• Welcome to TechPowerUp Forums, Guest! Please check out our forum guidelines for info related to our community.
  • The forums have been upgraded with support for dark mode. By default it will follow the setting on your system/browser. You may override it by scrolling to the end of the page and clicking the gears icon.

AMD Ryzen Threadripper 2970WX

W1zzard

Administrator
Staff member
Joined
May 14, 2004
Messages
28,803 (3.74/day)
Processor Ryzen 7 5700X
Memory 48 GB
Video Card(s) RTX 4080
Storage 2x HDD RAID 1, 3x M.2 NVMe
Display(s) 30" 2560x1600 + 19" 1280x1024
Software Windows 10 64-bit
AMD's Ryzen Threadripper 2970WX comes with a staggering 24-cores and 48-threads, clocked at up to 4.2 GHz. We take a closer look at application and gaming performance in this review, and test the new "Dynamic Local Mode", which automatically prioritizes busy applications.

Show full review
 
Last edited:
It really shows the difference between 2 dies and 4 dies and explains why the 2950x is better for many things.
 
needs a correction in the conclusion:

". The reason for that is the memory interface configuration, which was the only choice for these CPUs to fit on the existing X299 platform. " should be x399 no?
 
In this review, we present four results for the Ryzen Threadripper 2920X:

Should be 2970WX, on Test Setup page.
 
Chiplets FTW next year.
 
Last edited:
@Wizzard some doubts, is PBO Max and OC results in Distributed, Local or Dynamic Mode?

Now I realize why Intel didn't care at all, they knew that TR 24/32 cores will have plenty of bottlenecks and average performance will be worse than TR 12/16 cores. I hope the rendering niche market is worth it to release such an unbalanced product.

After reading this review I believe than Core X won't have price reductions any time soon.
 
@Wizzard some doubts, is PBO Max and OC results in Distributed, Local or Dynamic Mode?

Now I realize why Intel didn't care at all, they knew that TR 24/32 cores will have plenty of bottlenecks and average performance will be worse than TR 12/16 cores. I hope the rendering niche market is worth it to release such an unbalanced product.

After reading this review I believe than Core X won't have price reductions any time soon.
No it's because Intel didn't want to sell their $10k listed part cheaper, the main reason why TR2 WX chips are hobbling is because of 4 mem channels being disabled. That's not applicable to mesh based XCC chips from Intel, 20 cores & above. So like I said Intel cares for profit more than just the performance crown, that's why they've released the now infamous 28 core 5Ghz water chilled part as a $4k Xeon.
 
@Wizzard some doubts, is PBO Max and OC results in Distributed, Local or Dynamic Mode?

Now I realize why Intel didn't care at all, they knew that TR 24/32 cores will have plenty of bottlenecks and average performance will be worse than TR 12/16 cores. I hope the rendering niche market is worth it to release such an unbalanced product.

After reading this review I believe than Core X won't have price reductions any time soon.

The "average performance" is due to the die count, 4 dies have more latency than 2 dies. Plus, the bottlenecks are software not hardware. As I said on the other review, game selection favors single core performers hance why Intel did better on those charts. AMD just doesnt do well in single core apps and games, which comes back around to lazy or inept programmers not trying to optimize for more cores.
 
No it's because Intel didn't want to sell their $10k listed part cheaper, the main reason why TR2 WX chips are hobbling is because of 4 mem channels being disabled. That's not applicable to mesh based XCC chips from Intel, 20 cores & above. So like I said Intel cares for profit more than just the performance crown, that's why they've released the now infamous 28 core 5Ghz water chilled part as a $4k Xeon.

Any Skylake X has 4 memory channels connected to the die, instead of 2 or 0 as in TR WX, that is why they are hobbling, even with 8 memory channels it will be just 2 connected to each die.

Of course they care for the performance crown, any Skylake X from 14 cores and above is faster in average than any TR. I think that the new 28 core Xeon will be way faster than 2990WX.

The "average performance" is due to the die count, 4 dies have more latency than 2 dies. Plus, the bottlenecks are software not hardware. As I said on the other review, game selection favors single core performers hance why Intel did better on those charts. AMD just doesnt do well in single core apps and games, which comes back around to lazy or inept programmers not trying to optimize for more cores.

