AMD Threadripper 3990X CPU Review Code Compile, Render, & Compression
video description
Date: 2020-05-06
Related videos
Comments and reviews: 10
Ignacio
It's a nice review, I own 2990wx a and a 3990x, and usually work with 2970wx and 3970x as well. and there is something in this review that you didn't mention. There is apps that scale better and can use the full potential of the 3990x like Vray and Blender, and then there is the ones that won't scale that nicely because there is just SSD bottleneck or any other bottleneck due to software like Premiere and so. BUT: I think here there is another very important thing to mention. There is times where the bottleneck is the windows Scheduler. ( yeah I know you will do a Linux review ) but i'm talking about windows to windows here ( and yes I know they kinda patched windows pro, and also the statements of AMD about it ) but windows pro separate the threads in two groups in some applications we are noticing the CPU just uses at 100% power only, and exactly, half of the core counts. Yes this would be a software bottleneck but it's the SO so I think is worth mention, Looks like windows pro separate the cores in 2 groups and that the scheduler fails terribly into send jobs to both of them ( unless the software is properly set to do so like Vray or Blender) Looks like windows pro Enterprise/Workstation doesn't separate the threads in two groups so this specifics apps that are not programmed exactly to use all the cores, and that rely on the windows schedulers could get benefit from this. again, I can stress my CPU max power but just on half of it. If i could get the other half I would double the power. For instance I can open another instance of the same program, run it and will utilise the other half of the cores available and works like charm also. But simetimes you want more power in one instance of the program, not to run 2 instance of the program I think this would be worth testing cos that means 1, the scheduler of windows pro is shit. 2 if you own a 3990x and this enterprise version just can run soft in all cares and not like the windows pro then is worth the upgrade to windows enterprise/workstation. Sadly. very little information about this topic.
reply
It's a nice review, I own 2990wx a and a 3990x, and usually work with 2970wx and 3970x as well. and there is something in this review that you didn't mention. There is apps that scale better and can use the full potential of the 3990x like Vray and Blender, and then there is the ones that won't scale that nicely because there is just SSD bottleneck or any other bottleneck due to software like Premiere and so. BUT: I think here there is another very important thing to mention. There is times where the bottleneck is the windows Scheduler. ( yeah I know you will do a Linux review ) but i'm talking about windows to windows here ( and yes I know they kinda patched windows pro, and also the statements of AMD about it ) but windows pro separate the threads in two groups in some applications we are noticing the CPU just uses at 100% power only, and exactly, half of the core counts. Yes this would be a software bottleneck but it's the SO so I think is worth mention, Looks like windows pro separate the cores in 2 groups and that the scheduler fails terribly into send jobs to both of them ( unless the software is properly set to do so like Vray or Blender) Looks like windows pro Enterprise/Workstation doesn't separate the threads in two groups so this specifics apps that are not programmed exactly to use all the cores, and that rely on the windows schedulers could get benefit from this. again, I can stress my CPU max power but just on half of it. If i could get the other half I would double the power. For instance I can open another instance of the same program, run it and will utilise the other half of the cores available and works like charm also. But simetimes you want more power in one instance of the program, not to run 2 instance of the program I think this would be worth testing cos that means 1, the scheduler of windows pro is shit. 2 if you own a 3990x and this enterprise version just can run soft in all cares and not like the windows pro then is worth the upgrade to windows enterprise/workstation. Sadly. very little information about this topic.
