Blender released an offical benchmark

[Kenmitch]

I only did the short test.

3700x PBO on boosted to 4100 MHz during the test.


https://opendata.blender.org/benchmark/76a525ad-1c89-4e9c-862d-0dd2dca143e8

RX 5700 maintained boost in the 1770's


https://opendata.blender.org/benchmark/26ba01f4-cce9-40c1-9822-e786e9e19a92

Jeez....I had to have hwinfo64 running on the side as watching the progress bar was very boring.

 

[Soulkeeper]

14:54.20 cpu quick run
48min+ on my gpu, obviously it isn't running right with opencl in linux.

 

[DrMrLordX]

@Kenmitch

I'm wondering how your 3700x lost to a 2700x? Something doesn't seem right with the performance numbers there. My second (and best) run showed the two render scenes from the quick benchmark (bmw27 + classroom) taking approximately 9 minutes 55 seconds. My clockspeed was 4350 MHz fixed (had to back off from my usual CBR20 speeds, since Blender is more demanding). If I adjust for your boost clock of 4100 MHz and then adjust for you having 33% fewer cores:

((594.495 * 4350) / 4100) * 1.33333333 = 840.992 or ~14 minutes

I don't think you got the clockspeed you think you got.

edit: also one reason why @Soulkeeper might have done so well is that he's running Linux. If you're running Win10 then yeah that'll slow you down a bit. I was running Win10, though . . .

 

[Kenmitch]

I'm wondering how your 3700x lost to a 2700x?

I'm running Windows 10 Pro. It's probably semi whacked by now. I've disabled SMT, enabled SMT, flashed to newer uEFI, flashed back, had multiple reboots while playing with Ryzen Master and clocks, etc. I was planning on a reinstall when MSI dropped a new uEFI. The latest beta didn't like my b-die ram for some reason or another. Current uEFI I can run 3600 CL14, latest beta wouldn't boot 3600 with loose timings....Go figure?

I went with those two for now as the RX 5700 was exactly 10 minutes and a few milliseconds faster than the 3700x.

I'll have to run it again and double check the clocks.

Edit: I just ran it again to see. My clocks averaged 4094-4102 MHz during the run according to Hwinfo64. This run was 5 seconds faster than the above posted results. I noticed that I forgot to tweak my cpu fan speed in the uEFI after I flashed back. I'll have to go back in there and mess around with them to see if it'll maintain more boost as the clocks were higher in the beginning and slowly settled as the cpu temperature rose. Temp peaked at 76°C.

 

[lopri]

I've got 15:38.91 with my 3700X. bclk 102. Kenmitch's time looks normal to me?

 

[Markfw]

Full test on a 2990wx, linux, 16:53.92

Short test 4:18.31

1080ti 6:21.62

 

[john3850]

Ran it at 47 for a minute before going to 5000MHz which gave me a time of 14:41.37.

 

[Markfw]

Ran it at 47 for a minute before going to 5000MHz which gave me a time of 14:41.37.

Is that the short test time ?

 

[john3850]

Is that the short test time ?

I ran 50 from the start on the short test and got a 14:39.

 

[Markfw]

I ran 50 from the start on the short test and got a 14:39.

Yes, I was hoping that was compared to my 32 core 3.8 ghz 2990wx at 4:18. So that was your 9900s@5 ghz ?

 

[john3850]

Cores rule in blender and speed does not.

 

[Kenmitch]

Cores rule in blender and speed does not.

I got 15:55:60 on my 3700x which has the same amount of cores as you. You beat mine by 1 minute 16 seconds with a 900 MHz advantage. I was at 4.1 GHz.

Could you do another run while monitoring power consumption with hwinfo64? If so once the benchmark loads all cores to 100% reset hwinfo64 as I'd be interested in seeing the averages at the completion of the benchmark.

 

[Markfw]

My 3900x just did the quick test in 9:22.84 Linux

 

[Markfw]

So why aren't more people using this benchmark, at least for single-socket systems? I love CBR20 and all, but Blender seems more relevant.

Here's my current best results under win10 for my 3900x:

https://opendata.blender.org/benchmark/ba4d59e8-2458-40af-b80b-a39f3d5ea8ed

That's just the quick run though. Might do the full run later.

edit: did the full run, under Win10. Results:

https://opendata.blender.org/benchmark/cbf5951e-83f8-43b8-a0f6-ea34c8920954

Took me 35 minutes and 35 seconds to complete all the render tasks.

WOW, linux is faster....my 3900x@stock
https://opendata.blender.org/benchmark/32378a9e-9905-48cd-b042-fc9d6bf61a36

 

[Soulkeeper]

Yeah, linux is faster in most things that coders bother to target

 

[DrMrLordX]

I'm running Windows 10 Pro. It's probably semi whacked by now. I've disabled SMT, enabled SMT, flashed to newer uEFI, flashed back, had multiple reboots while playing with Ryzen Master and clocks, etc. I was planning on a reinstall when MSI dropped a new uEFI. The latest beta didn't like my b-die ram for some reason or another. Current uEFI I can run 3600 CL14, latest beta wouldn't boot 3600 with loose timings....Go figure?

I went with those two for now as the RX 5700 was exactly 10 minutes and a few milliseconds faster than the 3700x.

I'll have to run it again and double check the clocks.

