Crazy Efficient: AMD Threadripper 7980X & 7970X CPU Review & Benchmarks
video description
Date: 2023-11-21
Related videos
Comments and reviews: 20
Jonathan
Amazing tests, it's interesting to see how well those CPUs perform on code compilation, because despite being a highly parallelizable workload, at the end of the compilation pipeline, in which you have to assemble a single artifact (or a couple of individual artifacts), compilers and linkers are very reliant on the performance of individual cores.
And those CPUs are clearly worse at single-core performance (not only because of power and thermal concerns, but scheduling gets harder), but they save so much time by going through the parallel bits extremely fast that, even if those CPUs take longer to go through the non-parallel bits, it's still faster and more efficient on the job than the lower core-count counterparts.
It still like you said, you have to evaluate whether your workload takes advantage of it or not, and also be aware that there's a difference of highly multithreaded workload and highly parallelizable workload, the latter benefits way more from high core counts than the former, which commonly has more interdependency between the threads, so one thread stalling may negatively affect the others.
Also I'm a little curious to whether those CCDs are as powerful as the customer lineup or not, how the gaming performance fare if you were up to have only one CCD enabled. Yes, it's crazy to buy this monster and proceed to disable all but one CCD, however the question is whether those cores can hold higher clocks for longer if they were not power and thermal constrained.
reply
Amazing tests, it's interesting to see how well those CPUs perform on code compilation, because despite being a highly parallelizable workload, at the end of the compilation pipeline, in which you have to assemble a single artifact (or a couple of individual artifacts), compilers and linkers are very reliant on the performance of individual cores.
And those CPUs are clearly worse at single-core performance (not only because of power and thermal concerns, but scheduling gets harder), but they save so much time by going through the parallel bits extremely fast that, even if those CPUs take longer to go through the non-parallel bits, it's still faster and more efficient on the job than the lower core-count counterparts.
It still like you said, you have to evaluate whether your workload takes advantage of it or not, and also be aware that there's a difference of highly multithreaded workload and highly parallelizable workload, the latter benefits way more from high core counts than the former, which commonly has more interdependency between the threads, so one thread stalling may negatively affect the others.
Also I'm a little curious to whether those CCDs are as powerful as the customer lineup or not, how the gaming performance fare if you were up to have only one CCD enabled. Yes, it's crazy to buy this monster and proceed to disable all but one CCD, however the question is whether those cores can hold higher clocks for longer if they were not power and thermal constrained.
reply
PotatoHate
oh man i wish i could buy this for my personal projects but my actual work provides us with powerful pcs so i can only see this buy as if you do freelancing or you own a indie compony and my personal pc has 5950x so def insane jump in productivity and better game performance. This is a bit annoying area to be in, work in professional setting and for any personal projects you have to buy super expensive hardware yourself even though you are not going to making any money from these projects so you basically just wasting money in a way, i mean its the same with software like indie license is expensive for personal projects that don't make any money in return or if you want to learn it with full features or faster iterations time.
Maybe one day if it goes on sale, i could do some super fast fluid, smoke sim and cpu rendering also experiment with writing multiprocessing software and checking how this 64 core cpu would handle it.
reply
oh man i wish i could buy this for my personal projects but my actual work provides us with powerful pcs so i can only see this buy as if you do freelancing or you own a indie compony and my personal pc has 5950x so def insane jump in productivity and better game performance. This is a bit annoying area to be in, work in professional setting and for any personal projects you have to buy super expensive hardware yourself even though you are not going to making any money from these projects so you basically just wasting money in a way, i mean its the same with software like indie license is expensive for personal projects that don't make any money in return or if you want to learn it with full features or faster iterations time.
Maybe one day if it goes on sale, i could do some super fast fluid, smoke sim and cpu rendering also experiment with writing multiprocessing software and checking how this 64 core cpu would handle it.
reply
ChrisM541
1) Unfortunately, AMD haven't addressed - whatsoever - the rather massive Elephant in the room...that they royally shafted their entire non-pro/enthusiast HEDT userbase with that insane TRX40 'take the money and run' move. To publicly promise long term support for these expensive systems then laugh loudly as they immediately EOL TRX40 is about as braindead a move as you can get.
2) As for TR 7000, take a second to compare TRX40 vs TRX50 motherboards. Stunning....shocking, isn't it. TRX50 motherboards are severely cut down in IO and functionality compared to TRX40. The TRX40 Zenith II Extreme is a particular eye opener. No wonder MSI pulled out of TR.
