Benchmark your computer @4K with Handbrake 1.1 and H265! NOW WITH ZEN3!

Page 11 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Fir

Senior member
Jan 15, 2010
484
194
116
Mark: I've noticed this as well. It doesn't seem to be optimized to use that many threads. CB certainly does! One of my 3970s is hitting high 18k with close to 4.2GHz sustained across all 32 cores. That one has 256GB DDR4-3200MHz CL16 (8x32). Amazing kit for the price. Not long ago 256GB was a lot for a hard drive. :D
 
  • Like
Reactions: lightmanek

YBS1

Golden Member
May 14, 2000
1,945
129
106
This benchmark is highly flawed on todays high core count processors. My 64 core Rome processor was so "all over the place" that I guessed at 3 ghz, but it only used less than 50% of the power. I only got a little over 11 fps, so you beat me with 16 cores ! CB20 was much better in using all the cores.
Yeah, that's what drew my attention to begin with. I was sitting there watching Ryzen Master and you could pretty much see the cores "racing" to go to sleep. Like the encoder wasn't feeding them enough data to keep them busy. Most of the cores never did hit the full 4.2GHz during the encode.
 

Yeroon

Member
Mar 19, 2017
123
57
71
This benchmark is highly flawed on todays high core count processors. My 64 core Rome processor was so "all over the place" that I guessed at 3 ghz, but it only used less than 50% of the power. I only got a little over 11 fps, so you beat me with 16 cores ! CB20 was much better in using all the cores.

Have you considered trying it with linux and seeing what you get?
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,478
14,434
136
Have you considered trying it with linux and seeing what you get?
Does it work in linux ? Actually, I don't spend most of my time benchmarking my new rigs, just one or two before I set them off to work (in linux actually)
 

Bavor

Member
Nov 11, 2001
82
18
81
That 4.2GHz all core result is just about as identical to my 3.98GHz core average PBO result as it can be. I wonder if it's purely down to my memory speed advantage making up the 200MHz per core disadvantage or if it's something else (Infinity Fabric, motherboard differences, BIOS settings, etc.)? I'm going to run a 4.2 all core run and maybe a run with my memory brought down to 3600MHz and see what happens.

It could be UEFI/BIOS settings and memory speed. I haven't changed much except enabling XMP, enabling PBO +200, and raising the power limits. I'm sure there are other settings that I could change to speed things up.

I also noticed with PBO +200 that the core speeds vary during the handbrake test between 3.974 GHz and 4.1+ GHz. If I change the maximum speed of the cores to different speeds in Ryzen master, then more cores try to go idle during the test.
 

Hulk

Diamond Member
Oct 9, 1999
4,191
1,975
136
The Handbrake documentation reads that Handbrake will "scale well up to 6 cores and then with diminishing returns thereafter."
So on one hand it's technically not a great bench marking tool. But on the other hand it is widely used so practically it is a good bench marking tool.
Our results seem to suggest that if Handbrake is important to you then 12 to 16 cores with high clocks will get you optimum performance in Handbrake's current state of development.
 
Last edited:

Yeroon

Member
Mar 19, 2017
123
57
71
Does it work in linux ? Actually, I don't spend most of my time benchmarking my new rigs, just one or two before I set them off to work (in linux actually)

Handbrake installed 1.1.0 from APT for a 18.04 based install. Small package and didnt need to do anything else.
 

Micrornd

Golden Member
Mar 2, 2013
1,278
178
106
This benchmark is highly flawed on todays high core count processors. My 64 core Rome processor was so "all over the place" that I guessed at 3 ghz, but it only used less than 50% of the power. I only got a little over 11 fps, so you beat me with 16 cores ! CB20 was much better in using all the cores.
@Markfw - if you have time could you please run the benchmark again as I have in this post -
I 'm interested in seeing how EPYC CPUs perform when actually used for real workstation duties, not single tasks.
Maybe also with your second EPYC when you get it ?
Thanks !
 
Last edited:

HutchinsonJC

Senior member
Apr 15, 2007
465
201
126
That 4.2GHz all core result is just about as identical to my 3.98GHz core average PBO result as it can be. I wonder if it's purely down to my memory speed advantage making up the 200MHz per core disadvantage or if it's something else (Infinity Fabric, motherboard differences, BIOS settings, etc.)? I'm going to run a 4.2 all core run and maybe a run with my memory brought down to 3600MHz and see what happens.

In the cinebench thread and working with @Det0x he pointed out something to me about 1:1 infinity fabric and RAM. I actually bought two sets of RAM for my 3950x rig. The first set was a 2dimm 32GB 3600MHz C16 set and the other, which I bought maybe a couple or three days later because I thought it'd perform better (and before the rig was pieced together) was a 2dimm 16GB 4266MHz C19 set.

