[Artistus]

Hello!

It's strange after more than one year on the market and for the price, but on my both workstation systems and on other's too I can observe dramatic speed downgrade when rendering in Maxwell V4 (4.1.1.1 and all previous versions too) in comparison to V3 (3.2.1.5). Even the V4 version where it was claimed that speed issue was resolved and speed is the same as V3 - it's obviously not the same for V4 but 4.5 times longer rendering. Please, note that I don't talk about denoiser. Let's talk about workflow without denoiser, because it eats some small texture details when used.

My main workstation configuration is next:
2x CPU Xeon 2698 V4 (20 cores (40 threads) each)
MB ASUS Z10PE-D8 WS
8x16GB DDR4-2400 ECC Registered
etc.

I want to ask anyone who can help with this question, especially Maxwell dev team.
Can this speed issue be resolved soon, or is it impossible due to particular architectural changes made in V4 and maybe related to addition of GPU rendering?
Or generally can I get with some update for Maxwell V4 same speed +-5% as for Maxwell V3?
It depends on this question because I can't buy and use it for my needs with such speed flaws.

I attach Benchwell results on specified workstation for you to compare.

[choo-chee]

this is very interesting, and that's why I'm still stuck to v.3 . watching ...

[Mihai]

There must be something seriously wrong with either the testing scene or some other bug, maybe related to the nr of cores, or some of the channels. This isn't a general thing and I can't confirm such a huge difference in performance. Have you notified tech support of this post?

[Artistus]

This isn't a general thing and I can't confirm such a huge difference in performance. Have you notified tech support of this post?

I didn't notify tech support for that yet. But if you fire up Maxwell Render application on Windows 10 from V3 package and run Benchwell scene from it's menu, then fire up Maxwell Render application from any V4 package and run Benchwell from it's menu to compare speed, you will be blown away by the difference. Have you tried that? Please, check it if not. It's interesting if it has similar difference as for my rig. If so, then we must definitely notify tech support and it's weird why they haven't been notified before.

[Mihai]

Yes, I did. Got pretty much the same time and benchmark between V3 and V4. I think many more people would have noticed a 4.5 slower render time than before if it was a general thing.

[Artistus]

Yes, I did. Got pretty much the same time and benchmark between V3 and V4.

Thanks Mihai. Just checked on my laptop and results are the same on both versions. It's weird why are they so much different on 2CPU workstation. Maybe Windows 10 is the issue, or maybe some settings in motherboard's UEFI... I'll try to check with UEFI settings first, because resolving Windows issues usually requires much more time.

[Mihai]

I'd check with another scene besides the benchwell one so you can try setting different threads in render options to see if that changes anything. I have a single CPU machine.

[luis.hijarrubia]

In our tests we see no diference between v3 and latest v4 on cpu. And we have checked with 2 cpus machines too. Can it be related to some priority issue? Can you check if v4 is using all cores? Can you try benchwell with maxwell on high priority?

[Artistus]

Thanks for your advices, but I have some news on this issue on my system. First I tried Benchwell with different threads count as Mihai suggested. Thank you. I've got interesting results. As taken from my specs in 1st post I have 40 threads per CPU. So I tried setting 20, 40, 60, 70 threads to see the difference. I've checked how many cores are involved in work depending on settings - they all matched. 20 and 40 threads fit into 1st CPU, all above - involve cores from second CPU. So for Maxwell V3 in my case there is pretty linear increase in speed seen through all the range of settings tested. Maxwell V4 is increasing linearly up to 40 threads. So for each 20 or 40 threads set for benchwell there is pretty much the same result as Mihai mentioned. But all over 40 threads for V4 linearly decreases rendering speed and increases time like if I was not increasing threads value but decreasing it. So on 80 threads rendering time is a bit more then twice as bigger from that on 40 threads instead of being twice shorter. Can this issue be related with some setting in UEFI like CPU power management or NUMA setting (now in my rig it's set to Early Snoop)?

I also tried setting other priority as luis.hijarrubia adviced and if it had some influence, it was so minor that I wouldn't mention it.

[Artistus]

In our tests we see no diference between v3 and latest v4 on cpu. And we have checked with 2 cpus machines too. Can it be related to some priority issue? Can you check if v4 is using all cores? Can you try benchwell with maxwell on high priority?

Can you post your 2 CPUs machines specs and settings for CPU PM and NUMA in UEFI?

[Artistus]

Hi guys! I have news here.

I have tried changing QPI mode in UEFI from Early Snoop to other options which are next on my system:
1. Not Supported (Home Snoop Enabled & Cluster On Die Enabled)
2. Early Snoop (only Early Snoop Enabled)
3. Home Snoop on IVB...
4. Cluster On Die

All settings except Early Snoop offer less time similar for each (about 6 minutes opposite to 8.3 min on Early Snoop). But what is interesting here that on Not Supported mode and on Cluster On Die (both involve Cluster On Die supported) when rendering in Maxwell V4 Windows 10 task manager shows about 80% mean thread busy level on all threads, but if I look into Details tab - it shows only 55% mean CPU busy level on Maxwell application which floats between 50% min to 65% max. It's strange because Maxwell V3 shows in Task Manager 100% on CPU threads and 97-99% CPU usage on Details tab.

[luis.hijarrubia]

We have a big machine (I think 128 cores), and will test more cases. But the person on charge of that machine is off until monday, so we will tell you next week.

[Artistus]

We have a big machine (I think 128 cores), and will test more cases. But the person on charge of that machine is off until monday, so we will tell you next week.

I'll be looking forward to that. Thanks.

[Artistus]

the person on charge of that machine is off until monday, so we will tell you next week.

Hello. Was that person you mentioned able to make some testing this week on your big machine? Thanks.

[Forum Moderator]

I would try deactivating NUMA support in BIOS.

We have had some cases with computers with many cores that suffered from this issue too and NUMA was the key. NUMA is supposed to deal with how memory is handled to the different cores but it seems it doesn't work well with Maxwell. Once deactivated, the CPU usage was around 100% again (If I'm not wrong we had to deactivate it, but it could be the opposite )

Best regards,
Fernando