PC Processors

[s_vayner]

To AMD why your new processors never hit advertised boost speeds in single or multi core performance. I have Ryzen 3900x with Asus X570 Prime Pro board running latest bios and AMD chipset drivers. My single core boost is maxed at 4.2Ghz and multi core at 4Ghz. Doesn't matter if PBO enabled or not. There are 1000's of people complaining about this and you have not even acknowledged this as a problem. Your false advertisement about boost speeds is completely unacceptable. I was always Intel customer but this time I took a chance on your company and trust me this is first and last time i will buy anything from you. You hand selected CPUs for reviewers to get the hype going and stuck your paying customers with garbage. I will be filing a RMA claim and will be exchanging this processor until i get one that performs to advertised specifications.

[eddiez]

Temperature is a limiting factor when hitting the advertised speeds. If your airflow or CPU cooling sucks (which it never does of course, we are all PC experts :-/) and temps rise, the CPU will limit the boost margin. Simple as that. Oh, and other reasons could include thinking you set up the BIOS perfectly manually ...

[s_vayner]

Single core boost is never impacted by temp limit. i barely hit 65 on single core loads. but boost is low.

[eddiez]

I think you're wrong with that assumption. The temp of both cpu and soc, combined with voltage and RAM speed do influence the boost algorithm. But this does not necessarily have to be a cpu issue. Also the mobo and agesa can be the culprits. 

So I think the OP is a bit to pessimistic by blaming the cpu. Too many other factors are involved. 

[s_vayner]

My CPU is operating withing threshold to meet all boost requirements. I.e. voltage/temp/bus speed are below thresholds and below max advertised by AMD. AGESA microcode is also owned by AMD, motherboard manufacturers have nothing to do with its crappiness. Whether its CPU or AMDs fault i don't really care i just want my CPU to run at the advertised boost speed using hardware that was approved by AMD to run it.

[ajlueke]

Did running at a lower RAM speed help at all?

[eddiez]

I understand your frustration. The resolution is pretty simple: If you think your cpu is defective, RMA it. The "lots of people" do not constitute a generically failed product. Apparently lots of people have good working cpu's. Individual issues are always present, whatever product you buy.

So stop whining and RMA the thing. It's not getting better just by complaining. 

[ajlueke]

Just picked up an 3900X and installed over the weekend.  Updated UEFI on my Crosshair VII Hero to 2501 and ran the 3900X on stock settings.  No RAM overclock yet when I attempted my Cinebench R20 run.

The core does quite hit 4.6, but it is close enough that I would say I don't seem to have to abnormally low boost clocks you are seeing.  Only other thing I did in UEFI was disable CSM.

The thread was primarily loaded on the fastest core, but I did notice a few instances where the process was bounced to a different slower core.  The boost clock pictured here was held pretty consistently.

[ajlueke]

What's really interesting is what happens when I overclock my RAM.  Currently, I have it set to 3,400MHz and 1,700MHz on the fabric.  Now, when I run Cinebench R20 single core test, the fastest core in Ryzen Master is not used.  Instead, the thread bounces around from core to core but never to my "fastest core". 

This limits the boost clock I see to about 4.3 GHz as opposed to 4.6 GHz.  This is the problem with having software overhead to preferentially load the fastest core vs just have all fast cores, it's just another thing that can break.

[brucer]

So you are not hitting advertised frequencies either... This is still a common problem, AMD needs to get their crap right.

[ajlueke]

I did hit the advertised frequencies, until I overclocked my RAM.  Now, Windows doesn't seem to be able to load the fastest core anymore, which limits the boost clock I see.  Seems that if I reset CMOS, and load into Windows with default settings, the boost is fine.  If I start OCing my RAM and setting fabric/UClK speeds, then this happens.

Not sure how to get overclocked RAM/Fabric AND the fastest core loaded in single boost.

[eddiez]

That's a perfect example of how setup and environment do impact the performance of a cpu and whether or not it will reach its design limits. 

A Ferrari also doesn't reach its advertised top speed on every occasion or road... Some will never reach it

[ajlueke]

Now with the latest UEFI, 2606, single core boost is still problematic.  Now the processor will boost the correct core even with overclocked RAM, but not to 4.5 GHz + as it did with 2501.

 So, in summary on my Crosshair VII...

UEFI 2501

 C02 Fastest core boosted to 4.542 GHz.  Everything else at stock.

Overclock the RAM, and the C02 is no longer used, instead slower cores are used.

UEFI 2606

C02 is used even when RAM is overclocked.  However, speed is now limited to 4.274 GHz.  This is actually a performance downgrade from 2501.

All core boosting seems to work fine, all my cores boost to about 4.15 GHz as opposed to the 3.8 GHz base clock and stopped when it hit the EDC limit I had set.

