The AMD Ryzen 9 3950X Review - LLC Testing

Intro

This right here is a subset of the results I have from the 3950X review. Today we will focus on the motherboard setting for load line calibration. During my testing I found that lowering the load line strength resulted in higher scores in most tests. For the MSI X570 Meg Unify there are LLC settings from 1 to 8 with 8 being the lowest load line with the most droop. So, I will be showing my benchmark results for every load line setting in The Chip Collective’s standard benchmark suite.

I spent a lot of my own money on this and a couple other upcoming things for the website. Please throw some money my way either directly VIA PayPal or with my Ko-Fi maybe don’t do my Patreon until I fully set it up.

Test System Specs

• AMD Ryzen 9 3950X
• MSI X570 Meg Unify
• Corsair H115i PRO
• 2x G.Skill TridentZ RGB DDR4-3466 F4-3466C16D-16GTZR (Samsung B-Die) (Single Rank) (Provided by Intel for the 9900K review)
• 2x G.Skill TridentZ RGB DDR4-3466 F4-3466C16D-16GTZR (Hynix C-Die) (Single Rank)
• EVGA GeForce RTX 2080 Ti XC BLACK EDITION GAMING 11G-P4-2282-KR
• SanDisk Ultra II 480GB SSD
• PNY CS1311 960GB SSD
• Samsung 970 Evo Plus 250GB NVMe (Provided by Intel for the 9900K review)
• Cooler Master MasterBox NR600
• SeaSonic FOCUS Plus Gold 850W 80+ Gold (Provided by Intel for the 9900K review)

Tests Covered

• AIDA64 (3 iterations)
• Cinebench R20 (6 multi threaded iterations & 1 single threaded iteration)
• Cinebench R15 (6 multi threaded iterations & 1 single threaded iteration)
• GeekBench 4.3.4 (1 iteration)
• Asus Realbench (6 iterations)
• PassMark PerformanceTest (3 iterations)

System Config

The computer is running with the BIOS in default auto mode for everything except the LLC setting. The memory is running with XMP enabled which is at 3466 and no special tuned timings. Tuned timing stuff is coming in another part of this review to be published soon. The fans on the AIO are fixed to 100% and the pump is set to performance mode and there are no case panels installed on the case.

Test Results

Cinebench R15

Cinebench R20

Asus Realbench Image Editing

Asus Realbench Encoding

Asus Realbench Heavy Multitasking

AIDA64 Memory

AIDA64 CPU

Geekbench 4.3.4 Single Core CPU

Geekbench 4.3.4 Multi Core CPU

Passmark CPU

Passmark Memory

Aggregate Single Threaded Score

Aggregate Multi Threaded Score

Aggregate Memory Score

Thoughts and Conclusions

It would seem as though either heat or voltage prevents the CPU from boosting as high as it could. The lower load line settings help this, and you get a faster CPU. Simple right? Well here is some conflicting information for ya. If you set an offset voltage even one decrement from stock voltage all your scores will be lower across the board. Sorry I don’t have those test results right now, but you just gotta believe me and that is why I didn’t even continue with testing voltage offsets. It also has nothing to do with heat because the massive liquid cooler I’m using. This is especially true under these conditions with boosting set to auto in the BIOS. It simply has to do with those unmeasurable with software measurements transient voltage changes skewing the boosting algorithm.

For the most part you have LLC7 in the lead for the multi threaded tests and LLC6 in the lead for the single threaded tests. No significant changes on any of the memory tests. More testing is needed from other motherboards before a blanket conclusion can be made obviously. In another part of this review I will also have benchmarks that show LLC’s effect with different boosting preset modes that MSI has in the BIOS.

Thank you for reading this and don’t forget what I said earlier in the intro.

I spent a lot of my own money on this and a couple other upcoming things for the website. Please throw some money my way either directly VIA PayPal or with my Ko-Fi maybe don’t do my Patreon until I fully set it up.

I’ll be working tirelessly to bring you the rest of my review. I just wanted to push this little bit out to you. I actually think it will all be in sections instead of one long review article page because I have so much data.