Intel 28-Core W-3175X Revisit vs. Threadripper 3970X, 3960X: Power, OC, & Benchmarks
video description
Date: 2020-05-06
Related videos
Comments and reviews: 10
Dudamoos
14: 08 - We think this comes down to an instruction set that AMD doesn't have or some architectural difference. - I can assure you it's the latter. The information below is rather detailed, so TL; DR for the question at hand is that the W-3175X likely scores higher than the TR 3970X on the 7-Zip compression benchmark because this benchmark much prefers Intel's cache organization/behavior over AMD's. There's no AVX or anything fancy like that in 7-Zip, just boring 32-bit integer computations and some large memory buffers. As a programmer somewhat familiar with how general purpose compression software works and who has read the documentation for these benchmarks - this is what I can say: (1) The benchmark only runs two threads per compression task. Newer general purpose compression algorithms can use more than two threads per compressed block, but 7-Zip doesn't. If you make 7-Zip use more two threads, it instead divides the data into chunks to compress independently (a speed vs compression trade-off. For the benchmark, there is one compressor instance per two threads. However, this test doesn't scale as well to more cores as the decompression benchmark does. (2) When you use larger dictionaries (as 7-Zip tends to, you are very dependent on memory latency (not bandwidth, just latency, cache behavior, and how well your CPU's out-of-order execution engine can hide stalls when it needs to fetch data from RAM. The documentation for the compression benchmark explicitly mentions this. The Intel W-3175X is likely beating the AMD TR 3970X because Intel's cache organization and behavior is much more friendly to this test than AMD's (likely the L3 cache specifically. (3) The benchmark only runs one thread per decompression task. Decompressing a single compressed block is typically a completely serial process and for general purpose compression is fast enough that you won't gain anything from multi-threading (you'd hit an I/O bottleneck first. For special purpose compression with high complexity and parallelism integral to the design (e. g. new video codecs like AV1) you could see an improvement from multi-threaded decompression, but that's not what you're testing here. This also means that the decompression benchmark is embarrassingly parallel (yes that's a real term) where having more cores is a significant advantage. (4) The decompression benchmark is entirely dependent on the integer compute performance of the CPU. The benchmark documentation says it's most dependent on 32-bit integer ALU latency (multiply, add, bit-shift, etc) and branch mis-prediction penalty ( The decompression task has a very high number of unpredictable branches. The AMD TR 2990WX likely scores poorly for its core count because its cores are so much slower than the competition. The AMD TR 3960X likely beats the Intel W-3175X here because AMD has really invested in improving their IPC with Zen 2 (better ALU, better instruction cache, better branch predictor, better out-of-order execution engine, etc) - to the point it can overcome the four core deficit. I encourage you to read the benchmark documentation if you want more nitty-gritty than I've given here. It's a little old and the English isn't great, but the benchmark itself really hasn't changed in the past ten years. The benchmark technically uses three threads per compression task, but the last thread is negligible for CPU utilization and the benchmark doesn't count it for scoring or anything else. If configure the benchmark to use N threads, there will be N/2 compressor instances. dictionary is a common term for a buffer of past data that can be referenced instead of repeated (though this is somewhat over-simplified. 7-Zip, like Zip, will look for sequences of data it has seen before and replace them with a reference to the last time it saw the data (which takes less space. The size of the dictionary is how far back it can look for repeated data. The algorithms/structures used to make this process faster can be quite memory hungry - 7-Zip will normally use about 188 MB of memory per compressor instance for it.