[ajlueke]

This is the danger with having fast and slow cores.  To have the same performance as a processor with all fast cores, there is now software overhead that needs to load the correct cores at the correct time.  That is definitely not working correctly currently.

On the plus side, I was able to clock my 4X8 GB kit of RAM at 3600 CL16 relatively easily and keep the fabric clock at 1800 MHz as well.  All core boosting seems to work just fine.  At my EDC limit of 145A I was using about 1.4V sustained on the CPU for a clock speed of 4.13 GHz on all cores.  

[brucer]

my 3800x will do 4.4 all core stable  at 1.4v,  at default settings I will hit boost clocks, but its random across 3 different cores, and my fastest core according to Ryzen Master is ccx #6 yet it never hits the max boost clock, its either 1, 3 or 5 that hits it..

[shinkojiro]

Primary problem with Manual OC is too many people are scarred to go passed 1.35v as there is a maddening misconception that 1.325 is the safe voltage. If it were, AMD would never have default voltages so close to 1.5v.

My 3900x too does 4.4ghz at 1.4v all core stable, and achieves the same single core score I would if leave it to PB2/PBO.

What you should be able to do is use Ryzen master to set CCX0 preferrably to 4.5ghz, which may require a bit more voltage, and Check your single core score against what you would get with stock settings. This is what I am doing Currently:

Increases score from 205 to 216, showing that a manual clock actually reaches and maintains the speeds.

[ajlueke]

My issue is similar.  C02 is my fastest core, but other cores on CCD0 are similar.  CCD1 on the 3900X doesn't appear to be nearly as good on any core compared to CCD0.  So for the 3900X anyway, there appears to be one "good" CCD of 6 cores and one "bad" one.

All core loads appear to boost correctly and give performance in line with what is expected (about 4.15 GHz across all 12 cores).  Single core loads are the problem, and most of that seems to stem from the Windows scheduler.  This was supposed to be fixed in 1903, but it appears that in single core loads that the thread bounces around to suboptimal cores, even to those on my "bad" CCD.

Not sure if this can be remedied with a fresh install of Windows 10 1903.  Guess I can give it a shot.

[brucer]

if your fclk and memclk are different there is an issue with the memory controller and it degrades performance.. I seen this on a buildzoid video..

[ajlueke]

Thanks for the info, but I think it shows up in the Ryzen Master shots that I do have "coupled" mode active.

[shinkojiro]

Thats the almost the same speed I get with mine, 4544mhz. However, its really not just about seeing a reading close to the boost speed, its maintaining that speed through a single core load as well which isnt happening.

[ajlueke]

Maintaining the boost isn't really the issue for me.  The issue is more getting the thread on the fastest core and keeping it there.  My system sends that Cinebench thread all over the place, and it is rarely on the gold star core of CCD0 for long.

I didn't do a clean install of 1903 when I flashed and updated the CPU, so I may have some registry holdovers messing with appropriate scheduling.  I guess I could use something like Process Lasso to send all my background processes to CCD1 and see if that helps.  Windows really should just do that on it's own though.  Or, failing that, AMD could just give us cores that perform the same, and none of that would be necessary.  One or the other would be great, right now we have neither.

There does seem to be a conflict with the algorithm and generating the max single boost clock.  Form my observations the boost clock multiplier is determined at boot based primarily on the EDC value and the idle temp.  The cooler your CPU is at boot, the better the multiplier, and the higher the EDC the higher the multiplier.  Raising the EDC using PBO from stock to motherboard values adds voltage but also raises idle temps, so there is definitely a sweet spot there.  Adding a voltage offset cools things down and allows for a better all core boost, but then the system no longer has sufficient voltage for the low current single core boosts and those suffer.  It doesn't appear to be possible to set voltage offsets individually for high/low current situations.  Manual overclocks have the same problem, you can get a solid all core clock going at a decent voltage, but then again the low core performance is lost. 

If you turn of PBO entirely, then the single core boosts appear to be at their best, but multicore performance is lower than it would be with PBO on or an manual overclock.  Guess you have to chose one over the other.

[shinkojiro]

Looking at Ryzen master, it shows the same bouncing around between the 2 fastest cores of CCX0 with my manual OC like it does in Default, PBO, or Auto Overclocking. However my Manual OC leads to a higher score either way. That leads me to believe that None of the non-manual settings are maintaining a boost anywhere near 4.6 long enough for any performance gain that could be to be gotten. It may get close to there every now and then, but it is generally clocked lower most of the time.

That is indeed a problem with the algorithms, because that clock can be maintained for a higher performance score but they just wont do it. I still believe it is PB2 being way too strict with with whatever temperature it thinks the processor is at. Tho, the actual BIOS that is supposed to use that AGESA code across the boards is to blame as well.