reply
Daniel
Best content creation builds 1. Threadripper 3990X 64 core 128 thread 2. 9Ghz-4. 3Ghz is 3990 2. GIGABYTE TRX40 ATX AORUS PRO WIFI 400 3. G. Skill Trident Z Neo 64 GB (4 x 16 GB) DDR4-3600 Memory ram X 2 for 128GB 820 4. 2TB Aorus NVME PCIE 4. 0 SSD M. 2 NVME hard drive 5GB/s read and write speeds 424 5. Titan RTX graphics card 2500 6. Matrexx 55 deepcool case 40 7. 90 + rating 2400 watt power supply 136 8. Fractal Design S36 360 AIO 117 9. Windows 10 Pro 64 bit (you could get a hold of Windows 10 Pro 64 bit key for 15 range) Total: 8342 Or for graphics card use an Radeon Instinct MI60 1. Threadripper 3990X 64 core 128 thread 2. 9Ghz-4. 3Ghz is 3990 2. GIGABYTE TRX40 AORUS XTREME TRX40 850 (comes with the adaptor to attach 4 2TB Aorus NVMEs) 3. G. Skill Trident Z Neo 64 GB (4 x 16 GB) DDR4-3600 Memory ram X 2 for 128GB 820 4. 4X 2TB Aorus NVME PCIE 4. 0 SSD M. 2 NVME hard drive 5GB/s read and write speeds 1696 5. 1X Titan RTX 2500 or 4X Titan RTX 10, 000 6. Thermaltake View 91 Plus 405 7. 90 + rating 2400 watt power supply 136 (get from Frys online) 8. Fractal Design S36 360 AIO 117 9. Windows 10 Pro 64 bit (you could get a hold of Windows 10 Pro 64 bit key for 15 range) Total: 10, 429 or 17, 629 Or for graphics card use Radeon Instinct MI60. I don't know if this graphics card has deep machine learning like Titan RTX. If you want to use 4 graphics cards and it only accepts an AMD graphics card for multiple graphics cards, XFX Radeon VII 16 GB Video Card 550 is a good choice. If you want to go Nvidia for graphics cards the 2080 Ti is the best choice. 2500 Titan RTX is better for content creation than 2080 Ti, but has a lower bottom out framerate consistency than MSI GeForce RTX 2080 Ti GAMING X TRIO Graphics Card 1234 for gaming.
reply
Best content creation builds 1. Threadripper 3990X 64 core 128 thread 2. 9Ghz-4. 3Ghz is 3990 2. GIGABYTE TRX40 ATX AORUS PRO WIFI 400 3. G. Skill Trident Z Neo 64 GB (4 x 16 GB) DDR4-3600 Memory ram X 2 for 128GB 820 4. 2TB Aorus NVME PCIE 4. 0 SSD M. 2 NVME hard drive 5GB/s read and write speeds 424 5. Titan RTX graphics card 2500 6. Matrexx 55 deepcool case 40 7. 90 + rating 2400 watt power supply 136 8. Fractal Design S36 360 AIO 117 9. Windows 10 Pro 64 bit (you could get a hold of Windows 10 Pro 64 bit key for 15 range) Total: 8342 Or for graphics card use an Radeon Instinct MI60 1. Threadripper 3990X 64 core 128 thread 2. 9Ghz-4. 3Ghz is 3990 2. GIGABYTE TRX40 AORUS XTREME TRX40 850 (comes with the adaptor to attach 4 2TB Aorus NVMEs) 3. G. Skill Trident Z Neo 64 GB (4 x 16 GB) DDR4-3600 Memory ram X 2 for 128GB 820 4. 4X 2TB Aorus NVME PCIE 4. 0 SSD M. 2 NVME hard drive 5GB/s read and write speeds 1696 5. 1X Titan RTX 2500 or 4X Titan RTX 10, 000 6. Thermaltake View 91 Plus 405 7. 90 + rating 2400 watt power supply 136 (get from Frys online) 8. Fractal Design S36 360 AIO 117 9. Windows 10 Pro 64 bit (you could get a hold of Windows 10 Pro 64 bit key for 15 range) Total: 10, 429 or 17, 629 Or for graphics card use Radeon Instinct MI60. I don't know if this graphics card has deep machine learning like Titan RTX. If you want to use 4 graphics cards and it only accepts an AMD graphics card for multiple graphics cards, XFX Radeon VII 16 GB Video Card 550 is a good choice. If you want to go Nvidia for graphics cards the 2080 Ti is the best choice. 2500 Titan RTX is better for content creation than 2080 Ti, but has a lower bottom out framerate consistency than MSI GeForce RTX 2080 Ti GAMING X TRIO Graphics Card 1234 for gaming.