reply
14: 08 - We think this comes down to an instruction set that AMD doesn't have or some architectural difference. - I can assure you it's the latter. The information below is rather detailed, so TL; DR for the question at hand is that the W-3175X likely scores higher than the TR 3970X on the 7-Zip compression benchmark because this benchmark much prefers Intel's cache organization/behavior over AMD's. There's no AVX or anything fancy like that in 7-Zip, just boring 32-bit integer computations and some large memory buffers. As a programmer somewhat familiar with how general purpose compression software works and who has read the documentation for these benchmarks - this is what I can say: (1) The benchmark only runs two threads per compression task. Newer general purpose compression algorithms can use more than two threads per compressed block, but 7-Zip doesn't. If you make 7-Zip use more two threads, it instead divides the data into chunks to compress independently (a speed vs compression trade-off. For the benchmark, there is one compressor instance per two threads. However, this test doesn't scale as well to more cores as the decompression benchmark does. (2) When you use larger dictionaries (as 7-Zip tends to, you are very dependent on memory latency (not bandwidth, just latency, cache behavior, and how well your CPU's out-of-order execution engine can hide stalls when it needs to fetch data from RAM. The documentation for the compression benchmark explicitly mentions this. The Intel W-3175X is likely beating the AMD TR 3970X because Intel's cache organization and behavior is much more friendly to this test than AMD's (likely the L3 cache specifically. (3) The benchmark only runs one thread per decompression task. Decompressing a single compressed block is typically a completely serial process and for general purpose compression is fast enough that you won't gain anything from multi-threading (you'd hit an I/O bottleneck first. For special purpose compression with high complexity and parallelism integral to the design (e. g. new video codecs like AV1) you could see an improvement from multi-threaded decompression, but that's not what you're testing here. This also means that the decompression benchmark is embarrassingly parallel (yes that's a real term) where having more cores is a significant advantage. (4) The decompression benchmark is entirely dependent on the integer compute performance of the CPU. The benchmark documentation says it's most dependent on 32-bit integer ALU latency (multiply, add, bit-shift, etc) and branch mis-prediction penalty ( The decompression task has a very high number of unpredictable branches. The AMD TR 2990WX likely scores poorly for its core count because its cores are so much slower than the competition. The AMD TR 3960X likely beats the Intel W-3175X here because AMD has really invested in improving their IPC with Zen 2 (better ALU, better instruction cache, better branch predictor, better out-of-order execution engine, etc) - to the point it can overcome the four core deficit. I encourage you to read the benchmark documentation if you want more nitty-gritty than I've given here. It's a little old and the English isn't great, but the benchmark itself really hasn't changed in the past ten years. The benchmark technically uses three threads per compression task, but the last thread is negligible for CPU utilization and the benchmark doesn't count it for scoring or anything else. If configure the benchmark to use N threads, there will be N/2 compressor instances. dictionary is a common term for a buffer of past data that can be referenced instead of repeated (though this is somewhat over-simplified. 7-Zip, like Zip, will look for sequences of data it has seen before and replace them with a reference to the last time it saw the data (which takes less space. The size of the dictionary is how far back it can look for repeated data. The algorithms/structures used to make this process faster can be quite memory hungry - 7-Zip will normally use about 188 MB of memory per compressor instance for it.
reply
Winnetou17
I'll sound very disrespectful, but this kind of a review (I know this is a refresh, it doesn't matter this time) is. not good for this type of a CPU. First thing: too much gaming benchmarks. It's a waste of time for everybody. Not even streamers should look for this CPU. So, gaming benchmarks for this kind of a CPU should be between none or at most 2 games, with several seconds of airtime. The other thing: the productivity benchmarks. are too few. This kind of CPU is rarely for one person and rendering jobs are not everything. This kind of CPU is mostly used in servers. Besides rendering, there's also databases, of many kinds, applications, web servers and above all. virtual machines. That photoshop score was kind of meh. But how does it do with 6 photoshop at once, each in a different virtual machine? How about 7? Or 8? How does the threadripper or R9 3950 fare in this? How many queries per second can it do? Requests per second? How much RAM can it have? How well does it run a special algorithm? Or another algorithm, but in 50 instances? Or Docker container farms? In this video, W-3175X won confortably the 7-Zip compression benchmark. How many other applications/workloads does it win in? Probably not many, if any. But we don't know. And this video sheds way too little light. If you start to factor all the things said above, you start to realize that this kind of a review misses the point for this kind of a CPU. It spends too much time on benchmarks that are not relevant, misses a lot of benchmarks or workloads that are relevant, and I guess it also kind of speaks to the wrong audience. All in all I think this is just mostly a time waster. The folks at the big corporations that buy these CPUs don't decide based on this review. And those of us who look at this never buy something like this.
