Blender released an offical benchmark

[Gideon]

https://www.reddit.com/r/hardware/comments/968a3h/introducing_blender_benchmark_blenderorg/

Seems to be a good candidate for Cinebench replacement. The workload itself is loooooong though , taking about an hour for the full benchmark with 8 cores . The quick benchmark is more manageable. At least no short chiller shenanigans would be possible with this.

I guess we could also use this thread to compare the quick benchmark results (will run it on my Ryzen 1700X shortly)

Benchmark itself can be found here

 

[Gideon]

Quick run @ stock Ryzen 1700X took me:
21:34:39

 

[Justinus]

Quick run on my 5960x at 4.7: 17:01.97

Going to play with it a bit and get a full run, as well.

A lot of people play it fast and loose when it comes to overclock stability. I've had my share of experience with overclocks where you could run cinebench back to back indefinitely but as soon as something meaningfully stressful was going it would crash.

I think this has potential to be an excellent benchmark because it should weed out unstable or poorly tested overclocks fairly well.

My overclock is rock solid stable, will run 24 hours of non avx prime95 and two 8 hour realbench runs back to back, as well as 10 passes on memtest86.

My first attempt to run the quick benchmark resulted in the results page saying it crashed, but I am unsure if it was overclock related or beta software related. I've been running the benchmark for over an hour after that incident with zero issues, so I am inclined to blame the software given how thoroughly I test my overclock.

 

[richaron]

Seems Linux is giving an advantage with Blender.

1700@~3.7, ECC CL17@~3000, Ubuntu 18.04: 16:17.05 for the quick test.

Edit: 1:01:05.47 for the long test. Was running lightweight stuff and anime on a second screen at the same time so could shave off a little time if I was really trying... Just posting to give another reference point.

 

[24601]

Someone needs to do a run with a 1080 Ti, a Vega 64, and a Titan V (And perhaps the Quadro/Firepro counterparts of those to see if they have special driver segmentation that produces huge differences).

If a CPU is slower than 1080 Ti or Vega 64 for the particular Blender benchmark, the particular benchmark is realistically pointless for real world performance comparisons for those particular CPUs for obvious reasons.

We don't need more pointless benchmarks like Cinebench (I won't get into why it's pointless again).

 

[Justinus]

Seems Linux is giving an advantage with Blender.

1700@~3.7, ECC CL17@~3000, Ubuntu 18.04: 16:17.05 for the quick test.

Edit: 1:01:05.47 for the long test. Was running lightweight stuff and anime on a second screen at the same time so could shave off a little time if I was really trying... Just posting to give another reference point.

Yeah, it looks like the Linux version is better optimized.

My 17:01.97 score is W10. I booted up Ubuntu 18.04 and quick ran in 13:51.75.

Wow! Shaved off over three minutes.

 

[Justinus]

I did some testing between Windows, Ubuntu, and Manjaro and the results are very interesting.

All these results have been confirmed with 2 or more runs each.

Specs are 5960x at 4.7 core, 4.6 cache, DDR4 3200 14-15-15-30-1T.

Windows 10
Quick: 17:01.97
Full: 59:36.28

Ubuntu
Quick: 13:49.53
Full: 49:58.69

Manjaro
Quick: 13:22.44
Full: 48:01.36

 

[ub4ty]

I did some testing between Windows, Ubuntu, and Manjaro and the results are very interesting.

All these results have been confirmed with 2 or more runs each.

Specs are 5960x at 4.7 core, 4.6 cache, DDR4 3200 14-15-15-30-1T.

Windows 10
Quick: 17:01.97
Full: 59:36.28

Ubuntu
Quick: 13:49.53
Full: 49:58.69

Manjaro
Quick: 13:22.44
Full: 48:01.36

Linux for the win... Far less OS overhead as you'd expect vs Windows 10 thus why serious development is done atop it.

 

[moinmoin]

Anybody willing to test with Intel's Clear Linux? Phoronix keeps mentioning it having "out-of-the-box optimizations from PGO/LTO/FMV compiler optimizations, kernel patches, and a range of other optimizations compared to most other Linux distributions out there."

 

[Justinus]

Anybody willing to test with Intel's Clear Linux? Phoronix keeps mentioning it having "out-of-the-box optimizations from PGO/LTO/FMV compiler optimizations, kernel patches, and a range of other optimizations compared to most other Linux distributions out there."

I'll take a look when I have a chance, since I've already got test data from other OS's

 

[Justinus]

I'll take a look when I have a chance, since I've already got test data from other OS's

I'll admit I'm stumped. Blender simply won't launch on a new install of Clear Linux. The benchmark launches, and when I start it, it immediately stops with no error or message whatsoever. I went into the Blender folder to launch blender directly and it appears Blender simply won't start. I verified the OS is installed/running in x86_64 mode, but I'm not sure what else to try. Both Ubuntu and Manjaro worked booting from the install image and immediately upon install with no further configuration.

It's possible I'm simply missing some dependency, so I'll take another look tonight.

 

[Markfw]

I did it in 5:42:52 (5 minutes 42 seconds) AMD 2990WX@3550 with 3000 memory and win 10.

 

[Hitman928]

Mark, can you try the long test? Curious to see if it performs the same relative to other CPUs in long and short.