reply
Paul
AMD Threadripper 3990X 128-thread monster is too much for Windows 10 Pro to handle AMD s Threadripper 3990X is undoubtedly a mighty high-end desktop processor bristling with 64-cores and 128-threads, but Windows 10 Pro can t actually deal with the latter it s just too many threads to handle, apparently. As you may be aware, Windows 10 Home can t handle any more than 64-cores (or threads, but Windows 10 Pro can supposedly cope with 128-threads, at least according to the official specs of the OS. However, as discussed in Anandtech s review, what actually happens when you run a Threadripper 3990X in a Windows 10 Pro PC is that the operating system splits those 128-threads into two separate processor groups of 64-threads. And the OS even refers to these as two separate sockets, as if it was a dual-CPU machine. Sadly this isn t a simple labeling issue: splitting the 3990X into two CPU groups actually impacts on real-world performance. It can be the case that an application might be launched inside one group, and even though it s a multithreaded app, it may only be able to access those 64-threads in its group, rather than the full 128-threads effectively detracting from its performance. The end result is that Anandtech observed a lot of slowdowns in some benchmarks, although disabling SMT (simultaneous multithreading) cured the problem, but of course then you are only running with 64-cores, even if you still get most of the performance of the 3990X.
reply
AMD Threadripper 3990X 128-thread monster is too much for Windows 10 Pro to handle AMD s Threadripper 3990X is undoubtedly a mighty high-end desktop processor bristling with 64-cores and 128-threads, but Windows 10 Pro can t actually deal with the latter it s just too many threads to handle, apparently. As you may be aware, Windows 10 Home can t handle any more than 64-cores (or threads, but Windows 10 Pro can supposedly cope with 128-threads, at least according to the official specs of the OS. However, as discussed in Anandtech s review, what actually happens when you run a Threadripper 3990X in a Windows 10 Pro PC is that the operating system splits those 128-threads into two separate processor groups of 64-threads. And the OS even refers to these as two separate sockets, as if it was a dual-CPU machine. Sadly this isn t a simple labeling issue: splitting the 3990X into two CPU groups actually impacts on real-world performance. It can be the case that an application might be launched inside one group, and even though it s a multithreaded app, it may only be able to access those 64-threads in its group, rather than the full 128-threads effectively detracting from its performance. The end result is that Anandtech observed a lot of slowdowns in some benchmarks, although disabling SMT (simultaneous multithreading) cured the problem, but of course then you are only running with 64-cores, even if you still get most of the performance of the 3990X.
