Threadripper 3990x vs 3970x

[ADVenturePO]

I think the best person to ask this would be Juraj since from what I remember from the forum threads he has the 3990x on both air and loop but I might be wrong :)

@ADVenturePO

I am pretty sure that Blender, Cinebench, and Corona benchmarks don't scale the same way. Maybe I am wrong but from what I've seen on youtube reviews Corona was the only benchmark that was giving 2xperformance versus 3970x.
I never ever ran a Blender benchmark in my life so I can't really talk about that. What I can talk about is real-world performance, and in my case, the results are not always in line with the Corona benchmark.

I won't even mention animation frame rendering (especially frames that are under 20mins) where at least in "my case" the 5950x works the best as it's always the first machine that loads the scene due to the best max turbo on one core. By the time, the 3990x has loaded the scene the 5950x is already a few passes in :).

I would like to hear the opinion of the people who ran their 3990x's on-air and then upgraded to a custom loop. If it's actually that much worth it and their use case scenario since mine is purely from a rendering node perspective :).

Dear VUK,

I am a company that already made 70+ Threadripper Workstations. Few on air, most on LC. I have been comparing times for air and LC. And I'll try to present you my findings.
The problem with 3990x is that for some operations it has to have 8 sticks of RAM - example is video recoding - HWBOT x265 benchmark with 4 sticks 270 FPS and 8 sticks 370 FPS. In rendering that is not the case but with intensive RAM operations the memory configuration is much more influencing.
Also if there's no proper cooling to 3990x its speed will be reduced - also SC Clocks. That is why 5950x which is ultrafast in SC will prepare scene in 3D Max much more faster. 3990x  is without OC at best able to speed up single core to 4.3GHz and 5950x is able to run on good cooling up to 5025Mhz and it's memory controller is not demanding at all. 2 sticks? OK because it is Dual Channel architecture.

As for now, I will show you how rendering times of frames drops on TR3 3960 on air compared to i9-7960x on LC.
alpha is i9 and beta is TR3. These are not my results but my clients, who had forced me to make his WS on Dark Rock 4 Pro.  The MB was MSI TRX40 PRO Wifi. 8 sticks of RAM - 128GB. When cooler is heated up and AMD CPU find itself heating up faster it will drop Effective Clocks.
Please see attached pictures.

Now, I'm sorry I haven't made this video in English it's in Poilish. But here you have Effective Clocks on air:
2.63 on average.

Here You have effective clocks on LC without OC: 2.875 on average

And here are OC Tests at 3.3GHz with 1300W PSU (which screams and dies a little)

You can see that average Effective Clocks are 3.24+GHz.

From my perspective and from enormous amounts of tests OC is boosting TR3 3970x and 3990x rendering speed. For 3990x the amount of boost  is really big.
I'll try to find on which testing M.2 those test are in pictures and send them here.

Cheers

[Juraj]

These are not my results but my clients, who had forced me to make his WS on Dark Rock 4 Pro

A true client from hell :- )

Using "Dark Rock 4 Pro" as example of reference Air-Cooling skews thing negatively for air-cooling Threadripper, (DR4P) it's the only sTRX socket air-cooler which has horizontally running heatpipes across heatspreader, i.e they only cover the four middle dies. It makes sense, it was designed at 1st generation of chips, still, Noctua made correct design decision at same time..

Dark Rock 4 Pro was never a suitable cooler for 3990X, it does somewhat work for 3960X & 3970X because those have dies in middle of the chip. Which is also why those can be very well cooled by any generic Asetek AIO.

The difference proper TR4 air-cooler like Noctua UH-14s TR4 and Dark Rock 4 Pro on 3990X is simply massive.
Just a reminder, these two coolers had delta of 20 C already with past-generation 2990WX, the first chip with dies that were spread across the full chip.

I've put 3990X under OptimusPC water-block, state-of-art block that is even 2-3C better than the previous winner, Heatkiller IV. After that, it doesn't matter how many radiators it is connected to (in my case 2x420mm) since the issue is heat-transfer from chip, not dissipation from the loop. At room ambient of 25C, the effective boost clocks on long renders are effectively identical between this build and the one with Noctua UH-14s.

