Binning my 2 3900x chips

kamikazi

[H]ard|Gawd
Joined
Jan 19, 2006
Messages
1,592
I've actually purchased a second 3900x and I hope to bin it tomorrow. I've had the first chip running since August, but haven't had a ton of time to mess with it. The time I have had has been a bit frustrating, so I decided to pick up a second chip. Here is what I've done so far on the first chip as a quick and dirty binning test.

This is taken from my thread in the hot deals forum when 3900x's popped back up on Amazon for $499.99 https://hardforum.com/threads/3900x-for-499-99-at-amazon.1989392/#post-1044411252

Chip #1
So, I did some quick and dirty OC'ing to find the limits of my chip by CCX. First, I set infinity fabric fclk to 1800 and RAM to 3600 (I had been running at 3733 16-15-15-32-52 1.43v). I backed off on timings to 16-16-16-16-36-52. In Ryzen Master, I set cpu voltage to 1.3125. I set all cores to 4250, then I went back and went through each CCX one at a time seeing how high I could go before throwing errors in a 5 minute CB20 loop. On the "good" chiplet, CCD0, CCX1 maxed at 4400 and CCX2 maxed at 4400. On the "craplet", CCD1, CCX0 maxed at 4300 and CCX1 maxed at 4400. I noticed that I got slightly higher temps running looking for maxes on CCD1. Once I found these, I set each CCX to max and tried a to loop CB20 for 5 minutes, which is 8 loops. It scored 7689, but failed on pass 4. I raised CPU voltage to 1.31875 and it made the 5 minutes with a score of 7685, but on the very last pass, there are artifacts. Probably needs one more voltage bump for true stability Temps maxed around 75C. All power settings are set per recommended settings in the Ryzen DRAM calculator. LLC is set to 3, power levels set to 130, VRMs set to extreme.

If I set everything to default CPU wise, running CB20 for 5 minutes sees max clocks on any one core of 4017.8 which is the 40.3 multi with Bus clock at 99.8, but just for a millisecond. All cores seem to sit at 3992.8 which is the 40 multi. So what level silicon am I dealing with here? I got my other 3900x in and I'll be running the same exact tests Friday I hope.

So, any suggestions on changes or things to try on both chips? Thanks for any input.
 
Try out boost tester
See which one has the fastest core...

Tangoseal was going to run on his TR3 but I think he forgot.
You should be able to get higher on the goodlet vs the shitlet lol.
Good luck!

P.s. please tell us the production date of the CPUs 1 and 2.. be interesting to see if the sillicon improvement theory is correct. Some people with late 3600xs getting better OCs than early production.
Same thing happens usually but this fast is unusual. Further credence to the rapid and unprecedented ramp of 7nm.
 
Try out boost tester
See which one has the fastest core...

Tangoseal was going to run on his TR3 but I think he forgot.
You should be able to get higher on the goodlet vs the shitlet lol.
Good luck!

P.s. please tell us the production date of the CPUs 1 and 2.. be interesting to see if the sillicon improvement theory is correct. Some people with late 3600xs getting better OCs than early production.
Same thing happens usually but this fast is unusual. Further credence to the rapid and unprecedented ramp of 7nm.

I'll do it. Thanks for reminding me about that utility.

Back in the Bloomfield days with Intel, their early silicon was better. I had a very, very late core i7-930 d0 that was one of the most power hungry, lowest clocking examples I ever saw. I think it came down to Intel binning as they got further along and introduced faster models. I'm hoping for better silicon this time.
 
I couldn't figure out how to run that utility mentioned above to find max core speed. All I can find is a Visual Studio project file and I have no idea what to do with it.

Didn't get as much done today as I would have liked. A new BIOS dropped today, so I had to rebuild BIOS again to match old settings. The new BIOS indicates that it changed overclocking rules, whatever that means. I did a full passmark memtest86 6 hour run before retesting, just to verify there are no a RAM issues. I retested the same way I did yesterday and it appears nothing has changed. The CCX0 still won't go over 4400 stable in CB20 for 300 seconds at 1.3125v. I have also noticed that just running CB20 at stock, the chip seams to boost a little lower. No core ever breaks into the 4000s now. 40 multi is as high as it goes, no 40.3 multi. CCX0 at 4400 and the rest of the chip at 4250 nets 7551 in CB20. Some odd behavior to note. SOC voltage is different. I ran at the same voltage I did previously, 1.11250. This keeps SOC voltage just above 1.1 in HWinfo and at 1.1 even in Ryzen Master. When I did this, CB20 would fail after about 2 passes. I lowered it on bin which then shows exactly 1.1 in HWInfo and Ryzen Master and it failed after 4 passes. I then changed it to auto in BIOS and it shows 1.094 in HWINFO, but it shows 1.363 in Ryzen Master, but it completed all 8 passes. I have a hard time believing that the BIOS would auto it as high as Ryzen Master shows. I'm hoping that Ryzen Master is wrong and HWINFO is correct. There are two different sensors in HWINFO showing essentially the same voltage. I think I'll go ahead and move on to the new chip to see what it can do.
 
Is that VS file from GitHub, or ?
Link me to it and I'll take a look.
It's be ironic to fork it and build a toolchain out of it.
 
Got the utility working:

Annotation 2019-12-07 155915.jpg


Edited to show the new chip.

New3900x.jpg
 
Last edited:
It looks like a script we'd use to load a VM instance or scaling group to test auto scaling.
 
I edited a couple of posts above to show the difference between the chips with the boost utility. Looks like every CCX but one is superior on the new chip and that one is a tie with the worst chip from the other one...at least running that utility. This chip seems to run hotter though. I've mounted twice and I guess I'll try a third time just to be sure.