Same as above, any Skylake X has 4 memory channels connected to the die, instead of 2 or 0 as in TR WX, that is why Skylake X parts don't suffer from latency as TR.

If programmers don't want more cores, then AMD should make faster ones and stop sailing against the current.
 
Any Skylake X has 4 memory channels connected to the die, instead of 2 or 0 as in TR WX, that is why they are hobbling, even with 8 memory channels it will be just 2 connected to each die.

Of course they care for the performance crown, any Skylake X from 14 cores and above is faster in average than any TR. I think that the new 28 core Xeon will be way faster than 2990WX.



Same as above, any Skylake X has 4 memory channels connected to the die, instead of 2 or 0 as in TR WX, that is why Skylake X parts don't suffer from latency as TR.

If programmers don't want more cores, then AMD should make faster ones and stop sailing against the current.
The Xeon dies, from which Skylake is derived, has 6 mem channels & ECC among other things. For the consumer platform 2 channels & ECC are disabled across the board i.e. LCC & HCC dies. It's the same as TR & EPYC, AMD chose to disable 4 memory channels probably because 6 mem channels couldn't work with 4 dies.

No it's not, have you looked at reviews? In applications making heavy use of AVX(2 or 512) I expect Intel to come out on top more often than not, elsewhere AMD should win most of the times.
102297.png
102300.png

102295.png
102296.png

102293.png
102294.png


Yet Skylake X has worse latency than Skylake, this is a feature not a bug for higher core count CPUs & Intel will face the same issues as AMD if/when they go with the MCM approach.

Not every programmer wants less cores, are we making assumptions now?
 
Intel and AMD are giants in CPU manufacturing much like Barcelona and Real Madrid are in Spanish Football, and they have completely different "styles":

- Intel is using monolithic die approach while AMD has changed into multi-chip die approach (albeit recently)
- Barcelona uses passing around the ball until they manage to break through opposite defense, as a team or with an individual brilliance, while Real Madrid uses a more "brute force" and counter-attacks much more often

Both approaches have their own merits and their own drawbacks:

- Intel's generally have better latency and single thread performance but have problems with scaling to high core counts while AMD have high core counts but have latency and single thread issues
- Barcelona ability to circulate the ball enables them to frustrate the opposite side into making mistakes and those end up costing the opponents dearly while Real Madrid scores many goals (until recently, LOL) but has issues with defense because of the way they attack and that enables opponents to score goals which sometimes leads ties or even losses.

CPU manufacturing has reached a point where they either change the materials the chips are made of and are finally able to have big leaps in speed or find ways to have more cores. AMD has already gone the more cores route while Intel hasn't, for now.

As @R0H1T has said, unless whatever being run makes heavy use of AVX, doesn't scale well with cores or is heavily taxing on a single thread, then this 2970WX totally beats Intel's current offerings.
 
The Xeon dies, from which Skylake is derived, has 6 mem channels & ECC among other things. For the consumer platform 2 channels & ECC are disabled across the board i.e. LCC & HCC dies. It's the same as TR & EPYC, AMD chose to disable 4 memory channels probably because 6 mem channels couldn't work with 4 dies.

No it's not, have you looked at reviews? In applications making heavy use of AVX(2 or 512) I expect Intel to come out on top more often than not, elsewhere AMD should win most of the times.

Yet Skylake X has worse latency than Skylake, this is a feature not a bug for higher core count CPUs & Intel will face the same issues as AMD if/when they go with the MCM approach.

Not every programmer wants less cores, are we making assumptions now?

All the benchmarks you posted are rendering (without counting 7-zip) that I already agreed that is where TR shines. But most 95% other apps are slower on TR.
relative-performance-cpu.png

Well I can assume that since 2970WX is slower on average than 2920X or 9900K, yes most programmers prefer less cores but faster ones.

As @R0H1T has said, unless whatever being run makes heavy use of AVX, doesn't scale well with cores or is heavily taxing on a single thread, then this 2970WX totally beats Intel's current offerings.

Which are 95% of apps.
 
Same as above, any Skylake X has 4 memory channels connected to the die, instead of 2 or 0 as in TR WX, that is why Skylake X parts don't suffer from latency as TR.

If programmers don't want more cores, then AMD should make faster ones and stop sailing against the current.
Thats pretty ignorant.