reply
I'll sound very disrespectful, but this kind of a review (I know this is a refresh, it doesn't matter this time) is. not good for this type of a CPU. First thing: too much gaming benchmarks. It's a waste of time for everybody. Not even streamers should look for this CPU. So, gaming benchmarks for this kind of a CPU should be between none or at most 2 games, with several seconds of airtime. The other thing: the productivity benchmarks. are too few. This kind of CPU is rarely for one person and rendering jobs are not everything. This kind of CPU is mostly used in servers. Besides rendering, there's also databases, of many kinds, applications, web servers and above all. virtual machines. That photoshop score was kind of meh. But how does it do with 6 photoshop at once, each in a different virtual machine? How about 7? Or 8? How does the threadripper or R9 3950 fare in this? How many queries per second can it do? Requests per second? How much RAM can it have? How well does it run a special algorithm? Or another algorithm, but in 50 instances? Or Docker container farms? In this video, W-3175X won confortably the 7-Zip compression benchmark. How many other applications/workloads does it win in? Probably not many, if any. But we don't know. And this video sheds way too little light. If you start to factor all the things said above, you start to realize that this kind of a review misses the point for this kind of a CPU. It spends too much time on benchmarks that are not relevant, misses a lot of benchmarks or workloads that are relevant, and I guess it also kind of speaks to the wrong audience. All in all I think this is just mostly a time waster. The folks at the big corporations that buy these CPUs don't decide based on this review. And those of us who look at this never buy something like this.
reply
Royameadow
The 3175X and its 3275 counterpart, in my view, are a precursor to what we will be expecting from Intel's Ice Lake and Meteor Lake series of Xeon SP and W series chips, and with LGA 3647 becoming a necessity to bring to the Consumer and Prosumer front to compete with Threadripper, I've a feeling that the hypothetical 3375 to 3575 built on the 0I0 and 07 Nanometer architectures will show optimizations far beyond what we have seen from Zen II, Cascade Lake is definitely a beast in its own right when you have north of twenty cores on a chip but the power draw is far too much that I would say waiting for Ice/Tiger/Alder Lake, Golden Cove, and Meteor Lake for a 028 to 064 Core Xeon is probably the best route to take at this time. In the field of CPUs with north of 032 cores, Intel likely has something up their sleeve to take on the 064 Core/0I28 Thread 3990X proper, but when we will get that is still a mystery, especially with Ice Lake SP/W not coming anytime soon, from the looks of it; one would hope though that when we do get that branch of the 0I0 Nanometer chips that cost factor and common availability would be significantly improved, Team Blue knew that they had to drop the prices of Cascade Lake X due to its performance similarities to the Ryzen 09 parts of the Zen II Era and I feel that we will see a similar move going forward, at least for Core, though Xeon's features do need to finally come to Consumer Grade Core and Core: X Series chips to improve competition on all fronts, what they're doing with Comet Lake is a good start but we need more than that and that's why I'm holding out hope that Meteor Lake will address all of the issues that they have had since Zen+ came to the market.