No non-pro enthusiast, not even The Braindead, are going to invest in AMD's TR ever again. To repeat, AMD have no interest in compensating TRX40 owners.
reply
1) Unfortunately, AMD haven't addressed - whatsoever - the rather massive Elephant in the room...that they royally shafted their entire non-pro/enthusiast HEDT userbase with that insane TRX40 'take the money and run' move. To publicly promise long term support for these expensive systems then laugh loudly as they immediately EOL TRX40 is about as braindead a move as you can get.
2) As for TR 7000, take a second to compare TRX40 vs TRX50 motherboards. Stunning....shocking, isn't it. TRX50 motherboards are severely cut down in IO and functionality compared to TRX40. The TRX40 Zenith II Extreme is a particular eye opener. No wonder MSI pulled out of TR.
No non-pro enthusiast, not even The Braindead, are going to invest in AMD's TR ever again. To repeat, AMD have no interest in compensating TRX40 owners.
reply
Ellej
As a molecular simulation PhD, I also loved seeing these workstation tests included. Especially Lammps is really relevant for me. I actually think that the testcase in lammps is also molecular dynamics (just like namd). They just implemented some of the calculations/parallizing different.
What might be important to note is that in my field we topically run simulation 3 till 5 times with identical settings, just different initial configurations (results here have statistical value). So to me seeing the results of the 32cores compared to the 64 cores set up as 2 simulations next to each other would be very informative as well. (Or both running 16 thread simulations up till the cpu is 100% loaded would be good too). I would be interested in what the CFD people here would think of that.
reply
As a molecular simulation PhD, I also loved seeing these workstation tests included. Especially Lammps is really relevant for me. I actually think that the testcase in lammps is also molecular dynamics (just like namd). They just implemented some of the calculations/parallizing different.
What might be important to note is that in my field we topically run simulation 3 till 5 times with identical settings, just different initial configurations (results here have statistical value). So to me seeing the results of the 32cores compared to the 64 cores set up as 2 simulations next to each other would be very informative as well. (Or both running 16 thread simulations up till the cpu is 100% loaded would be good too). I would be interested in what the CFD people here would think of that.
reply
Pitipicqou
Thanks for the detailed review as always! Are you planning on a thorough review of the 7960X as well?
Especially if you're interested in video production workloads, the 7960X looks on paper like it might hit a sweet spot: you get the lanes of Threadripper for a second GPU, dedicated video output, high speed networking or storage, and higher clocks than the 7970X or 7980X, at a much lower cost (according to Puget, the 5965WX was almost tied with the higher tier 5th gen Threadripper for content creation).
Those features would make it the ideal CPU for a lot of video professionnals.
The inclusion of Resolve benchmarks would also be really cool, although your test suite is already a lot of work, it might help out quite a bit of people
reply
Thanks for the detailed review as always! Are you planning on a thorough review of the 7960X as well?
Especially if you're interested in video production workloads, the 7960X looks on paper like it might hit a sweet spot: you get the lanes of Threadripper for a second GPU, dedicated video output, high speed networking or storage, and higher clocks than the 7970X or 7980X, at a much lower cost (according to Puget, the 5965WX was almost tied with the higher tier 5th gen Threadripper for content creation).
Those features would make it the ideal CPU for a lot of video professionnals.
The inclusion of Resolve benchmarks would also be really cool, although your test suite is already a lot of work, it might help out quite a bit of people
reply
dsledoux
Not adding Threadripper Pro to the review seems more than a little odd to me. Clearly anyone that would be interested in this review and/or 7xxx series Threadrippers would certainly compare/be interesting in how they stack vs 5xxx series Threadripper Pro.
We re-tested the old Threadripper 3970X and 3960X to include on these charts, but skipped the 5000-series PRO CPUs (as they were not HEDT parts).
Come on now, why not just be straight with us and say something along the lines of, we didn't have time or the chips needed to test 5xxx series or sorry, we have severe OCD surrounding product naming conventions/classification
I being 50% serious and 50% cheeky... maybe like 60/40... Thanks for the hard work
reply
Not adding Threadripper Pro to the review seems more than a little odd to me. Clearly anyone that would be interested in this review and/or 7xxx series Threadrippers would certainly compare/be interesting in how they stack vs 5xxx series Threadripper Pro.
We re-tested the old Threadripper 3970X and 3960X to include on these charts, but skipped the 5000-series PRO CPUs (as they were not HEDT parts).