Since I'm not a fanboi of either side, I will just say all multicore cpus have latency, its different for each chip. AMD is still refining its "magic fabric".
 
Thats pretty ignorant.

Since I'm not a fanboi of either side, I will just say all multicore cpus have latency, its different for each chip. AMD is still refining its "magic fabric".

Yes that is true, but AMD TR has much more latency than Intel. If you mean Infinity Fabric, they should increase its max transfer rate from 25 G/s to at least 50. Also they should wire all 4 memory channels to each die, instead of 2 or 0. It will be more expensive but some users will be willing to pay for it.
 
All the benchmarks you posted are rendering (without counting 7-zip) that I already agreed that is where TR shines. But most 95% other apps are slower on TR.
View attachment 109640
Well I can assume that since 2970WX is slower on average than 2920X or 9900K, yes most programmers prefer less cores but faster ones.


Which are 95% of apps.
True for ST performance, the exact opposite for AVX.
Yes that is true, but AMD TR has much more latency than Intel. If you mean Infinity Fabric, they should increase its max transfer rate from 25 G/s to at least 50. Also they should wire all 4 memory channels to each die, instead of 2 or 0. It will be more expensive but some users will be willing to pay for it.
Except it's not as simple as that.
IF%20Power%202990WX_575px.png
IF%20Power%20EPYC_575px.png

IF%20Power%202700X_575px.png
IF%20Power%202950X_575px.png

Uncore%208700Ka_575px.png
Uncore%207980XE_575px.png
 
True for ST performance, the exact opposite for AVX.

Except it's not as simple as that.

Aren't most AVX apps faster on Intel too?

Sorry, I didn't get your point with the graphs, could you elaborate more please?
 
Aren't most AVX apps faster on Intel too?

Sorry, I didn't get your point with the graphs, could you elaborate more please?
If i understood correctly, he meant AMD's IF uses way more power then Intel's Uncore. It get's worse with more memory channels, as can be seen with the Epyc chart VS TR 2990WX.
 
Aren't most AVX apps faster on Intel too?

Sorry, I didn't get your point with the graphs, could you elaborate more please?
True, but the point I was making is that if 95% of apps are (mostly) single threaded then the opposite is also true wrt AVX i.e. 95% of apps don't make use of AVX including the most popular ones i.e. web browsers.

As you can see from the graphs, with increasing core count the power requirements of uncore shoots up. This is true for AMD & Intel, pretty sure the amount of power used by 28 core Xeon is disproportionately higher for it's uncore wrt their 18 core variant. The case with AMD is a bit different because their dies are limited to 4 core per CCX, so when Intel goes the MCM route you'll probably see the same charts as highlighted by AMD's HEDT.
 
  • Like
Reactions: HTC
True, but the point I was making is that if 95% of apps are (mostly) single threaded then the opposite is also true wrt AVX i.e. 95% of apps don't make use of AVX including the most popular ones i.e. web browsers.

As you can see from the graphs, with increasing core count the power requirements of uncore shoots up. This is true for AMD & Intel, pretty sure the amount of power used by 28 core Xeon is disproportionately higher for it's uncore wrt their 18 core variant. The case with AMD is a bit different because their dies are limited to 4 core per CCX, so when Intel goes the MCM route you'll probably see the same charts as highlighted by AMD's HEDT.

I think that is why Intel designed the mesh architecture, to avoid going the MCM route, at least for some years, so that 28 core Xeon is still a monolithic 28 core design, instead of 32 core TR or EPYC that are 4 dies of 2 native quad cores each all glued together. AMD can't even produce a true native 8 core. I can see on the charts that the 7980X uncore parts is between 2700X and 2950X, besides having more cores.
 
I know this isn't the main point of the review, but I gotta ask - what the hell is Far Cry 5's problem with this CPU? Across the resolutions the different modes trade places, at 1440p the difference between them is absurd, 720p yields 31% FPS less than 1080p on average - it's a mess.

If in every other game the 2970WX does a good enough job for a rendering workhorse - this being the odd exception. If the game has some thread scheduling issue, isn't Dynamic Local Mode supposed to remedy this? Or is it so badly or specifically coded that is just exposes some very rare use case scenario disadvantage, where even Dynamic Local Mode and/or overclocking can't prevent it from fumbling?

FC5_WTF.png
 
Last edited:
Back
Top