reply
The 3175X and its 3275 counterpart, in my view, are a precursor to what we will be expecting from Intel's Ice Lake and Meteor Lake series of Xeon SP and W series chips, and with LGA 3647 becoming a necessity to bring to the Consumer and Prosumer front to compete with Threadripper, I've a feeling that the hypothetical 3375 to 3575 built on the 0I0 and 07 Nanometer architectures will show optimizations far beyond what we have seen from Zen II, Cascade Lake is definitely a beast in its own right when you have north of twenty cores on a chip but the power draw is far too much that I would say waiting for Ice/Tiger/Alder Lake, Golden Cove, and Meteor Lake for a 028 to 064 Core Xeon is probably the best route to take at this time. In the field of CPUs with north of 032 cores, Intel likely has something up their sleeve to take on the 064 Core/0I28 Thread 3990X proper, but when we will get that is still a mystery, especially with Ice Lake SP/W not coming anytime soon, from the looks of it; one would hope though that when we do get that branch of the 0I0 Nanometer chips that cost factor and common availability would be significantly improved, Team Blue knew that they had to drop the prices of Cascade Lake X due to its performance similarities to the Ryzen 09 parts of the Zen II Era and I feel that we will see a similar move going forward, at least for Core, though Xeon's features do need to finally come to Consumer Grade Core and Core: X Series chips to improve competition on all fronts, what they're doing with Comet Lake is a good start but we need more than that and that's why I'm holding out hope that Meteor Lake will address all of the issues that they have had since Zen+ came to the market.
reply
horton
Greta Thun has left the country so you can go on like this! Man, you are from some neanderthal errortime where Overclock is missing a thermal parameter called global warming caused in part by data center hot spots. And yet you claim this Intel monster just could be a business proposition if if if if if it could be justified. I'm sure the intake of a small jet engine would be sufficient for both your centigrade and decibel levels. Like almost every Youtube guru I've assiduously followed lately, you are recommending the 3950x, but no one has ever seen it, including yourself. The CEO of AMD is a genius! I purchased a 3960x for Photoshop. You clearly don't understand that it is a lack of multiprocessing by the Bridge in Photoshop and Lightroom that leave Adobe hurt by the most negative criticism in their forums. If another company would take advantage of 3960x/ 3970X power to implement Bridge-type thumbnail software it could make Adobe's entire software empire wide open for easy pickings! There are big uses coming down, get ready.
reply
Greta Thun has left the country so you can go on like this! Man, you are from some neanderthal errortime where Overclock is missing a thermal parameter called global warming caused in part by data center hot spots. And yet you claim this Intel monster just could be a business proposition if if if if if it could be justified. I'm sure the intake of a small jet engine would be sufficient for both your centigrade and decibel levels. Like almost every Youtube guru I've assiduously followed lately, you are recommending the 3950x, but no one has ever seen it, including yourself. The CEO of AMD is a genius! I purchased a 3960x for Photoshop. You clearly don't understand that it is a lack of multiprocessing by the Bridge in Photoshop and Lightroom that leave Adobe hurt by the most negative criticism in their forums. If another company would take advantage of 3960x/ 3970X power to implement Bridge-type thumbnail software it could make Adobe's entire software empire wide open for easy pickings! There are big uses coming down, get ready.
reply
Siegbert
Funny to see that even Intels HEDT CPUs destroy AMDs dedicated gaming CPUs like the 3700X. Imo Intel is not lost at all, as AMD is still not able to achieve these high clocks, games highly profit from. Even with obsolete 14nm lithography, Intel is still able to further raise their clock speeds which is just impressive. Thanks to AMDs 3900x(12cores, Intels next top of the line consumer chip will come with 10 cores, still at 14 nm. Intels 14nm technology is now almost 6 years old and got several improvements. Now its more compareable to Global Foundries 12 nm, i guess.
reply
Funny to see that even Intels HEDT CPUs destroy AMDs dedicated gaming CPUs like the 3700X. Imo Intel is not lost at all, as AMD is still not able to achieve these high clocks, games highly profit from. Even with obsolete 14nm lithography, Intel is still able to further raise their clock speeds which is just impressive. Thanks to AMDs 3900x(12cores, Intels next top of the line consumer chip will come with 10 cores, still at 14 nm. Intels 14nm technology is now almost 6 years old and got several improvements. Now its more compareable to Global Foundries 12 nm, i guess.