I started with the 16GB 4266MHz set and only just Sunday night swapped to the 32GB 3600MHz set as a result of the Cinebench thread. On a source I was encoding in vidcoder, I was getting ~15fps encode speed on the 4266MHz RAM. After I swapped to the 3600MHz RAM I was getting ~23fps. Cinebench numbers only went up like 100 points by comparison.

I'm confused by the numbers I saw, not sure if it even makes sense to me, and I'm pretty tempted to swap RAM back around for further testing, as it's not like I documented everything in a spreadsheet. It's all in my head and the only real programs I used were Cinebench, Vidcoder, PassMark Performance Test, and Ryzen Master. The memory scores were a good bit higher in PassMark. I don't remember the scores, but I remember the 4266MHz test said I was like 93 percentile and the 3600MHz RAM said 99 percentile.

I would say Ryzen folks should look at their Ryzen Master at the RAM area and see if it's coupled or decoupled.

The infinity fabric basically maxes out at 1833/1866 as near as I can tell, for most people. So RAM at 3600 to 3733 with low timings, I'm thinking is a sweet spot.

Here's a couple things I was reading yesterday:


 
  • Like
Reactions: lightmanek

moinmoin

Diamond Member
Jun 1, 2017
4,933
7,619
136
@HutchinsonJC If left on auto the IF frequency (fClk to mClk) is 1:2 of the RAM starting at DRR4 3800. The ratio can be set manually as well, but I'm not aware of tests where somebody successfully used a ratio of 1:1 with DDR4 4266 RAM.
 
  • Like
Reactions: HutchinsonJC

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,478
14,434
136
@HutchinsonJC If left on auto the IF frequency (fClk to mClk) is 1:2 of the RAM starting at DRR4 3800. The ratio can be set manually as well, but I'm not aware of tests where somebody successfully used a ratio of 1:1 with DDR4 4266 RAM.
I think 3766 (or is it 3733) is the max. Otherwise any motherboard will go to 1:2.
 

HutchinsonJC

Senior member
Apr 15, 2007
465
201
126
Yeah, I think the default is 1800MHz on IF, but a lot are getting 1866 without too much trouble, mostly. So 3733MHz RAM.
 

thesmokingman

Platinum Member
May 6, 2010
2,307
231
106
Mark: I've noticed this as well. It doesn't seem to be optimized to use that many threads. CB certainly does! One of my 3970s is hitting high 18k with close to 4.2GHz sustained across all 32 cores. That one has 256GB DDR4-3200MHz CL16 (8x32). Amazing kit for the price. Not long ago 256GB was a lot for a hard drive. :D

Handbrake is limited in multicore threading, however you can run multiple instances of Handbrake. With 32 cores you could run simultaneous tests and still lose only a few % from a single run. Try it out.
 
  • Like
Reactions: Fir

Fir

Senior member
Jan 15, 2010
484
194
116
Good suggestion. That will come in handy for 64 core systems. (and beyond!) ;)
 

Hitman928

Diamond Member
Apr 15, 2012
5,177
7,628
136
@Markfw Can you try adding 'pmode' to the additional options to see if that helps better utilize your high core count CPU?
 

Kenmitch

Diamond Member
Oct 10, 1999
8,505
2,248
136
I think 3766 (or is it 3733) is the max. Otherwise any motherboard will go to 1:2.
Yeah, I think the default is 1800MHz on IF, but a lot are getting 1866 without too much trouble, mostly. So 3733MHz RAM.

Not sure what the default is, but my MB will keep 1:1 ratio up to 1900MHz. I run my ram at 1:1 3800 CL15 on my MSI x570 MB. Anything beyond 1900MHz and the 1:2 kicks in and kills my latency/performance. It kind of sucks as my b-die will go way beyond 3800MHz if I loosen the timings. I stopped testing at 4333MHz CL17 which had lower bandwidth and high latency of 68.9ns compared to my 3800 CL15 results with 62.8ns and higher bandwidth. You're on your own when the ram exceeds the Dram Calculators suggested settings. I played around a little bit trying to wing it on my own at 4333MHz, but came to the conclusion it was a battle to be lost in the end anyways and thru in the towel.
 
  • Like
Reactions: HutchinsonJC

HutchinsonJC

Senior member
Apr 15, 2007
465
201
126
Yeah, I think we're almost better off leaving things 1:1 (3600MHz RAM up to 3800MHz RAM motherboard depending?). I wonder at what point RAM frequency/timing can make up for the decoupling between IF and RAM.
 

Bavor

Member
Nov 11, 2001
82
18
81
I have a Fractal Design Define R6 case for my system. I decided to try the handbrake test with the door open after seeing the reviews of the Define R7 where there was a significant temperature drop in CPU(~10C) torture tests when they ran the test with the door open compared to having the door closed.