reply
Mark
Steve when compiling anything memory is king. In a perfect world, a 3990x would use 128gb for compilers alone. This is important - 1 THREAD 1 Gig of memory. When compiling code, 2 things are key beyond the CPU Itself, one is rapid disks, you will (depending on the build) do a lot of disk thrashing, so SSD / NVMe, the second is enough RAM to back it all, and even then from what i have learned so far, the 3990x is so heavily memory-constrained in workload using all 128 threads results in a slower build, using make -j 96 or equivalent gives a faster compile and uses just 96 gigs of ram. If you have less than 1 gig per thread you'll find the system starts swapping to disk, and as soon as that shit starts its game over for performance, match the build commands to the system being used, and then a week it. If like me you can afford it, strap 256 gigs of ram with it. Then go nuts on your benchmarking compiles. Youll be amazed how hardcore the memory constraints are on the 3990, you simply cant feed it large amounts of data fast enough, as evidenced by make -j 96 being about 20% faster than make -j 128
reply
Steve when compiling anything memory is king. In a perfect world, a 3990x would use 128gb for compilers alone. This is important - 1 THREAD 1 Gig of memory. When compiling code, 2 things are key beyond the CPU Itself, one is rapid disks, you will (depending on the build) do a lot of disk thrashing, so SSD / NVMe, the second is enough RAM to back it all, and even then from what i have learned so far, the 3990x is so heavily memory-constrained in workload using all 128 threads results in a slower build, using make -j 96 or equivalent gives a faster compile and uses just 96 gigs of ram. If you have less than 1 gig per thread you'll find the system starts swapping to disk, and as soon as that shit starts its game over for performance, match the build commands to the system being used, and then a week it. If like me you can afford it, strap 256 gigs of ram with it. Then go nuts on your benchmarking compiles. Youll be amazed how hardcore the memory constraints are on the 3990, you simply cant feed it large amounts of data fast enough, as evidenced by make -j 96 being about 20% faster than make -j 128
reply
Thanny
Fifteen years ago, when dual-core CPUs were first coming out, reviewers had a problem. Very few of the programs they use to test were able to use more than one thread for processing. So what did they do? They ran more than one program at a time! Amazing, right? What's amazing is that basically none of the current generation of reviewers have figured out this basic concept. Even when reviewing a 64-core processor, they still just run one program at a time, as if they're providing any meaningful information whatsoever by so doing. With a 3990X, you could run two or three Premiere renders at once, and still have plenty of performance left over to play a game while waiting for the renders to finish, without losing much of any performance at all, in either the renders of the game. But you won't find anyone doing any tests of that sort, for some unfathomable reason.
reply
Fifteen years ago, when dual-core CPUs were first coming out, reviewers had a problem. Very few of the programs they use to test were able to use more than one thread for processing. So what did they do? They ran more than one program at a time! Amazing, right? What's amazing is that basically none of the current generation of reviewers have figured out this basic concept. Even when reviewing a 64-core processor, they still just run one program at a time, as if they're providing any meaningful information whatsoever by so doing. With a 3990X, you could run two or three Premiere renders at once, and still have plenty of performance left over to play a game while waiting for the renders to finish, without losing much of any performance at all, in either the renders of the game. But you won't find anyone doing any tests of that sort, for some unfathomable reason.
reply
Jorge
Even now as I sit here and watch this review, there's only one use case I can think of where it would feasibly make sense for a person like myself to have a build with something like these threadrippers, and the 3990X. Namely, in an mATX or maybe even an ITX build, using it for real-time CPU encode to be able to get some high-quality x264 encode with capture card footage to then broadcast live to twitch or similar. With something like the 3990X, I would think, there would be enough cores and threads for OBS to basically be like yeah, okay, you can do real-time x264 and have it output to a stream at something approaching BluRay quality. but that would be quite the overstated investment for such a singular use case XD still, I have to wonder if this is something feasible on such a high-core CPU, or if even a lower-tier chip could do this finally.
reply
Even now as I sit here and watch this review, there's only one use case I can think of where it would feasibly make sense for a person like myself to have a build with something like these threadrippers, and the 3990X. Namely, in an mATX or maybe even an ITX build, using it for real-time CPU encode to be able to get some high-quality x264 encode with capture card footage to then broadcast live to twitch or similar. With something like the 3990X, I would think, there would be enough cores and threads for OBS to basically be like yeah, okay, you can do real-time x264 and have it output to a stream at something approaching BluRay quality. but that would be quite the overstated investment for such a singular use case XD still, I have to wonder if this is something feasible on such a high-core CPU, or if even a lower-tier chip could do this finally.
reply
Ted
Interesting stuff. Youmentioned in a recent video wanting questions for the pro overclockers. It is not so much about the actual aspect of what they are doing and why it is more about the actual physics of the situation. Absolute zero is the abscence of any movement, I think, and the point of overclocking the cpu is to reduce temperatures to possibly increase the performance out of it. I am wondering what the lowest temperature you could get down to before the speed the electricity can move across transistors becomes negatively impacted and stops functioning. Also why liquid nitrogen? Is that because it only gets temps down to the point it can and using something colder like liquid helium goes too far and past the point nitrogen becomes a solid?