reply
Spartan0536
3000 just for the CPU, about 1000 IF you can get the motherboard, and likely a 1500W PSU if that CPU is eating 600W. Not to mention you need to drop another 1500 in exotic cooling just to maintain OC clocks without making your soul die on the inside. The W-3175X CPU is effectively worthless at it's price and availability point, especially when comparing other Intel and AMD CPU's that are more readily available and considerably cheaper, but it is interesting to see Intel's top 0. 001% silicon in action.
reply
3000 just for the CPU, about 1000 IF you can get the motherboard, and likely a 1500W PSU if that CPU is eating 600W. Not to mention you need to drop another 1500 in exotic cooling just to maintain OC clocks without making your soul die on the inside. The W-3175X CPU is effectively worthless at it's price and availability point, especially when comparing other Intel and AMD CPU's that are more readily available and considerably cheaper, but it is interesting to see Intel's top 0. 001% silicon in action.
reply
ISOHaven
28: 00 - you never need a 28 core CPU for gaming anyway. you realize they used to say the same thing about 4 core, right? Then again about 8. Then again about 12. don't be so cliche. It's becoming clear that high core/thread count CPUs are the near future. Once game engines get updated to better handle them then all these high core count CPU benchmarks will be worthless and will have to be redone. All of a sudden their gaming productivity will go through the roof.
reply
28: 00 - you never need a 28 core CPU for gaming anyway. you realize they used to say the same thing about 4 core, right? Then again about 8. Then again about 12. don't be so cliche. It's becoming clear that high core/thread count CPUs are the near future. Once game engines get updated to better handle them then all these high core count CPU benchmarks will be worthless and will have to be redone. All of a sudden their gaming productivity will go through the roof.
reply
Νίκος
You sare actually comparing an actual cpu, that can run perfectly on air, to an imaginary one, that requires a 1Hp and change (1360+ Watts) of industrial chiller to cool down when overclocked. By the way, without all this overclocking or silicon lottery stuff. isnt it obvious what the choice for us consumers is? You shouldn t mock the verge and their funny, nontheless harmless videos. 2019's greatest disappointment is you buddy.
reply
You sare actually comparing an actual cpu, that can run perfectly on air, to an imaginary one, that requires a 1Hp and change (1360+ Watts) of industrial chiller to cool down when overclocked. By the way, without all this overclocking or silicon lottery stuff. isnt it obvious what the choice for us consumers is? You shouldn t mock the verge and their funny, nontheless harmless videos. 2019's greatest disappointment is you buddy.
reply
Guest
I love you! Other reviewers are alright, but you are by very far, the most objective and scientific. Very refreshing to see someone who doesn't just try to peddle and sell the highest powered thing to someone who clearly doesn't need it. People like you can really help make a decision as to what someone could buy to fit their use case, with objective data to back it up whilst covering a lot of topics.
reply
I love you! Other reviewers are alright, but you are by very far, the most objective and scientific. Very refreshing to see someone who doesn't just try to peddle and sell the highest powered thing to someone who clearly doesn't need it. People like you can really help make a decision as to what someone could buy to fit their use case, with objective data to back it up whilst covering a lot of topics.
reply
Multitalent
Hey, GN could you consider adding DAWBench (it's a digital audio workstation benchmark) for those of us who do audio editing? It would be helpful to have it as a test for workstation chips in addition to your 3D graphics and video editing benchmarks. It would be hugely helpful to those of us who work in environments where audio editing is a thing. Thanks for taking the time to read this!
reply
Hey, GN could you consider adding DAWBench (it's a digital audio workstation benchmark) for those of us who do audio editing? It would be helpful to have it as a test for workstation chips in addition to your 3D graphics and video editing benchmarks. It would be hugely helpful to those of us who work in environments where audio editing is a thing. Thanks for taking the time to read this!
reply
Add a review, comment
Other channel videos