That obviously very quickly changes with PBO/Curve Optimizer/Static Voltages/etc.. (any other OC technique), where loop or even the funky tower from Ice-Giant pull quickly ahead, by a large delta. Past 280W (and it's not hard to take 3990X into 700+ W territory), Air is no longer viable.

99perc. of users on this forum are running their PCs stock. That's matter of life for most workstations (discounting XMP/DOCP profile).

Ultimately, my personal suggestion comes down to budget. If you can afford 3990X, get it, even if it scales very little in certain workloads, or particular working conditions.

3970X might indeed be more prudent, more universal option, it does maintain better boost clocks with simple math of diving its 280W power budget only between 4 Dies + 1 MC, a benefit which is partly negatively offset by worse cooling (4 centered dies are harder to cool than 8 spread-out dies).

TL:DR :

Buy the one you won't regret spending budget for.
If you have sizeable render-farm, your choice is more partial towards 3970X. If you have only single-workstation that acts as your main rendering machine as well, absolutely go for 3990X.
Air cooling is absolutely viable, zero-issue choice for 3990X. But you have to choose either Noctua NH-U14s TR4, or Ice-Giant. You can ignore the existence of everything else.

Few graphs to show how much Dark Rock Pro 4 sucks, and why it should never be used as argument against air-cooling Threadripper.
Also article from KitGuru that explains the failed design of this cooler.

https://www.kitguru.net/components/cooling/luke-hill/threadripper-3990x-cpu-cooling-comparison-how-to-tame-the-beast/11/

[ADVenturePO]

These are not my results but my clients, who had forced me to make his WS on Dark Rock 4 Pro

A true client from hell :- )

Using "Dark Rock 4 Pro" as example of reference Air-Cooling skews thing negatively for air-cooling Threadripper, (DR4P) it's the only sTRX socket air-cooler which has horizontally running heatpipes across heatspreader, i.e they only cover the four middle dies. It makes sense, it was designed at 1st generation of chips, still, Noctua made correct design decision at same time..

Dark Rock 4 Pro was never a suitable cooler for 3990X, it does somewhat work for 3960X & 3970X because those have dies in middle of the chip. Which is also why those can be very well cooled by any generic Asetek AIO.

The difference proper TR4 air-cooler like Noctua UH-14s TR4 and Dark Rock 4 Pro on 3990X is simply massive.
Just a reminder, these two coolers had delta of 20 C already with past-generation 2990WX, the first chip with dies that were spread across the full chip.

I've put 3990X under OptimusPC water-block, state-of-art block that is even 2-3C better than the previous winner, Heatkiller IV. After that, it doesn't matter how many radiators it is connected to (in my case 2x420mm) since the issue is heat-transfer from chip, not dissipation from the loop. At room ambient of 25C, the effective boost clocks on long renders are effectively identical between this build and the one with Noctua UH-14s.

That obviously very quickly changes with PBO/Curve Optimizer/Static Voltages/etc.. (any other OC technique), where loop or even the funky tower from Ice-Giant pull quickly ahead, by a large delta. Past 280W (and it's not hard to take 3990X into 700+ W territory), Air is no longer viable.

99perc. of users on this forum are running their PCs stock. That's matter of life for most workstations (discounting XMP/DOCP profile).

Ultimately, my personal suggestion comes down to budget. If you can afford 3990X, get it, even if it scales very little in certain workloads, or particular working conditions.

3970X might indeed be more prudent, more universal option, it does maintain better boost clocks with simple math of diving its 280W power budget only between 4 Dies + 1 MC, a benefit which is partly negatively offset by worse cooling (4 centered dies are harder to cool than 8 spread-out dies).

TL:DR :

Buy the one you won't regret spending budget for.
If you have sizeable render-farm, your choice is more partial towards 3970X. If you have only single-workstation that acts as your main rendering machine as well, absolutely go for 3990X.
Air cooling is absolutely viable, zero-issue choice for 3990X. But you have to choose either Noctua NH-U14s TR4, or Ice-Giant. You can ignore the existence of everything else.