Come on now, why not just be straight with us and say something along the lines of, we didn't have time or the chips needed to test 5xxx series or sorry, we have severe OCD surrounding product naming conventions/classification
I being 50% serious and 50% cheeky... maybe like 60/40... Thanks for the hard work
reply
Ismail
Hey Gamers Nexus. Weird request. The world of workstation computers is starting to quickly involve Artificial Intelligence type workloads. I know a majority of those workloads are Pytorch type Nvidia GPU bound workloads. Since this was a Workstation category review, I think it'd be interesting to see things like Image Generation, Training, and Task Completion time for an AI workload. Question would be, can I get gains on those workloads by splurging on a big Workstation CPU? Or is it still better to go some traditional type CPUs?
Not too familiar with testing in that space, but I do believe there are benchmark type tools out there. Peace and Thank you for all the great content
reply
Hey Gamers Nexus. Weird request. The world of workstation computers is starting to quickly involve Artificial Intelligence type workloads. I know a majority of those workloads are Pytorch type Nvidia GPU bound workloads. Since this was a Workstation category review, I think it'd be interesting to see things like Image Generation, Training, and Task Completion time for an AI workload. Question would be, can I get gains on those workloads by splurging on a big Workstation CPU? Or is it still better to go some traditional type CPUs?
Not too familiar with testing in that space, but I do believe there are benchmark type tools out there. Peace and Thank you for all the great content
reply
vizintini
I would have added the Cinebench suite scores R23 and R24 plus V-ray benchmark plus Corona Benchmark , Houdini particle simulations, etc .. and got rid of games benchmarks which are irrelevant for this CPUs. Most of the people who buy this kind of CPU's at home are working as 3D render artist or in VFX industry. So rendering and fluid / particle simulations is important and not VM stuff , server stuff ,or molecular thing, financial blah blah... . 90% of the people who buy this CPU are using it in 3D aplications. This CPUs are made exactly for this.
Also i would have added the 3990x and de 5995wx in the comparison charts. There is no other 64cores cpu in the charts...
reply
I would have added the Cinebench suite scores R23 and R24 plus V-ray benchmark plus Corona Benchmark , Houdini particle simulations, etc .. and got rid of games benchmarks which are irrelevant for this CPUs. Most of the people who buy this kind of CPU's at home are working as 3D render artist or in VFX industry. So rendering and fluid / particle simulations is important and not VM stuff , server stuff ,or molecular thing, financial blah blah... . 90% of the people who buy this CPU are using it in 3D aplications. This CPUs are made exactly for this.
Also i would have added the 3990x and de 5995wx in the comparison charts. There is no other 64cores cpu in the charts...
reply
Kristopher
24:05 the financial and probability simulations are very important for machine learning - whether that's training or inference, depending on the model architecture. It's a very big deal for ML but even bigger for time series data analysis in market dynamics when you assess a large system of signals and indicators which compare to one another and are compared across several time scales in addition to many different products whose price action is being recorded. Log complexity on millisecond updates across dozens or hundreds of items needs a ton of processing if you don't want to miss out on an arbitrage or trade intraday algorithmically.
reply
24:05 the financial and probability simulations are very important for machine learning - whether that's training or inference, depending on the model architecture. It's a very big deal for ML but even bigger for time series data analysis in market dynamics when you assess a large system of signals and indicators which compare to one another and are compared across several time scales in addition to many different products whose price action is being recorded. Log complexity on millisecond updates across dozens or hundreds of items needs a ton of processing if you don't want to miss out on an arbitrage or trade intraday algorithmically.
reply
Andrew
I kind of skipped through some of it, but for scientific applications, many can bottleneck on things like file bandwidth, memory bandwidth, or even cache size. I'd recommend trying to find a couple types of analyses that scale linearly with the number of cores for the benchmarks (i.e., on the same CPU, run the test with different numbers of cores and see if plots as a straight line). One that is math heavy and one that is logic/conditional heavy, without hitting other system bottlenecks. These two benchmarks would then represent the ideal case for scaling and serve as sort of a reference point for comparing across CPUs.
reply
I kind of skipped through some of it, but for scientific applications, many can bottleneck on things like file bandwidth, memory bandwidth, or even cache size. I'd recommend trying to find a couple types of analyses that scale linearly with the number of cores for the benchmarks (i.e., on the same CPU, run the test with different numbers of cores and see if plots as a straight line). One that is math heavy and one that is logic/conditional heavy, without hitting other system bottlenecks. These two benchmarks would then represent the ideal case for scaling and serve as sort of a reference point for comparing across CPUs.
reply
falldownandgoboom
Another comprehensive review of new processors. Once again the 14900, 14700, and 14600 K skus are represented. Ditto the 13900, 13700, and 13600 K skus. And, of course, the 12900K as well as the 12600K but again, no 12700K! I'm starting to question my sanity. Maybe I only imagined building a PC around an Intel 12700K processor. I know GN has an F variant in a test bench - or did until recently. Maybe in the next video that fails to include the 12700K Steve can at least explain the reasoning. I hope the explanation is not that, with many thousands of dollars worth of CPUs on hand, GN doesn't own a 12700K.