[s_vayner]

This is my update after latest agesa aab release for my motherboard. After tweaking it and trying every possible combination of settings i was able to get my processor to boost to 4.5 on single and around 4.1 on multi. i think at this point this is as good as it will ever get. If anyone has Prime x570 Pro board these are the settings that will get you most performance, at least from my experience. 

1. Install latest BIOS, Chipset Driver, set power profile to AMD Balanced, Load defaults in bios.

2. make sure memory and infinity fabric is 1:1 ratio. Keep it on Auto if memory is at or below DDR3600

3. set CPU v-core to offset mode and set offset to around (negative) -0.05v. You can try values between -.025 and -.075 (at -0.075 you might get better boost but CPU may become unstable)

4. Enable all 3 PBO boosts in BIOS. One in AI Tweaker, One in AMD Overclock and one in AMD CBS (the last one is buried really deep in bios somewhere) i have trouble finding it every time. Keep everything on AUTO except setting PBO to Enabled.

With those settings i was able to get CB20 multicore to 7212 and single core to 519. I hope this could help someone with similar issues.

[shinkojiro]

Can you verify the same Single core score if you use a manual OC and use Ryzen master to set the speed of CCX0 to the rated boost speed?

[s_vayner]

i dont want to experiment with OC. These CPUs have no OC headroom and that core only barely sustains 4.5ghz for a few seconds. running it  locked at 4.6ghz will probably damage it.

[shinkojiro]

I guess you'll be stuck with what you got until AMD adjust the algorithm in AGESA updates.

[ajlueke]

Update for mine as well.

On UEFI 2606, it was definitely the RAM that was breaking the boost.  If I ran with just the optimized defaults loaded, my core boosts to around 4.5 on single core loads.  As soon as I upped the RAM clock speed, single core boost hard locked at 4.25GHz.

I did notice something strange though, even if I set RAM to 3600 MHz, Fclk would stay at 1,200Mhz instead of auto scaling to 1,800 MHz like it should.

I found two other places where I could set the RAM speed.  One, you set as the double data rate (3600Mhz), the other two you set the clock speed (1800).

I set the RAM in all places and the fabric now scaled correctly with my RAM.  On top of that, single core boost clock did not break!  It held at the same 4.5 GHz I had been seeing.  Sounds like this is pretty typical for the 1.0.0.3XX AGESA, while 1.0.0.2 boosted higher.  Either way, pretty happy with the result.

 

[shinkojiro]

That is what I get too, however the max processor speed I am seeing is 4550. I get higher in Multi-core tho.

To better show this is not holding that speed at all, here is my single core score when I am manually clocking CCX0 to 4.6 ghz:

And here is what I will leave as my safest Multi-Core score:

This maxes out the 200A EDC so i'm pretty sure that will go no further.

[ajlueke]

I'll probably leave mine boosting to 4.544 and wait for an AGESA update to get it to 4.6.  Manual overclocking can get higher multicore scores, especially with the CCD overclocking tool, but it sounds like voltages can get pretty nuts.   

[shinkojiro]

Its really no different that whats set at stock. In fact, I have it set lower. Stock does lower voltage for whenever it does all-core tasks, but for single-core its typically really that 1.487 you see. Up close to that is really what is required to run the cores at their boost speed, and ofc AMD like to overvolt a bit so you can do with a little less than that.

A perfect bios setup would be to be able to set those cores at 4.625 while setting the rest that are at 4.4ghz to 1.4v. I wish they would do something like that.

[ajlueke]

Further update.

I played around with CCX overclocking using Ryzen Master.  I can get CCX0 stable at 4.5GHz, CCX1 is 4.4GHz, with CCX2 and CCX3 are at 4.1 GHz.  All at 1.3V, which is within FIT spec for the silicon.  The second CCD can actually do higher than 4.1, but I cut it down to fit into my voltage profile.

Using this setup, I get identical single core performance to PBO, but significantly higher multithread performance.  I tried to get CCX stable at 4.6GHz, but some of the three cores can pull it off.  Trying to set an individual core in the CCX at a different frequency from the others leads to errors.  Overall, not bad.

[shinkojiro]

Yea 4.6 requires more voltage than people are generally willing to set :3 Imo whatever AMD will have set in stock/auto is safe enough for me. To me it means that boost clock is nearly the highest we'll be able to clock them, with some very slight headroom like 4650.

[ajlueke]

I think it demonstrates pretty clearly how poorly XFR2 is working with the 3000 series currently.  The weakness of manual overclocking is always that it necessitates setting a single CPU voltage.  Which means, it is impossible to have the high voltages at low current boosting algorithms can have. 

As such, the single core boost score should greatly exceed what I could get manually.  But it doesn't.  At 1.3V, I can run CCX0 at 4.5 GHz, meaning at 1.5V the fastest core alone should boost higher.  But I get the same score using the algorithm as my 1.3V manual boost.

