Ryzen 3800x Much Slower with XMP (very odd)

[JCBiggs]

I ran my new 3800x build through UserBench today. I LIke the benchmark for its comparison purposes across a generic workload.   it scored verywell, The CPU was in the 93rd percentile.   but when I got to looking at the memory, the latency was a little high on the ram.

 

So I opened HWinfo and sure enough, even though the ram was set to 3600 mhz manually, It didnt pick up the advertised timings (obviously since i didnt enable xmp)  So for reference  the timings that are active are 3600mhz, 26-25-25-58  (horrible)

 

So I enabled XMP and my motherboard reports that the ONLY changes that are being made is the memory timings to 18-22-22-42.    I boot the system with no problem and run the benchmark again and it scored CONSIDERABLY less. the cpu performance dropped all they way to the 7th percentile. (FROM (93!)   the gpu also dropped, as well as the ram score.  

 

I cut the XMP profile off and re-ran the test, and got the performance back.   Both modes  had the infinity fabric coupled at 1:1 (1800mhz)

 

 

I have no explanation for this unless its some kind of very misconfigured sub timing.      Anyone? 

 

 

(EDIT _ appologies, the Ram score actually did go up a couple of points. from 97% to 101% with  XMP. its just the CPU that dropped. The gpu did as well but negligible)

 

 

BUILD:

3800x

Vengeance 3600mhz cas 18 Ram (8gbx4)

ws pro Ace x570

 

[JCBiggs]

NON XMP

UserBenchmarks: Game 80%, Desk 167%, Work 144%
CPU: AMD Ryzen 7 3800X - 98.5%
GPU: Nvidia Quadro P4000 - 67.5%
SSD: Samsung 970 Pro NVMe PCIe M.2 512GB - 387.5%
HDD: Seagate Barracuda 7200.14 2TB - 115.2%
RAM: Corsair Vengeance LPX DDR4 3600 C18 4x8GB - 97%
MBD: Asus Pro WS X570-ACE
 

 

 

XMP
UserBenchmarks: Game 77%, Desk 158%, Work 133%
CPU: AMD Ryzen 7 3800X - 88.1%
GPU: Nvidia Quadro P4000 - 67.2%
SSD: Samsung 970 Pro NVMe PCIe M.2 512GB - 390.7%
HDD: Seagate Barracuda 7200.14 2TB - 115.9%
RAM: Corsair Vengeance LPX DDR4 3600 C18 4x8GB - 100.7%
MBD: Asus Pro WS X570-ACE
 

[Jurrunio]

5 minutes ago, JCBiggs said:

infinity fabric coupled at 1:1 (1800mhz)

The one you couple is the UCLK, or memory controller frequency. Infinity fabric frequency (FCLK) is another number somewhere else.

[Harry Voyager]

What motherboard, which BIOS, and what RAM are you running? 

 

When I built my machine, it needed the Bios flashed to even boot with the full memory kit in it, even with it on the QVL for that MB. 

[JCBiggs]

Just now, Jurrunio said:

The one you couple is the UCLK, or memory controller frequency. Infinity fabric frequency (FCLK) is another number somewhere else.

its 1:1. both test were coupled. checked in Ryzen master to make 100% sure when i saw the odd result. 

[JCBiggs]

3 minutes ago, Harry Voyager said:

What motherboard, which BIOS, and what RAM are you running? 

 

When I built my machine, it needed the Bios flashed to even boot with the full memory kit in it, even with it on the QVL for that MB. 

added to post,

 

its a asus ws x570 ace  Im running latest bios .  Ram is vengeance lpx 3600 cas 18  8gb x 4

[Flying Sausages]

5 minutes ago, JCBiggs said:

added to post,

 

its a asus ws x570 ace  Im running latest bios .  Ram is vengeance lpx 3600 cas 18  8gb x 4

Can you test something like cinebench r20 or r15, unigine valley and heaven benchmark, and 3dmark?

[Jurrunio]

15 minutes ago, JCBiggs said:

3600 cas 18  8gb x 4

maybe it's unstable then. Keep XMP profile enabled, but select 3200MHz and see if score improves

[Mark Kaine]

Yeah. But userbenchmark is very unreliable.

 

Use some real benchmarks.  

[JCBiggs]

13 hours ago, Mark Kaine said:

Yeah. But userbenchmark is very unreliable.

 

Use some real benchmarks.  

it doesn't matter how "reliable" it is.  it consistently shows a lower score with the ram timings tightened.  I did run cinebench and it was 2165 average across 3 test with xmp memory on.  with xmp off, it scored 2204.

 

 

[JCBiggs]

13 hours ago, Jurrunio said:

maybe it's unstable then. Keep XMP profile enabled, but select 3200MHz and see if score improves

ill do some more testing after work. my bet is that at 3200 the score will drop further since the fclock will drop to match.  but just a hunch. 

 

 

[Jurrunio]

3 hours ago, JCBiggs said:

ill do some more testing after work. my bet is that at 3200 the score will drop further since the fclock will drop to match.  but just a hunch. 

 

 

just FYI, if you can push FCLK 266MHz or higher than half memory data rate withy stability, push FCLK higher and let it desync. Faster that way

[Mathieu9836]

Memory timings set to auto will give you all sort of weird and awful timings, set them manually using Typhoon Burner and DRAM Calcultator.

[JCBiggs]

I beleive the above poster was correct.  All I did was manually set 18-22-22-42  instead of using XMP and my score results went back to where they were.  Though oddly, I didnt improve any. latency did a few points. 

 

Good enough for me for now.  Ill do a consierable amount of tuning later. Im still not hitting 4.5ghz as often as I like. Solidworks is mostly single threaded so I would like to get that performance up a little. 

[Mathieu9836]

33 minutes ago, JCBiggs said:

 All I did was manually set 18-22-22-42 

 

Those timings are really loose at 3200mhz

 

[JCBiggs]

On 11/14/2019 at 3:16 PM, Mathieu9836 said:

Those timings are really loose at 3200mhz

 

3600

[JCBiggs]

i took it off XMP and just set the timings manually and it worked fine. The XMP settings must of changed something else.