reply
Another comprehensive review of new processors. Once again the 14900, 14700, and 14600 K skus are represented. Ditto the 13900, 13700, and 13600 K skus. And, of course, the 12900K as well as the 12600K but again, no 12700K! I'm starting to question my sanity. Maybe I only imagined building a PC around an Intel 12700K processor. I know GN has an F variant in a test bench - or did until recently. Maybe in the next video that fails to include the 12700K Steve can at least explain the reasoning. I hope the explanation is not that, with many thousands of dollars worth of CPUs on hand, GN doesn't own a 12700K.
reply
Robert
You should measure throughput using performance rather than Meaningless Information Propagated by Salesmen (MIPS). MIPS is meaningless because there is no single definition for Instruction . Only The Ultimate Ridiculous Instruction Set Computer (RISC) has only one instruction. For all other architectures, the MIPS figure would require testing only with a Jump to This Same Instruction load which does nothing but does use many clock cycles. Any other load would use many DIFFERENT instructions and thus render a useless measure of performance. All MIPS is good for is measuring clock speed, not computer performance.
reply
You should measure throughput using performance rather than Meaningless Information Propagated by Salesmen (MIPS). MIPS is meaningless because there is no single definition for Instruction . Only The Ultimate Ridiculous Instruction Set Computer (RISC) has only one instruction. For all other architectures, the MIPS figure would require testing only with a Jump to This Same Instruction load which does nothing but does use many clock cycles. Any other load would use many DIFFERENT instructions and thus render a useless measure of performance. All MIPS is good for is measuring clock speed, not computer performance.
reply
Benjamin
Desire to use server or workstation boards and CPUs for gaming is courtesy of EVGA when they made their SR-2 Classified dual 1366 socket server board capable of using dual Intel XEON processors, 12 dimms of memory, and up to 4x graphics cards. The other cool dream for cool-factor at that time was the Supermicro quad socket AMD Opteron server board. For gaming, my i5-6600 mini-ITX with GTX 1080 Ti still runs most games at high settings on 4k, and zero issues with 3D design rendering. The idea of having a T-rex monster system is always going to be way more awesome than a little gecko, though.
reply
Desire to use server or workstation boards and CPUs for gaming is courtesy of EVGA when they made their SR-2 Classified dual 1366 socket server board capable of using dual Intel XEON processors, 12 dimms of memory, and up to 4x graphics cards. The other cool dream for cool-factor at that time was the Supermicro quad socket AMD Opteron server board. For gaming, my i5-6600 mini-ITX with GTX 1080 Ti still runs most games at high settings on 4k, and zero issues with 3D design rendering. The idea of having a T-rex monster system is always going to be way more awesome than a little gecko, though.
reply
Krakenfall
Again, thank you for excellent coverage. I was facepalming when these threadripper CPUs were announced because I had JUST streamed my 7950X/AM5 upgrade build. I do gaming, editing, and ML workloads, so when I realized most X670 motherboards don't even run two full bandwidth x16 slots, I really thought I messed up not going for threadripper. With this review, I now know threadripper isn't exactly the magic bullet I was looking for with my use case. I guess I'll have to spread out my high-bandwidth PCIE components to my stream PC and make it up with 10G network cards or something. Ugh
reply
Again, thank you for excellent coverage. I was facepalming when these threadripper CPUs were announced because I had JUST streamed my 7950X/AM5 upgrade build. I do gaming, editing, and ML workloads, so when I realized most X670 motherboards don't even run two full bandwidth x16 slots, I really thought I messed up not going for threadripper. With this review, I now know threadripper isn't exactly the magic bullet I was looking for with my use case. I guess I'll have to spread out my high-bandwidth PCIE components to my stream PC and make it up with 10G network cards or something. Ugh
reply
itech
3:30 I have a major issue with that call out. I want to know what game actually makes use of more than 16 threads currently & massive ram bandwidth? How is that not a vailed question for gaming? No AM5 board support what I want/need, so I have to go to Threadripper because the WRX90/TRX50 boards actually still support S.L.I. Beside that my cards (RTX 2080 ti's in S.L.I, to soon RTX 3090 ti's S.L.I) won't be bottlenecked by the R7 5800x 3D I have, & being in x8 pci-express slots. I'm cpu bottlenecked with this cpu, mostly because of lower clocks.