 

[Markfw]

Mark, can you try the long test? Curious to see if it performs the same relative to other CPUs in long and short.

Running. Will edit this thread when it finishes.

Edit: 21:56:14 @Fir did it in 18 minutes, but he is at 4100, and I am at 3550 mhz

 

[Hitman928]

Running. Will edit this thread when it finishes.

Edit: 21:56:14 @Fir did it in 18 minutes, but he is at 4100, and I am at 3550 mhz

Compared to richaron's 1700, it is ~ a 2.82x speed up for the short test and 2.77x speed up for the long. So it does slightly better in comparison on the short test, but it's really close.

 

[tamz_msc]

Compared to richaron's 1700, it is ~ a 2.82x speed up for the short test and 2.77x speed up for the long. So it does slightly better in comparison on the short test, but it's really close.

Which makes sense because the longer a benchmark runs the more likely it is that Precision Boost will decrease clocks due to the accumulation of heat.

 

[Markfw]

Which makes sense because the longer a benchmark runs the more likely it is that Precision Boost will decrease clocks due to the accumulation of heat.

Well, on my end I am OC'ed, no precision boost, just 3550 mhz (until my custom water is up and running)

 

[guskline]

On my 5960X at 4.4 the Quick Benchmark took 18 minutes 17.7 seconds while the full benchmark took 1 hour 4 minutes and 51.7 seconds. On my 2700x at stock below the Quick Benchmark took 18 minutes 44.62 seconds while the full benchmark took 1 hour, 6 minutes and 58.01 seconds.

BTW Markfw & Fir, those scores for your 2990wx rigs are incredible.

A whole lot of core power.

 

[richaron]

Compared to richaron's 1700, it is ~ a 2.82x speed up for the short test and 2.77x speed up for the long. So it does slightly better in comparison on the short test, but it's really close.

My rig was set at ~3.7GHz, so the only clock deviation would be throttling. Which I'm guessing is less likely on my build with mid-high end air cooling compared to a ~250W beast.

And I did have a legit (light) background load on my long (not short) run so I expect it has suffered slightly.

If anyone actually cares for more accurate (Ubuntu 18.04) numbers I'm happy to try match clocks and run again properly.

 

[WelshBloke]

Just found this thread!

18:31.06 quick benchmark on Windows 10. Stock 2700X (pb2 doing its thing and boosting to 4ghz) with memory at 3200mhz.

01:06:08.94 for the complete one.

 

[Timmah!]

Just found this thread!

18:31.06 quick benchmark on Windows 10. Stock 2700X (pb2 doing its thing and boosting to 4ghz) with memory at 3200mhz.

01:06:08.94 for the complete one.

Same here! Will give it a try with my 7940x!

EDIT> So gave it a try - the quick test - and the result was 11:02.39. That is at 4,1GHz all core turbo, 2,7GHz mesh and 3000 MHz RAM.

Needless to say, while it takes longer than Cinebench, its not as taxing. If the monitoring apps are to be believed, during short Cinebench run all the cores run at the full speed and the CPU package power hovers around 260W with my clocks/OC settings. With this benchmark, the clocks vary and fluctuate between 3,3 to 4,1 GHz, reaching 4,1GHz only during certain spikes, while the power hovers mostly around 220W, spiking to 240 max. And that goes only for clasroom test, the car one is even less demanding.

Seeing that, its very likely i could OC perhaps up to 4,4 - 4,5 GHz and still be relatively OK temp wise (by that i mean under 90C) - seeing how most of my cores were at the end of those 10 minutes in the 70-75C range, the hottest one being 77C. That could potentially shave another minute and get me under 10 minutes. Still far cry from those AMD 32 cores though, obviously

I might try the GPU test as well, just out of curiousity to see how the 1080 does in comparison to CPU and then, when my 2080Ti´s finally arrive, to see how much speed-up they provide. Since i ordered them primarily for this task, rendering.

 

[rvborgh]

tried running it on my 3.3 GHz 48 core Opteron 61xx system... and it crashes on startup with a memory access violation Error message "system_info["num_cpu_cores"] = int(cpu-topology" blah blah... seems to indicate that its having some troubles querying my system topology? Its currently 4 NUMA nodes with 12 cores each.


Blender itself runs fine... must be a bug in the benchmark

 

[kjboughton]

tried running it on my 3.3 GHz 48 core Opteron 61xx system... and it crashes on startup with a memory access violation Error message "system_info["num_cpu_cores"] = int(cpu-topology" blah blah... seems to indicate that its having some troubles querying my system topology? Its currently 4 NUMA nodes with 12 cores each.


Blender itself runs fine... must be a bug in the benchmark

Can confirm. Makers were informed the benchmark does not properly initialize on systems with more than a single CPU socket but have failed to address. So they have a massively parallel benchmark that bugs out using precisely those systems they wish to target and are silent as to fix. Weird

 

[DrMrLordX]

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.

 

[DrMrLordX]

https://opendata.blender.org/benchmark/9059950e-54c0-49d1-a5b0-fac6accd40cf

That's Radeon VII mostly at stock (RAM @ 1200 MHz and +50% power limit). Overclocking the GPU causes it to downclock in the second test to 1600 MHz or so. Lame.