AMD Ryzen 3990X review

Started by KlausK, May 28, 2020, 10:38:22 AM

Previous topic - Next topic

KlausK

Since AMD Ryzen CPUs seem to draw a lot of attention these days this might be of interest to someone thinking about upgrading:

http://www.cgchannel.com/2020/05/review-amd-ryzen-threadripper-3990x/
CHeers, Klaus
/ ASUS WS Mainboard / Dual XEON E5-2640v3 / 64GB RAM / NVIDIA GeForce GTX 1070 TI / Win7 Ultimate . . . still (||-:-||)

Dune

Thanks Klaus. Seems overkill, and a little bit too costly. If I upgrade (and I will), I need to have something that is decently priced and decently fast, and also decently quiet ;)

WAS

I won't be going for a thread ripper anytime soon. The single-core performance isn't the best, and with TG will only have issues with core count.

Ulco, the Ryzen 5 3600x is at a really nice price right now for one of the most powerful 6 core 12 thread CPUs. The Ryzen 7 3700x 8 core 16 thread CPU isn't much more than the 3600x too, and much better. I'm probably going to go with the 3900x or 3950x when I can afford it. Really want the 3950x for the 16 cores and 32 theats. :)

Dune


D.A. Bentley (SuddenPlanet)

Here is what I just got done building.  It's a 24-Core Threadripper, and I turn OFF the AMD Simultaneous Multi Threading (SMT) in the BIOS so I don't have 48 threads (don't need or want that many).  The real-time preview seems to only use 16 threads efficiently, and I work with clouds a lot in the RTP.  It has a mild overclock so it pops up to 4.1GHz when at Full Load, and when less cores are being used it goes up as high as 4.3GHz.

Here are the specs I chose:
Mobo:  Gigabyte TRX40 AORUS PRO WiFi
CPU:  AMD Ryzen Threadripper 3960X 24-Core 3.8 GHz
RAM: Corsair VENGEANCE LPX 128GB (8x16GB) DDR4 3200 (PC4-25600) PC Memory Black (CMK128GX4M8X3200C16)
SSD Storage:  Samsung (MZ-V7E1T0BW) 970 EVO SSD 1TB - M.2 NVMe
CASE:  Fractal Design Define R6 USB-C Blackout
PSU:  EVGA Supernova 1000 P2 80+ Platinum, 1000W
CPU Cooler:  EVGA CLC 360mm All-in-one RGB LED CPU Liquid Cooler
VIDEO:  EVGA GeForce GTX 1080 Ti FTW3 GAMING, 11G-P4-6696-KR, 11GB GDDR5X, iCX


From doing a few benchmarks It's almost 3X as fast as my older Xeon E5-2697v2 @2.7GHZ (12-Core)

:)

-Derek

WAS

Quote from: D.A. Bentley on June 03, 2020, 06:44:01 PMHere is what I just got done building.  It's a 24-Core Threadripper, and I turn OFF the AMD Simultaneous Multi Threading (SMT) in the BIOS so I don't have 48 threads (don't need or want that many).  The real-time preview seems to only use 16 threads efficiently, and I work with clouds a lot in the RTP.  It has a mild overclock so it pops up to 4.1GHz when at Full Load, and when less cores are being used it goes up as high as 4.3GHz.

Here are the specs I chose:
Mobo:  Gigabyte TRX40 AORUS PRO WiFi
CPU:  AMD Ryzen Threadripper 3960X 24-Core 3.8 GHz
RAM: Corsair VENGEANCE LPX 128GB (8x16GB) DDR4 3200 (PC4-25600) PC Memory Black (CMK128GX4M8X3200C16)
SSD Storage:  Samsung (MZ-V7E1T0BW) 970 EVO SSD 1TB - M.2 NVMe
CASE:  Fractal Design Define R6 USB-C Blackout
PSU:  EVGA Supernova 1000 P2 80+ Platinum, 1000W
CPU Cooler:  EVGA CLC 360mm All-in-one RGB LED CPU Liquid Cooler
VIDEO:  EVGA GeForce GTX 1080 Ti FTW3 GAMING, 11G-P4-6696-KR, 11GB GDDR5X, iCX


From doing a few benchmarks It's almost 3X as fast as my older Xeon E5-2697v2 @2.7GHZ (12-Core)

:)

-Derek

That's pretty nice performance on those cores with SMT off. I guess would be pretty much fine on that except for some games that try to balance loads across all cores and actually suffer because of it. I think Tomb Raider has this issue.

Dune

Cool, Derek. Thanks for posting your specs. That might be something for me too. Is it a bit quiet or roaring off at full load? And what price range are we talking about roughly? I am hesitating between something a bit less speedy, but maybe quieter and (a lot) cheaper, or go for the ultimate.
Regarding such large SSD; I heard they have a limited timespan, less than a normal HD, so a lot of folks opt for a smaller SSD just for software (speedy startup) and a big normal HD for storage. Or has that changed in the meantime?

