EXPO dropping ryzen 7900 performance 12%

[mainframe89]

As per title, has anyone else had this, perhaps with MSI AM5 motherboards?

 

cinebench r23 multi - stock - expo off/ram at 4800 - 25500

cinebench  r23 multi - expo on 6000 - 22300

 

No other settings in BIOS changed, stock BIOS V1.0, PBO disabled

 

Specs

ryzen 7900 (non-x)

g.skill ddr5 6000 36-36-36

msi b650i edge

7900xt gpu

750w PSU

 

The only thing I can see is when running bench with EXPO enabled is the core voltage of the CPU tanks, the speed drops from 4.3ghz to 4.0ghz and the wattage used by the cores drops 20W from 65W to 45W, even though the PPT is 88W under both EXPO and non EXPO conditions.

 

Not sure whos at fault, the CPU, the motherboard or the RAM?

 

Any help much appreciated.

 

MF

[RONOTHAN##]

I'd say that's likely some sort of bug in the particular BIOS revision you're on. If you go to a newer or older revision, does it still happen? 

 

If the CPU is downclocking with XMP enabled it has to be a motherboard issue, not a RAM issue. The only question now is if it's a bad BIOS revision (these do pop up from time to time) or is it an issue with your specific board that requires an RMA?

[porina]

There is no fault. Cinebench only really cares about CPU clocks, and ram speeds has an insignificant effect on it. When you turn on EXPO more power goes to the IMC, leaving less for the cores, and thus lower core clocks. This has been seen in DDR4 era too on both AMD and Intel CPUs when running on a power limit.

[RONOTHAN##]

1 minute ago, porina said:

When you turn on EXPO more power goes to the IMC, leaving less for the cores, and thus lower core clocks.

IIRC AMD's PPT system doesn't factor in SOC power draws, only core power draws, but even if it did there's no way enabling EXPO would cause the IMC to draw 20W more. The SOC in total is lucky if it draws in the neighborhood of 15-20W total including the power it takes for handling all the PCIe lanes and other IO functionality, it's not going to more than double by just enabling EXPO, at most only going up another ~2W. 

 

There's something else going on with this.

[porina]

34 minutes ago, RONOTHAN## said:

IIRC AMD's PPT system doesn't factor in SOC power draws

It was my understanding that PPT is everything in the socket, which would include IMC. If it also alters the IF speeds that and anything synced to it will burn more power also.

 

In the past I also had trouble getting all the individual reported powers to add up to PPT, so there may be some reporting difference somewhere too.

[PDifolco]

6 minutes ago, porina said:

It was my understanding that PPT is everything in the socket, which would include IMC. If it also alters the IF speeds that and anything synced to it will burn more power also.

 

In the past I also had trouble getting all the individual reported powers to add up to PPT, so there may be some reporting difference somewhere too.

According to HWInfo64 the total power draw includes SOC power, and it' s that measure that may be limited by PPT afaik, at least that's what I've observed when tweaking PLs on my 5900X

OP results are pretty weird, EXPO should only affect RAM latency and timings if it works like XMP, unless it now tweaks as well IF clock or something else it makes no sense to lose CPU voltage and clocks while using it

 

 

[RONOTHAN##]

19 minutes ago, porina said:

If it also alters the IF speeds that and anything synced to it will burn more power also.

Ryzen 7000 always runs at 2000MHz FCLK with auto settings no matter what the RAM speed is, so that shouldn't be it. There is something else going on in this config. I'm still gonna go and say that the current BIOS revision OP is on has some sort of bug that happens when XMP is enabled, but there's definitely something up. You should be able to check though to see if the SOC is pulling a ridiculous amount of power though, since it should be on two different power planes there should be two different power draw figures reported by the board itself (not the CPU). 

[mainframe89]

Thanks for the feedback! Great knowledgeable community here.

 

I would've thought enabling ram XMP/EXPO wouldn't have tanked CPU speeds this much, no one would use it right? And my reference for these cinebench scores are all online youtube videos - there are multiple of them showing a 25000+ CPU score WITH overclocked DDR5-6000. I am suspecting the motherboard/BIOS but I don't have any other AM5 parts to reference with, and they aren't cheap. There is a BETA BIOS revision V1.13 on MSI's website which I tried, and there was no difference, so I rolled back to the V1.0.

 

Here are screenshots under cinebench loads - first without EXPO enabled and ram at 4800

 

 

And here with EXPO enabled - literally just the single click in the bios.

 

 

 

 

Here was just one of the setups that I found on a YT vid

 

 

 

Don't get me wrong I do understand cinebench is not a good representation of how a PC will perform day to day. But it is a good metric that has loads of online data to compare apples to apples. I just was just benching my new rig after putting it together to make sure everything was sweet, and to do that I used online R23 references. I found my CPU under performing compared with everyone else.

 

It does appear to be differences in SOC power... hmm..

 

[mainframe89]

Okay the plot thickens. Seeing those strange SOC numbers made me want to bench the memory controller. And lo and behold errors within seconds - do I have a lemon? But what, the CPU, the RAM, the motherboard? Keep in mind this is all default BIOS no EXPO no OC's

 

 

 

[mdk777]

How many sticks of memory ? brand (any brand can get the binning wrong, but some more than others)

Check the voltages,

Sometimes a small boost makes a difference.

 

My ddr3 3200 is 1.35

Just boosting to 1.37v makes them run at 3600 (optimal for AM4)

 

My guess is the memory.

 

EXPO boosts to the levels that the memory stick tells it to.

The memory doesn't always get it exactly correct.

 

 

https://www.newegg.com/g-skill-32gb/p/N82E16820374419?quicklink=true

 

Well the comments are mixed here.

One person says that the MB was the problem.

 

A second seems to think the memory was at fault.

 

Looks like you have your work cut out figuring out which one.

 

[mainframe89]

I've got the whole lot out of the case and im running tests now. Will report back when I have found something.

[mainframe89]

So the problem was the motherboard, I'm pretty sure. I think MSI will need to sort it in BIOS updates. I re-seated and pasted the CPU and swapped the RAM slots. This did not fix the cinebench issue but did fix the memory errors.

 

I was able to get things somewhat close to normal by manually setting the VDDR_SOC to 1.25V, or more correctly, an offset or -0.1V as with EXPO on my MSI board the voltage was 1.35 and was causing the memory controller to pull too much power, at least I think.

 

The PPT was 88W and the CPU and SOC power are included within this PPT limit. So the SOC was drawing more than it needed to. And it was stable with SOC voltage at 1.25V.

 

As for the Memory errors I got in prime95 with expo disabled, we Cell, who knows. I couldn't recreate that, however I had re-seated the CPU and RAM so this was likely the issue.

[yetsosimple]

Oh man, so glad that I found this post. I have exactly (!) the same problem with the same CPU and same Motherboard. I can tell you it’s not the RAM – I had the G.Skill kit first and replaced it with Kingston without any effect (apart from solving my POST issues).

 

I did come to the same conclusion as you, in that it is a BIOS bug related to power draw, resulting in lower frequency under load. Changing VDDCR SOC voltage had no effect for me. My workaround for now is to change PPT to 100 which results in same frequency boost under load and same temperatures as without RAM overclocking.

 

My reasoning why there is little discussion about this is that the non-X CPUs are fairly new, and this bug only manifests if PPT is limited. But it certainly would be nice if MSI would fix this.