reply
Interesting stuff. Youmentioned in a recent video wanting questions for the pro overclockers. It is not so much about the actual aspect of what they are doing and why it is more about the actual physics of the situation. Absolute zero is the abscence of any movement, I think, and the point of overclocking the cpu is to reduce temperatures to possibly increase the performance out of it. I am wondering what the lowest temperature you could get down to before the speed the electricity can move across transistors becomes negatively impacted and stops functioning. Also why liquid nitrogen? Is that because it only gets temps down to the point it can and using something colder like liquid helium goes too far and past the point nitrogen becomes a solid?
reply
ExtinctInsanity
Lmao, well, I can tell you're an Intel fanboy. I have both the 9900k and 3990x and the 3990x outshined the 9900k in almost every aspect I used it for. Better in games, rendering(video and CAD, several benchmark softwares, etc. Couldn't get the speed as high as the 9900k though. The 3990x is not designed for just servers. If you checked the AMD release notes you would have read yes they designed the 3990x for servers just like the gen 2 Threadripper but also designed it for desktop gaming as well unlike gen 2 to open this cpu for more of a broadened field of consumers. Anyways, I love your videos my dude and I'm looking forward to the factory tours.
reply
Lmao, well, I can tell you're an Intel fanboy. I have both the 9900k and 3990x and the 3990x outshined the 9900k in almost every aspect I used it for. Better in games, rendering(video and CAD, several benchmark softwares, etc. Couldn't get the speed as high as the 9900k though. The 3990x is not designed for just servers. If you checked the AMD release notes you would have read yes they designed the 3990x for servers just like the gen 2 Threadripper but also designed it for desktop gaming as well unlike gen 2 to open this cpu for more of a broadened field of consumers. Anyways, I love your videos my dude and I'm looking forward to the factory tours.
reply
Victor
Compilation tasks usually require up to 2GB RAM per _thread_. So, 64GB RAM for 64 cores worked for Chromium, but can be too low for other C++ projects. Hopefully, 128 threads is too much for those many C++ projects in the wild because they just don't have so many independent compilation units to load all those threads at the same time. If some arbitrary compilation task don't scale from 32 to 64 core, this probably not because hardware does not scale, but because the project itself just don't have enough number of independent compilation units to feed increased number of cores.
reply
Compilation tasks usually require up to 2GB RAM per _thread_. So, 64GB RAM for 64 cores worked for Chromium, but can be too low for other C++ projects. Hopefully, 128 threads is too much for those many C++ projects in the wild because they just don't have so many independent compilation units to load all those threads at the same time. If some arbitrary compilation task don't scale from 32 to 64 core, this probably not because hardware does not scale, but because the project itself just don't have enough number of independent compilation units to feed increased number of cores.
reply
BRUXXUS
I have a super niche program I use for my work (Pyware 3D, that I really wish I could find some hardware benchmarks on. It's a design and animation program that runs both 3D and 2D animations with audio. It's built from jME. For some reason it never pushes my CPU or GPU to utilize more than about 20-50% but there's no performance caps built in, so I don't know what's up there. If I force higher frequency on my CPU I do get better performance, but the program won't trigger boost on its own. yet, it will spread load across all threads on my 3950x. Very weird.
reply
I have a super niche program I use for my work (Pyware 3D, that I really wish I could find some hardware benchmarks on. It's a design and animation program that runs both 3D and 2D animations with audio. It's built from jME. For some reason it never pushes my CPU or GPU to utilize more than about 20-50% but there's no performance caps built in, so I don't know what's up there. If I force higher frequency on my CPU I do get better performance, but the program won't trigger boost on its own. yet, it will spread load across all threads on my 3950x. Very weird.
reply
Add a review, comment
Other channel videos