Nala1977

Quote from: D.A. Bentley on June 03, 2020, 06:44:01 PMHere is what I just got done building.  It's a 24-Core Threadripper, and I turn OFF the AMD Simultaneous Multi Threading (SMT) in the BIOS so I don't have 48 threads (don't need or want that many).  The real-time preview seems to only use 16 threads efficiently, and I work with clouds a lot in the RTP.  It has a mild overclock so it pops up to 4.1GHz when at Full Load, and when less cores are being used it goes up as high as 4.3GHz.

Here are the specs I chose:
Mobo:  Gigabyte TRX40 AORUS PRO WiFi
CPU:  AMD Ryzen Threadripper 3960X 24-Core 3.8 GHz
RAM: Corsair VENGEANCE LPX 128GB (8x16GB) DDR4 3200 (PC4-25600) PC Memory Black (CMK128GX4M8X3200C16)
SSD Storage:  Samsung (MZ-V7E1T0BW) 970 EVO SSD 1TB - M.2 NVMe
CASE:  Fractal Design Define R6 USB-C Blackout
PSU:  EVGA Supernova 1000 P2 80+ Platinum, 1000W
CPU Cooler:  EVGA CLC 360mm All-in-one RGB LED CPU Liquid Cooler
VIDEO:  EVGA GeForce GTX 1080 Ti FTW3 GAMING, 11G-P4-6696-KR, 11GB GDDR5X, iCX


From doing a few benchmarks It's almost 3X as fast as my older Xeon E5-2697v2 @2.7GHZ (12-Core)

:)

-Derek
im very confused on why did you turn off the SMT, did you try to run the processor with SMT on? why even buy a ryzen with 48 threads if you just want half of it?

D.A. Bentley (SuddenPlanet)

Quote from: Nala1977 on June 04, 2020, 09:09:25 AMim very confused on why did you turn off the SMT, did you try to run the processor with SMT on? why even buy a ryzen with 48 threads if you just want half of it?
I have noticed that at the end of large (8,000x4,000+) high quality renders with water in the scene Terragen will have a few buckets taking way longer than the rest of the image near the horizon and each bucket just uses one thread.  With SMT/Hyper-threading on a core is split into two threads, and that core is then only being used at 50%, or at least that's how it appears in Windows Performance Monitor.  It's possible Windows Performance Monitor is not correctly displaying the core usage.  I did do some tests in the past to determine if there is a performance difference, but it was so long ago I can't remember now, but there is one other reason I limited my threads to just 24 (One per Core).  The RTP seems to only fully use 16 threads, and anything more than 16 get partially used (maybe 90%).  Matt could probably say more on this.

Back to the "straggler bucket" issue... One method I discovered that can help alleviate this inefficiency is to render your frame upside down(flip camera/-180 on Z), so Terragen starts rendering the more complicated part (water) of the render first, and ends with faster rendering stuff at the end.  This totally works, but only if you have done some test renders to verify there are straggler buckets in the lower half of your image.  I do a lot of Spherical Camera renderings (Equirectangular Panoramas) for Skyboxes in games, or IBL for Ray-Tracing Rendering Engines.  So pretty much every image I render has an above horizon and below horizon area, and if lakes are in the scene it's a given the lower half will always take longer if the quality is high.

-Derek

D.A. Bentley (SuddenPlanet)

Quote from: Dune on June 04, 2020, 02:36:15 AMCool, Derek. Thanks for posting your specs. That might be something for me too. Is it a bit quiet or roaring off at full load? And what price range are we talking about roughly? I am hesitating between something a bit less speedy, but maybe quieter and (a lot) cheaper, or go for the ultimate.
Regarding such large SSD; I heard they have a limited timespan, less than a normal HD, so a lot of folks opt for a smaller SSD just for software (speedy startup) and a big normal HD for storage. Or has that changed in the meantime?

I'm not worried about SSD lifespans.  By the time an SSD fails I'll be dead.  hah.  The 970 EVO is rated at 1.5 Million Hours, which is 171 years.
If by chance it did prematurely fail I routinely make System Images to backup my OS drive, and I use Backblaze for everything else including my OS drive.

As far as sound, the machine can be as quiet as I want using fan control software.  Or I can rev it up to sound like a leaf blower.  Even on quiet settings the 360mm RAD keeps the CPU plenty cool.

Price range for a system like this will be $3,000 - $4,000.  You could save money by going with less RAM, and a lower end video card, like an RTX 2060.
I already had a GTX 1080 Ti, so I put that in my new system and got the RTX 2060 as a replacement.  You don't need a fast/powerful GPU like the GTX 1080 Ti, or RTX 2080 for Terragen, but I also run software that does use it like Modo/Blender.

