methodology? I've only seen 7zip results for that. anything else or you'll be releasing more metrics later on? edit: ok did a quick look on the written article. you used the the 2700 non-X so I can assume that the X might see some increased performance? seems that with new methodology it might have better results vs the 3600 edit2: also, totally forgot. many thanks for this massive overhaul on methodology.
reply
Add a review, comment
Other channel videos