reply
3:30 I have a major issue with that call out. I want to know what game actually makes use of more than 16 threads currently & massive ram bandwidth? How is that not a vailed question for gaming? No AM5 board support what I want/need, so I have to go to Threadripper because the WRX90/TRX50 boards actually still support S.L.I. Beside that my cards (RTX 2080 ti's in S.L.I, to soon RTX 3090 ti's S.L.I) won't be bottlenecked by the R7 5800x 3D I have, & being in x8 pci-express slots. I'm cpu bottlenecked with this cpu, mostly because of lower clocks.
reply
Anonymous
B-S, binomial and Monte Carlo are three of the most basic derivative (options) and forecasting models: they are very useful, and depending on the dataset Monte Carlo simulations could take a while. At university some teachers did not have enough horsepower to run them real time with lots of references and multiple (6+) graphs updating simultaneously. Running Excel with 30+ Binomial decision trees, while calculating the resulting Black-Scholes pricing for each step, and having their Greeks update live needs some CPU power to do it.
reply
B-S, binomial and Monte Carlo are three of the most basic derivative (options) and forecasting models: they are very useful, and depending on the dataset Monte Carlo simulations could take a while. At university some teachers did not have enough horsepower to run them real time with lots of references and multiple (6+) graphs updating simultaneously. Running Excel with 30+ Binomial decision trees, while calculating the resulting Black-Scholes pricing for each step, and having their Greeks update live needs some CPU power to do it.
reply
RFC3514
Comparing just CPU power consumption isn't entirely fair in relative terms, because the rest of the system is also powered during rendering. So even if a CPU consumes just 50% of the power to do the the same work, that doesn't mean the system as a whole has twice the efficiency. If the CPU was originally using (for example) 60% of the total system power, then a 50% reduction in _CPU_ power is really just a 30% reduction in total power consumption to do the same work. Still relevant, but it's an important distinction to make.
reply
Comparing just CPU power consumption isn't entirely fair in relative terms, because the rest of the system is also powered during rendering. So even if a CPU consumes just 50% of the power to do the the same work, that doesn't mean the system as a whole has twice the efficiency. If the CPU was originally using (for example) 60% of the total system power, then a 50% reduction in _CPU_ power is really just a 30% reduction in total power consumption to do the same work. Still relevant, but it's an important distinction to make.
reply
Ted
Appreciated the HPC (high-performance compute ) benchmarks being added to the mix for the high-core CPUs. I, too, have no idea what most of them mean in real-world terms, but it's cool to see some of the numbers these CPUs might actually be crunching. If you come across someone that _does_ know about these things, is doing something cool, and wants to show off the compute, I wouldn't object to seeing that. HPC is cool. ( -defined because too many things use the letters HP)
reply
Appreciated the HPC (high-performance compute ) benchmarks being added to the mix for the high-core CPUs. I, too, have no idea what most of them mean in real-world terms, but it's cool to see some of the numbers these CPUs might actually be crunching. If you come across someone that _does_ know about these things, is doing something cool, and wants to show off the compute, I wouldn't object to seeing that. HPC is cool. ( -defined because too many things use the letters HP)
reply
movax20h
7980X is clearly memory bandwidth limited in many of these benchmarks. You can see sometimes zero scaling, despite the workload being scalable well (things like parallel decompression, rendering), in places where there is less memory bandwidth need, like compilation or compression, it scales way better (and remaining difference can be explained due to lower base/boost clocks, and not perfect scaling of the benchmark itself - i.e. linking during compilation)
reply
7980X is clearly memory bandwidth limited in many of these benchmarks. You can see sometimes zero scaling, despite the workload being scalable well (things like parallel decompression, rendering), in places where there is less memory bandwidth need, like compilation or compression, it scales way better (and remaining difference can be explained due to lower base/boost clocks, and not perfect scaling of the benchmark itself - i.e. linking during compilation)
reply
Omizuke
So many questions!
Among them, now Im curious about their performance/power efficiency with cores or CCD disable. Does AMD have other CPU that could translate to 28w on 8 cores with good performance?
And the 7800X3D looks more amazing now.
I love how these kind of monstrous CPU reviews bot just inform us about them. But also gives you an even better perspective about where other CPU stands when it comes to performance and specially efficiency.
reply
So many questions!
Among them, now Im curious about their performance/power efficiency with cores or CCD disable. Does AMD have other CPU that could translate to 28w on 8 cores with good performance?
And the 7800X3D looks more amazing now.
I love how these kind of monstrous CPU reviews bot just inform us about them. But also gives you an even better perspective about where other CPU stands when it comes to performance and specially efficiency.
reply
Add a review, comment
Other channel videos