Edit: I just ran it again to see. My clocks averaged 4094-4102 MHz during the run according to Hwinfo64. This run was 5 seconds faster than the above posted results. I noticed that I forgot to tweak my cpu fan speed in the uEFI after I flashed back. I'll have to go back in there and mess around with them to see if it'll maintain more boost as the clocks were higher in the beginning and slowly settled as the cpu temperature rose. Temp peaked at 76°C.

Hmm.

I've got 15:38.91 with my 3700X. bclk 102. Kenmitch's time looks normal to me?

Maybe. Care to try running it at a static clockspeed? Like 4000 MHz? Obviously you'll have to dial it in carefully because of the bclk increase.

WOW, linux is faster....my 3900x@stock
https://opendata.blender.org/benchmark/32378a9e-9905-48cd-b042-fc9d6bf61a36

Linux makes this and GB4 run a lot faster. It's a bit scary.

 

[john3850]

I got 15:55:60 on my 3700x which has the same amount of cores as you. You beat mine by 1 minute 16 seconds with a 900 MHz advantage. I was at 4.1 GHz.

Could you do another run while monitoring power consumption with hwinfo64? If so once the benchmark loads all cores to 100% reset hwinfo64 as I'd be interested in seeing the averages at the completion of the benchmark.

I reread your post a little on the late side.
I did a 3 blender short 14.30 to 14.35.10 time
5000MHz vcore 1.356v-1.368v
Cpu package high 191.6w 186.3w
IA Core Power high 189.735w 188.633w
Riva Tuner read 190.6w
This was the time the cores were fully loaded till the end.

 

[Thala]

Linux makes this and GB4 run a lot faster. It's a bit scary.

it is certainly not Linux, which makes the benchmark run faster.

 

[DrMrLordX]

it is certainly not Linux, which makes the benchmark run faster.

I think it's really the compiler. GB5 uses the same compiler for their Linux and Windows binaries, and the results are performance parity between the operating systems.

 

[Ajay]

Gack! Ran the quick test on my GTX1070 (PCIe 2.0), pretty sad:



I don't dare run it on my old i7 970 (x58) after seeing this result.

Edit, don't know if this ran using CUDA or not.

 

[DrMrLordX]

Some of the GPU results are not necessarily that great. Radeon VII is a beast in this benchmark, though it takes a long time to do warm-up passes. Still waiting to get a proper driver that doesn't downclock RAM or GPU clocks for my card when overclocking in this bench, though.

 

[Ajay]

Wow, this seems like an awful time for a quick test on my i7 970 (3.84 GHz):



Man, I need a new computer. I've been having performance issues. It has been getting on my nerves anyway for a string of problems, won't upgrade to 1903 - even after using every troubleshooting trick I could find on the web. But, the joys of home ownership, roof leaked pretty badly end of this summer, so now I'll be spending money on that instead

 

[Kenmitch]

Some of the GPU results are not necessarily that great. Radeon VII is a beast in this benchmark, though it takes a long time to do warm-up passes. Still waiting to get a proper driver that doesn't downclock RAM or GPU clocks for my card when overclocking in this bench, though.

I did the quick run on my 5700 last night and it maintained 1819-1824 MHz this go around. Little tweaking of fan curve and used morepowertool to slightly tweak max clocks. Treading with caution with the tool.

I got 5:34:55 this time....Typo should be 5:34.55

 

[Ajay]

Some of the GPU results are not necessarily that great. Radeon VII is a beast in this benchmark, though it takes a long time to do warm-up passes. Still waiting to get a proper driver that doesn't downclock RAM or GPU clocks for my card when overclocking in this bench, though.

I wonder if the test defaults to OpenCL. AMD has robust OpenCL support, Nvidia doesn't. I'll have to look up and see if there is a command line switch for CUDA.

 

[DrMrLordX]

Wow, this seems like an awful time for a quick test on my i7 970 (3.84 GHz):

Man, I need a new computer. I've been having performance issues. It has been getting on my nerves anyway for a string of problems, won't upgrade to 1903 - even after using every troubleshooting trick I could find on the web. But, the joys of home ownership, roof leaked pretty badly end of this summer, so now I'll be spending money on that instead

Yeah fix the roof first.

Nehalem is getting a little long in the tooth. For reference, here's my 3900x doing the full test in ~35 minutes:

Blender Open Data

Free and Open Data for everyone. Download the official Blender Benchmark.

opendata.blender.org

That being said, how much serious Blender work do you plan on doing anytime soon?

I did the quick run on my 5700 last night and it maintained 1819-1824 MHz this go around. Little tweaking of fan curve and used morepowertool to slightly tweak max clocks. Treading with caution with the tool.

I got 5:34:55 this time.

Hmmm that seems a little slow. Is that with the full test? I think software optimizations for compute on RDNA just aren't there yet.

Actually now that I think about it, that isn't too bad. My jacked up Radeon VII does the same test in about 266s, or 4 minutes 26 seconds.

I wonder if the test defaults to OpenCL. AMD has robust OpenCL support, Nvidia doesn't. I'll have to look up and see if there is a command line switch for CUDA.

Blender has full CUDA support. If you look at the front of their opendata page, you'll see CUDA and OpenCL results being listed separately. Also, if you report your data online, you can get the post-run output and see which API it used during the run. For example, here's Radeon VII @ something like 2100 MHz GPU/1000 MHz RAM:

Blender Open Data

Free and Open Data for everyone. Download the official Blender Benchmark.

opendata.blender.org

At the bottom, it says:

device_infodevice_typeOPENCL

etc.