8-year-old Xeon with SLI Titans XPs is faster than 1-year-old 64core SLI 3090s

[NeoXanthus]

Recently I decided to make a Franken-computer from old server parts and left-over video cards. But after speed testing it using Time Spy it is faster than my main rig. Looking for some help from the community to understand why my main rig is so comparatively slow in this test.

My main rig specs:
Dual water-cooled Asus RTX 3090 EKWB
Water-cooled AMD 3990x with PBO enabled TRX40 Aorus Xtreme and 256gb of 3200mhz memory

Second rig Franken-computer build:
Dual water-cooled Titan Xps
Water cooled dual E5-2687W-v2 with X9DAE supermicro and 64gb of 1867mhz memory

The 3DMARK Time Spy Score of the older computer is 22% faster than my new computer. What is very odd when swap the video cards between my main computer and Franken-computer, my scores are relatively similar. These TitanXps should not by faster than Asus 3090s so I am at loss understanding the issue here. I thought it might be NVlink bridge so I have tried both an EVGA bridge and the factory Nvidia bridge. Same results on both. I ran nvidia-smi to make sure the nvlink is working and it is (I removed the UUID for this post.) Any help solving this issue would be greatly appreciated.

C:\WINDOWS\system32>nvidia-smi nvlink --status -i 1
GPU 1: NVIDIA GeForce RTX 3090 (UUID: GPU-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx)
Link 0: 14.062 GB/s
Link 1: 14.062 GB/s
Link 2: 14.062 GB/s
Link 3: 14.062 GB/s

C:\WINDOWS\system32>nvidia-smi nvlink --status -i 0
GPU 0: NVIDIA GeForce RTX 3090 (UUID: GPU-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx)
Link 0: 14.062 GB/s
Link 1: 14.062 GB/s
Link 2: 14.062 GB/s
Link 3: 14.062 GB/s

Here are my scorers from the rig with 1-year-old parts and 8-year-old parts.

[Braegnok]

I would need to see the link to the result for the SLI 3090 Time Spy run to say exactly what the issue is. Need the Detailed Scores, Settings, and Run Details.

Based on information in the screen shot it seems you are not running 2X 3090 in SLI. The Graphics Score is not even in line with a single 3090 @ default settings.

Single 3090 Time Spy Graphics Score @ default settings should be in the 19K range, your screen shot is showing 12K for SLI.

Single 3090 @ default settings. http://www.3dmark.com/spy/20243133

[NeoXanthus]

Braegnok wrote:
I would need to see the link to the result for the SLI 3090 Time Spy run to say exactly what the issue is. Need the Detailed Scores, Settings, and Run Details.

Based on information in the screen shot it seems you are not running 2X 3090 in SLI. The Graphics Score is not even in line with a single 3090 @ default settings.

Single 3090 Time Spy Graphics Score @ default settings should be in the 19K range, your screen shot is showing 12K for SLI.

Single 3090 @ default settings. http://www.3dmark.com/spy/20243133

I apprentice your response. I ran the test again so I could get the URL did not even know it did that which is very cool BTW. Learn something every day.

https://www.3dmark.com/3dm/65880585?

As you can see it is 2x in SLI. I really want to understand why it runs so slow.

[Braegnok]

Thanks NeoXanthus,

Having the detailed scores, settings helps. I only notice few things that could be causing an SLI performance hit, your system memory speed, memory timings impact your performance, but I would not think the memory speed is the only issue.

The third graphics card ATI Radeon HD 5450 could be causing an issue, aside from using PCIe lanes that type card does not use CF straps it uses the PCIe lanes for crossfire, this could be causing an issue with your SLI configuration. What is the PCIe lane speed your SLI is running @ is it running 3.0 x16 speed.

Which slot is the HD 5450 card in, can you pull the card, if so please do,.. run Time Spy without the HD 5450 and see if that resolves the issue.

[NeoXanthus]

Braegnok wrote:
Thanks NeoXanthus,

Having the detailed scores, settings helps. I only notice few things that could be causing an SLI performance hit, your system memory speed is 3,172 MHz that does impact your performance, but I would not think the memory speed is the only issue.

The third graphics card ATI Radeon HD 5450 could be causing an issue, aside from using PCIe lanes that type card does not use CF straps it uses the PCIe lanes for crossfire, this could be causing an issue with your SLI configuration. What is the PCIe lane speed your SLI is running @ is it running 3.0 x16 speed.

Which slot is the HD 5450 card in, can you pull the card, if so please do,.. run Time Spy without the HD 5450 and see if that resolves the issue.

The memory is 3200mhz, which is the correct speed for the 3990x Threadripper. I have already tested this via the SPD overclock to 3600mhz. However, it does not make much difference in performance as the latency is also increased in the profile.

https://en.wikichip.org/wiki/amd/ryzen_threadripper/3990x#:~:text=The%203990X%20operates%20at%20a,%2....

I will pull the 5450 ATI card to see if it makes a difference and rerun. I use this oddball card to light up two old monitors for my coding work.

[NeoXanthus]

I removed the ATI card.

https://www.3dmark.com/3dm/65888265?

No change in results.

[Grimpak]

NeoXanthus wrote:
The memory is 3200mhz, which is the correct speed for the 3990x Threadripper. I have already tested this via the SPD overclock to 3600mhz. However, it does not make much difference in performance as the latency is also increased in the profile

you can supertune it (fiddle with the primary and secondary timings). However that is a very time-consuming effort, I must tell you. Considering zen2 specs, there is benefit to bring the memory up to speed of the infinity fabric (Fclk). To note that since ram runs at double data rate, that actually means half of the ddr speed too. Fclck on zen2 usually sits on the 1800mhz, so pulling speed up to 3600mhz and then finetune timings can bring extra performance to the table.

Now for the issue at hand, maybe there's a bios setting we're ignoring?

[Braegnok]

That is vary strange indeed, your system should be getting 30K + graphics score with SLI 3090 4.0 X 16 configuration.

The graphics score weight, equal to 0.85 and the CPU score weight, equal to 0.15 is how 3DMark Time Spy score is calculated, your raw performance Graphics Test is taking a huge hit. https://support.benchmarks.ul.com/en/support/solutions/articles/44002136143-how-is-the-3dmark-time-s...

That is crazy, I have run SLI configurations in the past, and have never seen this type scaling, raw performance results in Time Spy.

Hopefully someone else will chime-in with some helpful advice.

[xeromist]

What's the score like with SLI disabled?

[NeoXanthus]

xeromist wrote:
What's the score like with SLI disabled?

https://www.3dmark.com/3dm/65894401?

Slightly faster with SLI disabled, I do not understand.