-Derek

jaf

Quote from: D.A. Bentley on June 04, 2020, 12:15:44 PM
Quote from: Nala1977 on June 04, 2020, 09:09:25 AMim very confused on why did you turn off the SMT, did you try to run the processor with SMT on? why even buy a ryzen with 48 threads if you just want half of it?
I have noticed that at the end of large (8,000x4,000+) high quality renders with water in the scene Terragen will have a few buckets taking way longer than the rest of the image near the horizon and each bucket just uses one thread.  With SMT/Hyper-threading on a core is split into two threads, and that core is then only being used at 50%, or at least that's how it appears in Windows Performance Monitor.  It's possible Windows Performance Monitor is not correctly displaying the core usage.  I did do some tests in the past to determine if there is a performance difference, but it was so long ago I can't remember now, but there is one other reason I limited my threads to just 24 (One per Core).  The RTP seems to only fully use 16 threads, and anything more than 16 get partially used (maybe 90%).  Matt could probably say more on this.

Back to the "straggler bucket" issue... One method I discovered that can help alleviate this inefficiency is to render your frame upside down(flip camera/-180 on Z), so Terragen starts rendering the more complicated part (water) of the render first, and ends with faster rendering stuff at the end.  This totally works, but only if you have done some test renders to verify there are straggler buckets in the lower half of your image.  I do a lot of Spherical Camera renderings (Equirectangular Panoramas) for Skyboxes in games, or IBL for Ray-Tracing Rendering Engines.  So pretty much every image I render has an above horizon and below horizon area, and if lakes are in the scene it's a given the lower half will always take longer if the quality is high.

-Derek

Couldn't you set the maximum number of threads to 16 or 24 in the Render|Advanced tab?  Then you would have the other threads available if needed without having to go to the bios.
(04Dec20) Ryzen 1800x, 970 EVO 1TB M.2 SSD, Corsair Vengeance 64GB DDR4 3200 Mem,  EVGA GeForce GTX 1080 Ti FTW3 Graphics 457.51 (04Dec20), Win 10 Pro x64, Terragen Pro 4.5.43 Frontier, BenchMark 0:10:02

D.A. Bentley (SuddenPlanet)

You could do that, but that would defeat the purpose of having 24 cores.  ;)

WAS

The cores TG uses would still be at 50% total load because they're still split by multi-threading. TG also wouldn't be aware of what it's grabbing, or able to choose. It could bind in to many virtual threads. I'm not positive but I think this effects multitasking performance.

Dune

Thanks very much, Derek! Good info to have.

D.A. Bentley (SuddenPlanet)

#14
I did some benchmark testing to see how SMT on/off affects rendering speed.

Just FYI:

AMD = SMT (Simultaneously Multi Threading)
Intel = Hyper Threading

So based on the render results from my benchmark tests my understanding of how SMT/Hyper Threading works might have been wrong.
First here are the test results using the TG4 Benchmark Scene:

Terragen version    |      SMT ON   |      SMT OFF  |    Diff
Terragen v4.4.67    |      3:37          |      4:11          |    34s
Terragen v4.4.46    |      3:23          |      3:58          |    35s

I attached some screen shots showing the results with SMT ON.  Wasn't taking screen shots with SMT off because I didn't know I was going to post this.

Those results are based on just one scene though.  Further testing with scenes with more water and more complicated clouds could produce different results, but I have some guesses on how SMT/Hyper Threading works but I don't know how to test (maybe a programmer could make a test).

But basically my gut is telling me even though each and every core is split into two threads when SMT is on, it may be that if only one thread on a core is being used (and at 100%), that core might still be getting 100% utilization.  Windows 10 Performance Monitor might be just providing a visual representation that is not quite accurate.  This is just a guess, but if a programmer could make a test that specifically targets specific threads, and leaves others unused this could determine the result I am looking for.

For example, if you could on my 24-core CPU target Thread 0,2,4,6,8,10,12,14,16,18,20,22 as the group of threads to use but leave the odd threads unused. I would expect if my theory is correct that the rendering results would be similar to if SMT was OFF which just a small performance decrees (not 50% less performance).

I also captured a Perf Monitor shot (attached) of when I was using RTP with the TG4 Benchmark scene.  Notice how the first 32 threads are at 100%, and the rest are somewhere around 80%.  When I had SMT off I saw the same representation but with 16 threads at 100% and the remaining 8 at about 80%

So for now I think I am going to leave SMT on, and if I ever encounter any scenes that have really long hangs on a few straggler buckets I'll try and negate that by rendering upside down.


-Derek