cancel
Showing results for 
Search instead for 
Did you mean: 

5950x extreamly low L3 cache speed on 3801 BIOS

Rimc_LT
Level 7
I dont know then it started but today I saw that in AID64 my L3 speed is not normal does anyone have same.89987
802 Views
13 REPLIES 13

Aysberg
Level 10
I am using the 3801 bios on my Dark Hero and everything looks fine with my 5950x. At least I think.

1264.7 GB/s Read 909.47 GB/s Write 1064.2 GB/s Copy 11.9 ns Latency

Silent_Scone
Super Moderator
I try not to get too hung up on synthetic tests like AIDA, as there's too many variables when measuring latencies this low, such as background activity and system overheads.


All depends if it's having a tangible negative effect on system performance.
13900KS / 8000 CAS36 / ROG APEX Z790 / ROG TUF RTX 4090

RedSector73
Level 12
Rimc_LT wrote:
I dont know then it started but today I saw that in AID64 my L3 speed is not normal does anyone have same.89987


Silent Scone@ROG wrote:
I try not to get too hung up on synthetic tests like AIDA, as there's too many variables when measuring latencies this low, such as background activity and system overheads.


If for example it was background activity and system overheads, I would expect all the scores inclusive the RAM to show the effect, not just L3 in my experience. Without knowing the test conditions it is hard to just write this off as normal in my view, even if there is background activity and system overheads above normal. I wonder if OP has any CPU OC running apart from DOCP.

What I see from the result given is something isn't right and I have always found Aida, a reliable and capable tool. If you look more closely, all the L3 scores are crap with the last two scores in L3 posting in MB/s not GB/s, and the latency should be around 10 ns not 35 ns.

OP - Just for comparison my 3900x scores for L3 are below from Aida. Your L1 & L2 scores are better than mine, and your RAM is where it should be as well. Also compare to Aysberg scores which also show significant difference in his system to your L3.

3900x L3 Scores
R 1110.8 GB/s
W 1041.6 GB/s
C 1179.3 GB/s
L 9.9 ns.

If you own Aida (not sure if the free version does it) then stress test would be worth running. It moniotrs background activity and system overheads and provides warnings if these are above normal ranges. While on this topic again, I vagely remember the memory test you ran (OP) in Aida, briefly locks background activity and system overheads during the test but could not confirm this thought, still nagging me that it does do something like this. As an example, the corsair software that runs RGB, actually stops cycling briefly (multiple times) during the test cycle which in my mind kinda confirms it does or maybe my system is having background activity and system overheads without the crap score L3 to match and I only have unicorn computer systems on which to test. More than happy to test 5950x for years if anyone want to give me one, Asus ??

Another choice that is free, you can run userbench and post the link (which provides us some useable information and maybe some insight into what is going on) https://www.userbenchmark.com/

Hope you find this helpful.

RedSector73 wrote:
If for example it was background activity and system overheads, I would expect all the scores inclusive the RAM to show the effect, not just L3 in my experience.



That would depend on what bus the workload is hitting, so bit of a nonsensical assumption! 😄


Makes more sense to test real world performance, that way you don't send the user on a goose chase.


Here's a quick comparison.

One run is with normal system background activity, the other is light load placed on the memory and cache. Most notably transfer speeds are not impacted as bandwidth is not an issue, however cache hierarchy takes a hit. L3 cache is impacted the most as this is not only the entry point in the cache hierarchy, but also the slowest, making it susceptible to transient changes in background activity.


This isn't to say this is the is issue here, but it's one reason why performance monitoring is best left to real world tasks, and more to the point whether or not there is a tangible loss in performance



90005

90006
13900KS / 8000 CAS36 / ROG APEX Z790 / ROG TUF RTX 4090

Silent Scone@ROG wrote:
That would depend on what bus the workload is hitting, so bit of a nonsensical assumption! 😄


Maybe, what would be good real world test.
Cinebench R20 or what flavour do you think would help OP determine if there is anything going on, to minmise the goose chase ?

RedSector73 wrote:
Maybe, what would be good real world test.
Cinebench R20 or what flavour do you think would help OP determine if there is anything going on, to minmise the goose chase ?


I agree with that too. just test it in a benchmark.
with my Crosshair viii formula version 3801 runs without problems.
values are the same as with 3602

Rimc_LT
Level 7
It might be Windows 11 problem i think.

Hey now, at least your Aida benchmark does complete without crashing at an instant, requiring a full power cycle to the PSU and the system. For that is how Aida has been on my X570 Impact since BIOS 3204. The last time I was able to complete the benchmark was with 3003 and since then as soon as the memory read test begins, black screen - RAM fault LED on the MB, and no turn back other than with a power cycle. For reference, I haven't adjusted my RAM settings beyond DOCP, funny that.

yttersta wrote:
Hey now, at least your Aida benchmark does complete without crashing at an instant, requiring a full power cycle to the PSU and the system. For that is how Aida has been on my X570 Impact since BIOS 3204. The last time I was able to complete the benchmark was with 3003 and since then as soon as the memory read test begins, black screen - RAM fault LED on the MB, and no turn back other than with a power cycle. For reference, I haven't adjusted my RAM settings beyond DOCP, funny that.


Hey. Its happening to me also. If I enable pbo vmax in BIOS, if I turn it off or leave on auto it can run. Try to turn it off or auto for yourself.