Furthermore, I can get a better multicore score simply by preferentially boosting the best CCXs on the chip.  XFR2 doesn't appear to dynamically push the better CCXs higher, it just tries to push all of them the same.  CCD1 needs a ton more voltage to push over 4.2 compared to CCD0.  So my power budget should focus on boosting CCD0 and not wasting it on CCD1.  Multicore boost would be better if it boost all cores together to 4.1, and only boosted CCD0 after that.

The algorithm still behaves it a way that would make sense if all cores on the chip were equal, which they were more or less in the 2000 series.  That isn't the case now, and it feels like this should have been worked on before the 3000 series launched.

[hardcoregames_]

I simply use my CPU at stock with a good cooling solution. For me stability is more important than squeezing a new more Hz out of a CPU.

[rodrox]

I know is not the same as a high end Ryzen 3900X, but since day one I never had any issues with boosting speed on single core with my Ryzen 3600, I did had the issue with high volatge and temps on idle and load, that really worried me, and still does. But as someone pointed out voltage is just one part of the equation, if high voltage but low current then your CPU is ok for the long run.

I have a Gigabyte B450 Gaming X, which is really a low/middle range mobo.

In the image you can see a Cinebench R20 single core run, I tried 3 times, everytime reseting min/max/avg values of hwinfo when the rendering started. And every time I reached 4192.3Mhz, which is basically the 4.2GHz advertised.

Also not all youtubers get selected AMd cpus, some of them have to go buy it like any of us.

So maybe, as some already wrote is some BIOS/Motherboard problem that still need to be fixed.And it should be fixed soon!

Also not all youtubers get "hand selected" AMD cpus, some of them have to go buy it like any of us. Anyways, more about boosting issues on dif motherboards on this:

Zen2's 7nm Complications: Why not All Ryzen 3000 Cores Are Created Equal - Tom's Hardware 

AMD 3rd Gen Ryzen Boost Clock Investigation - YouTube

I need YOUR Help! RYZEN 3000 Boost Analysis - YouTube

Cheers

[sqwerty]

What happened to the thread title? Moderated? Too damaging? 

AMD, if you don't want calling out on shenanigans don't mislead your customers.

[eddiez]

How about: Not true? Or just ranting? 

Unsubstantiated claims, blaming one piece of a chain. Regardless of environments and user interventions thrashing AMD. 

So that would justify adjusting the title a bit. 

[ajlueke]

I would offer that the boost clocks quoted by AMD aren't a "lie", as in a complete fabrication.  There are definitely AGESA problems as possibly also scheduler problems from making this a reality though.

As in my earlier post, the best core in my CCD0 can in fact reach around 4.6 GHz, but it doesn't for various reasons.  In UEFI 2501, it is seldom the core that is actually being used.  Instead Windows puts single threaded workloads on cores in the much slower CCD1.

In UEFI 2606, every core seems to be hard capped at 4.25 GHz.  Doesn't matter which core is boosting on a single core load, they just won't exceed that.  So now I have to wonder if we'll ever see these operate the way the 2000 series did.  I guess the title could be Ryzen 9 3900X Single Core Boost Frequency:  Not There yet.

I think they'll eventually get there, but XFR2 is definitely not ready for prime time on these chips currently.  Hopefully this doesn't go the way of primitive shaders in Vega. 

[shinkojiro]

Because for they arent a lie. The title was nothing more than a baseless accusation. In fact the author does not even want to verify himself that his processor can actually do the 4.6ghz on any of the cores because hes scared to go manual. Lucky for him, I've already verified that it can do 4.6, and maintain it, as long as its not Precision boost 2 doing the work.

If you dont own the hardware and/or youre not willing to verify that the advertised boost is not achievable through any means, then you guys have no legs to stand on. AMD isnt going to get sued, and the problem will eventually get rectified once bios matures. It wont take all that long.

[tarvaln]

I'm also having these issues with the 3900x. I can't get any of my cores to hit 4.6Mhz at any time. These are the best numbers I could get out of 3 test per setting. I tested after midnight so the room is cool and let the PC cool for 10 minutes between each set of test. All on stock cooler. Cores were observed with Ryzen Master. All CPU setting were set with Ryzen Master.

My MB is Asus Prime x570 Pro. BIOS 1005 with AGESA 1.0.0.3 Patch ABB.

So, what can we do? I feel a little disappointed and cheated.

*I made this post a few days ago with a link to my MB info and BIOS on asus.com. It was under "Currently being moderated" until I removed the link.I understand not wanting malicious links. But a link to their own MB vendors... SMH.

[ajlueke]

Hmm, your single core scores under default are pretty typical for the current AGESA.  However, your Multicore does seem low.  I haven't tried 1.0.0.3 ABB yet, I give it a whirl and see how it behaves.