Few graphs to show how much Dark Rock Pro 4 sucks, and why it should never be used as argument against air-cooling Threadripper.
Also article from KitGuru that explains the failed design of this cooler.

https://www.kitguru.net/components/cooling/luke-hill/threadripper-3990x-cpu-cooling-comparison-how-to-tame-the-beast/11/

Please check it out.
Thank You.

I do not care that you have seen pictures on the Internet so you will make a dispute with a guy that built 60+ TR3 Workstations.

Is there any other proof that you need?

[Vuk]

@ADVenturePO

I see your point and thanks for sharing these videos they are informative. I also agree with what you say about not testing short and fast benchmarks like the Corona one but I am also sure that the performance in Corona vs Blender is not the same. And that the cpu effective clocks don't behave the same. I can't say with certainty because I don't use Blender myself but from all the benchmarks seen thus far with different types of processors they just don't seem to scale the same way ( Corona and the Blender rendering engine ).


I ordered an Ice Giant cooler a few days ago and I should be getting it any day. Will conduct the same tests you did only in Corona and in a real-world scenario using a big project scene to see what is the difference. As soon as I have something interesting I will post it on the forum when I grab some time :).

[Majeranek]

@ADVenturePO

There's one thing that I'm wondering about.
In first scenario where you have air cooling and temperatures are quite comfortable but effective clocks are still lower. Why is that happening?

[ADVenturePO]

Dear All,

I cannot provide the exact answer why Threadrippers behave differently on different cooling as they do.
I'm just a guy that is sitting all the time with them and making Workstations.
I'm an overclocker who after only 4,5 years in 2020 was no 1 in ambient cooling - Enthusiast Group on HWBOT.

Here You have another 3990x Workstation. Effective Clocks are again much higher than on air cooling.

 

From my observations - the faster the TR is heating up the lower the Effective Clock. It's a cheeky mechanism that lets user to go with almost any bad cooling.
The real answer if cooling is good is the outcome of a long benchmark.

That is all I can really say.

[Majeranek]

Hi guys,

I know that it's been a while here but I've just built my 3990x workstation.
I was curious about those effective clocks on air cooling which I have - Noctua UH-14s with 2 fans installed and everything else is just a stock build without any OC.

Temps were pretty comfy with max 61 on the CPU and max 69 on the CPU Package and Averege Effective Clock was around 3,24GHz.
Everything was measured during the Aida64 Stress Test (CPU, FPU, Cache, Memory).

I think it's just a matter to choose right air cooling like Juraj has written.

Cheers!

[Nejc Kilar]

@Majeranek

Thank you for sharing! Just out of curiosity, could you give the Corona benchmark a couple of concurrent run and report back the effective clock numbers? Just interested in seeing what you are getting there :)

[Majeranek]

@Majeranek

Thank you for sharing! Just out of curiosity, could you give the Corona benchmark a couple of concurrent run and report back the effective clock numbers? Just interested in seeing what you are getting there :)

In the Corona Benchmark the effective clock number reaches 3,05GHz (too quick to catch average here). Windows task manager shows that utilization of CPU varies between 91-97%.

I've made simple scene to test it on the production rendering and here it averages 3,29GHz and utilization is 100%. However while using IR on the same scene the utilization is again between 91-97% and cpu clocks 2,75Ghz on average.

I think we need a new benchmark ;). The old one is too simple for today CPUs.

[Basshunter]

@ADVenturePO

I see your point and thanks for sharing these videos they are informative. I also agree with what you say about not testing short and fast benchmarks like the Corona one but I am also sure that the performance in Corona vs Blender is not the same. And that the cpu effective clocks don't behave the same. I can't say with certainty because I don't use Blender myself but from all the benchmarks seen thus far with different types of processors they just don't seem to scale the same way ( Corona and the Blender rendering engine ).


I ordered an Ice Giant cooler a few days ago and I should be getting it any day. Will conduct the same tests you did only in Corona and in a real-world scenario using a big project scene to see what is the difference. As soon as I have something interesting I will post it on the forum when I grab some time :).

Hey Vok. Just wondering if you managed conduct the test with Corona. I've been following the thread and I'm really interested in the results you could provide. Thanks.