AMD 3900X Strange Core Ranking Cinebench Uses Slow Cores

[ltron]

I have a strange issue with my 3900X, on average the cores in the slow CCD 1 chiplet (4300-4400MHz) are ranked as faster in the CPU's firmware than the cores in the fast CCD 0 chiplet (4500-4650MHz). When applications are not using my fastest core 0 (correctly ranked) they will use a core pair that are ranked #2 and #3 fastest (incorrectly), these are cores 6 and 8 from the slow CCD 1 chiplet and they are only capable of 4.425GHz max which results in low Cinebench R15 and R20 single core scores of 205 and 510 respectively, multi core is fine at 3200 and 7250. Other people get 520-535 from what I've seen.

Here is a HWInfo screenshot so you can see what I mean, the #numbers on the right are how the cores have been ranked in the factory and written into the firmware and the #numbers on the left are what Windows thinks the ranking is. Can you check this on your CPU, particularly those with low Cinebench scores?

Here is Ryzen Master:

I have confirmed this ranking in Linux too so it's not a Windows issue and all screenshots I've seen from other people have shown CCD 0 correctly ranked above CCD 1. I have confirmed that all cores in CCD 0 do at least 4500 MHz by disabling CCD 1 (which resulted in 521 in Cinebench R20 and 210 in R15) whereas all cores in CCD 1 do at most 4425 MHz. I think someone may have made a mistake in the factory and ranked the cores wrong. Is there a plausible explanation for this ranking being correct and does anyone else have this problem? I'm asking for educational purposes, not that I can fix it as that's up to AMD (hopefully they can fix it if I can determine the problem is not unique to me) and I certainly wouldn't RMA a CPU over this.

This CPU core ranking is replicated across both an MSI X570 ACE and a Gigabyte X570 Master and multiple BIOS versions and AGESAs. CPU-z is fine at 541 single core (it does its own thing and uses core 0 which is my fastest core so that's why it's not affected by the ranking) and Geekbench 5 is fine probably for the same reason and I get 1315 single core. I ran Aida 64 Memory Latency test which loads core 0 and gives me 4.6GHz+ so I could show the CPU is capable of these speeds in the screenshot.

My specs:
AMD R9 3900X
Gigabyte X570 Master
32GB Crucial Ballistix CAS 16 DDR 4 3600MHz
Noctua NH-D15S
Palit GTX 1080 Gamerock Premium
Corsair RMi 1000W PSU
Corsair MP510 960GB
Windows 10 1909 64bit Pro

 

[ltron]

Are CPPC Preferred Cores and CPPC enabled under AMD CBS in the BIOS?

Yes, both are enabled. The problem is that the two preferred cores (6 and 8) are some of the slowest in the CPU. My screenshot shows that the supposedly worst core in the CPU (according to AMD) is reaching 4500MHz at idle (core 5) whereas cores 6 and 8 (two of the best cores according to AMD despite being in the slow CCD 1) reach only 4.425GHz max and these are the cores used by Windows in sustained singe core loads like Cinebench.

From Ryzen Master it would make much more sense to use cores 0 and 2 (1 and 3 in RM), both capable of 4.6GHz+ from my testing, so I'm struggling to see why the slow cores 6 and 8 (7 and 9 in RM) are used instead. I would think AMD would have used a sanity check in the factory to make sure the fast chiplet (CCD 0) is always ranked above the slow chiplet (CCD 1) given they went to the effort of binning them in the first place (every 3900X has one fast chiplet and one slow chiplet).

 

[ltron]

Yeah I agree it's very weird... My top 6 are all on CCD0, as they are in every other screenshot I've ever seen of 3900X chips..

Do you happen to know the date code etc from your chip? Mines 1925SUT.

No idea what to suggest really - perhaps ask the maker of HWiNFO as he's quite active on Reddit? He'll have seen lots of reports from these chips and might be able to shed some light.. see if your complete chip is a total outlier..

I don't know the batch code unfortunately, is there any way to get this info without taking off the cooler? Thanks for the advice I'll look into posting this on reddit/HWInfo's forum. I bought the CPU from a well known competitor (only place that had it in stock in early August that I could find) not the rainforest, so it's genuine and not an engineering sample. At least I hope so.

I'm not going to RMA the CPU over this, but it's just very annoying and all I can think of is that a mistake was made in the factory. What do you get in Cinebench R20 single core out of interest?

 

[ltron]

No way to tell without looking at the heatspreader - was just curious.. HWiNFO (in the full view, not just sensors) can detect ES CPU's.. I doubt it is one though.

Last CB20 result I got was 521 single, and 7206 multi. PBO disabled with 4* 8GB B-Die 3200C14 at 3800C16 but otherwise everything at stock.

Are you using AGESA 1.0.0.4B? It's not an ES according to HWInfo.

 

[ltron]

Maybe worth trying the 1usmus custom power plan as this is supposed to prioritise the best cores.

Thanks, I tried this when it was new and it made no difference. It may be worth trying it again now it has been improved upon, however the core ranking itself is the problem. It is using what it thinks are the best 2 cores in the same CCX, the problem is they are only capable of 4.4GHz and there are at least 6 other cores in the CPU that are capable of 4.5GHz+ but are incorrectly ranked as slower by AMD.

 

[ltron]

I'm on an X570 Taichi on the 2.73 beta bios - though those scores were likely on the earlier 2.50 with 1.0.0.4B. I don't think it was so long ago that it was going to be on 1.0.0.3ABBA.

Edit: Probably also with RAM at 3600C14 too.

Thanks, I'll just use the PC for now, hope that AMD can fix it and not worry about Cinebench. Having the cores ranked incorrectly only results in a 2-4% performance hit anyway.

 

[ltron]

I agree, the ranking is the core issue here.. out of interest, does the ranking order change at all if you enable PBO? I see you've got it disabled, but I've seen another post on a different forum where someone with a 3700X got quite a different order with and without PBO enabled..

I didn't try that because I saw a post on reddit where someone said it didn't change the order, only the values in the Windows event log were different but the order was the same (I spend far too much time on the internet researching this stuff). Thanks for the suggestion and I appreciate the replies given it's Christmas day.

 

[ltron]

No worries - its a curious case though. I like making sense of these weird edge cases.

The fact is though, you've had this persist across two motherboards which has to say that it's just the CPU, and nothing you change is going to make any real difference to it.. maybe you got a weird chip with only a half-decent 'good' CCD so it's all a bit confused.. it maybe prioritised cores on CCD1 because it thought for lightly threaded (but more than one thread) that CCD1/CCX0 was a better bet as a whole package (keeping things within a single CCX 16MB L3 cache) than CCD0/CCX0..

Still a good chip, definitely looks to be an outlier, and I seriously doubt there's anything that can be done to correct the oddness..

Thanks, I'm very grateful for your indulgence of me because I too like to understand edge cases and drilling into the detail. I suspect you are right that there is some reason for the ranking rather than it being a mistake on AMD's part, I can't say this for sure though. I think I might at some point install Cinebench under Wine on Linux just to see what happens. I confirmed the ranking of the cores using a Linux LiveUSB (i.e. without installing it) and some instructions on the web.