3950X PBO+200
32GB DDR4 3600 MHz
Windows 10 Pro
Files were on a 2TB Micron SATA III SSD if that matters

Door Closed:
HandBrake 1.1.0 (2018040700) - 64bit
OS: Microsoft Windows NT 10.0.18363.0 - 64bit
CPU: AMD Ryzen 9 3950X 16-Core Processor
Ram: 32685 MB,
GPU Information:
NVIDIA GeForce RTX 2080 - 26.21.14.4187
encoded 1806 frames in 169.73s (10.64 fps), 11820.04 kb/s, Avg QP:29.09

Door Open:
HandBrake 1.1.0 (2018040700) - 64bit
OS: Microsoft Windows NT 10.0.18363.0 - 64bit
CPU: AMD Ryzen 9 3950X 16-Core Processor
Ram: 32685 MB,
GPU Information: NVIDIA GeForce GTX 1080 Ti - 26.21.14.4187
encoded 1806 frames in 164.87s (10.95 fps), 11820.04 kb/s, Avg QP:29.09

The increase was 0.31 FPS in encoding speed using PBO+200 with raised CPU power limits. With how temperature sensitive Ryzen CPUs seem to be with clock speed, the results makes sense. Its probably a 25-50 MHz speed difference with the drop in CPU temperature.

If i have time today, I'll run the same test with a 4.2 GHz all core overclock to compare it to my previous results.
 

mopardude87

Diamond Member
Oct 22, 2018
3,348
1,575
96
Got this on a 7700k stock with 16gb 2x8gb 2666mhz cl15 ripskill kit at 2400 cl15 atm. System freezes with 2666mhz with xmp for some reason. Maybe investigate and push to 2666mhz and test the difference. Doubt there will be much of a difference.

encoded 1806 frames in 588.27s (3.07 fps), 11820.04 kb/s, Avg QP:29.09
 

Bavor

Member
Nov 11, 2001
82
18
81
3950X @4.2 GHz
32GB DDR4 3600 MHz
Windows 10 Pro
The difference from previous results was that I left the front door of the Fractal Design Define R6 case open to allow more airflow. The CPU stayed cooler and didn't reduce speed as often to maintain temperature which gave me higher results.

HandBrake 1.1.0 (2018040700) - 64bit
OS: Microsoft Windows NT 10.0.18363.0 - 64bit
CPU: AMD Ryzen 9 3950X 16-Core Processor
Ram: 32685 MB,
GPU Information:
NVIDIA GeForce GTX 1080 Ti - 26.21.14.4187

encoded 1806 frames in 160.13s (11.28 fps), 11820.04 kb/s, Avg QP:29.09

41IYVS5.jpg
 
  • Like
Reactions: lightmanek

Fir

Senior member
Jan 15, 2010
484
194
116
3990x is not doing so well due to OS issues. I did load 10 Enterprise 1909 but it still shows TWO NUMA groups in task manager. All cores are around 4.05-4.18GHz in benchmarks that push 100% across 128 threads (CB20) but the scaling is wrong.

In any case, OT with this thread here is the handbrake results:

HandBrake 1.3.1 (2020010400)
OS: Microsoft Windows NT 10.0.18363.0
CPU: AMD Ryzen Threadripper 3990X 64-Core Processor
Ram: 130944 MB,
GPU Information:
NVIDIA GeForce GTX TITAN X - 26.21.14.4250
Screen: 4096x2160


encoded 1806 frames in 94.86s (19.04 fps), 11820.04 kb/s, Avg QP:29.09
 
  • Like
Reactions: lightmanek

Hulk

Diamond Member
Oct 9, 1999
4,191
1,975
136
3990x is not doing so well due to OS issues. I did load 10 Enterprise 1909 but it still shows TWO NUMA groups in task manager. All cores are around 4.05-4.18GHz in benchmarks that push 100% across 128 threads (CB20) but the scaling is wrong.

In any case, OT with this thread here is the handbrake results:

HandBrake 1.3.1 (2020010400)
OS: Microsoft Windows NT 10.0.18363.0
CPU: AMD Ryzen Threadripper 3990X 64-Core Processor
Ram: 130944 MB,
GPU Information:
NVIDIA GeForce GTX TITAN X - 26.21.14.4250
Screen: 4096x2160


encoded 1806 frames in 94.86s (19.04 fps), 11820.04 kb/s, Avg QP:29.09


Since it is possible to run multiple instances of Handbrake I'm curious as to how your system would do running say 3 or 4 instances of this benchmark as close to simultaneously as possible? By that I mean get 4 instances set and ready to run and then select start on all of them within a few seconds.

If the combined load is distributed more or less evenly across your cores and the benchmark results are good then it would certainly be a good workaround for transcoding multiple files quickly since one instance of handbrake won't utilize more than 8 or so cores very effectively.

If you have the time/inclination to do this I would be really curious to see the results.
 
  • Like
Reactions: lightmanek