Old chip was week 26 2019, new chip is week 36 2019.
 
  • Like
Reactions: N4CR
like this
Still running very hot compared to the first chip. It peaks at 87 during a 5 minute Cinebench run, but typically stays around 80-81, which is about 5 degrees warmer at the same settings. CCD0 runs about 10 degrees hotter at max than CCD1. I've now applied kryonaut 3 times. How does my TIM application look to you?

2019-12-09 14.08.49.jpg
 
Looks fine, it'd probably have shaded the edges bc Tim will spread when you clamp down your cooler.

What kind of cooler are you using?
 
Looks fine, it'd probably have shaded the edges bc Tim will spread when you clamp down your cooler.

What kind of cooler are you using?

H100i SE Platinum in push pull. Corsair LL120 push and Noctua AF-12x25 in pull.
 
I found the chiplet positioning in 3900x/3950x interesting from a cold plate standpoint.
What happens if you popped off the am4 bracket and reattached it so it was 90deg from where you have it right now?

Might want to be doing this type of testing with a graphite pad for consistency.

I wonder if you'll find an orientation over the hottest chiplet that matters in respect to an Asetek aio.

This would make for a good read: https://www.igorslab.de/en/ryzen-30...asymmetrical-design-with-interesting-results/
 
I found the chiplet positioning in 3900x/3950x interesting from a cold plate standpoint.
What happens if you popped off the am4 bracket and reattached it so it was 90deg from where you have it right now?

Might want to be doing this type of testing with a graphite pad for consistency.

I wonder if you'll find an orientation over the hottest chiplet that matters in respect to an Asetek aio.

This would make for a good read: https://www.igorslab.de/en/ryzen-30...asymmetrical-design-with-interesting-results/

The rotation is a good idea, but I don't know if the cooler will allow it because of how the bracket and the coolant lines interact. The coolant lines would be on the same side as one of the thumbscrews. It's very odd because the other chip just ran much cooler even on the same settings. I actually ordered one of the IC graphite pads earlier today. Great minds think alike. This version of the Corsair H100i is actually a Coolit unit instead of Asetek. Based on the picture in the article and the graphic on Corsair's page, it would appear this is how mine is currently lined up as my chip is rotated 90 degrees from vertical in the motherboard. I would think this would be the better of the two possible alignments.

AiO-Position-1-768x764.png
collapsible-1.png


On another note, I got a CB20 pass, just one run, at 4.325 GHz all core at 1.3v which is better, but hotter than the other chip. 4.350 GHz reboots the machine immediately with a high temp warning, which makes no sense as it's under 85 C.
 
Here is like 10 mins and I am using the system to do stuff like internet browsing etc.... nothing heavy duty.

Using AMD Ryzen Max Performance Profile but I normally use Balanced.


I am having several cores hit over the spec of 4.5ghz by hitting 4.55 but I have seen one the other day hit 4.65 ghz for a brief moment. Also my loop is running a little warm. I have 7x120 worth of rads but right now 4 fans are removed and I am cooling all of this on just 3x120 worth of fans in push only. I ordered 7 Delta fans (pics below) to place in push on my loop. These fans should knock it out of the park even on low speed. Delta Brushless dual ball bearing fans. Will last me a decade or more with up to 14mmHg of pressure. Better than any other fan you can find that is a 1 amp with PWM.

Also with my loop running @ 30% or so cooling capacity because I have 4 fans removed I finally hit over 70C. With all my fans online running even @ low RPM this chip never hit 70c no matter how hard I hit it. So this should be a testament that even in a custom loop at 3x120 with a 2080ti is NOT an effective number of radiator slots.

upload_2019-12-10_1-39-32.png


upload_2019-12-10_1-35-34.png


upload_2019-12-10_1-37-51.png
 
Last edited:
  • Like
Reactions: N4CR
like this
^ likes to wear ear pro by his rig, but I can appreciate Delta or sunon industrial fans on an open loop.
Lots of guys went that way back when peptides were a thing

I can't wait for any of the TR3 guys I used to work with that were on x299 + multi gpu + external square rads packed with 9x120mms to show me their next builds. Those video editors and vfx guys would benefit from higher density local nvme storage options. They'll generate a ton of heat but require quieter open loop solutions bc they're in the studio.

OP guess flipping an Asetek bracket isn't an option.
It would be a worthwhile exercise if you could get your hands on a 280mm or 360mm aio to see if there was an optimal can patch and water flow path.
If I had the same issues as you, I'd go grab one from Best Buy or similar local chain retailer with generous return policy and generate the data.
 
  • Like
Reactions: N4CR
like this
Old chip was week 26 2019, new chip is week 36 2019.
Some more evidence towards later chips doing better and more creedence to why AMD may have waited for ramping before releasing these...

Thanks for running the utility and keeping us updated!!

Seeing that microchannel overlay I'm wondering if the outer edges of the chiplets are impacted. Probably not as I believe the limit is the sillicon transfer side of things.

tangoseal approach is great it shows that a high thermal gradient benefits these chips to minimise silicon heat transfer issues.
On that line I want to look at if cooling these at daily sub ambient temps will further help.
When I do my next build ( looking q1 at this rate) I want to test chilled water on a spare chiller, then if promising mill a custom copper block with a 500w pelt or stacked pelts and water chiller and see what happens.. Ultimately would love to develop an embedded pelt that can do pure water or pelt+ water without much efficiency loss but the design will be very challenging... Whole different way to do a waterblock that I have not seen before.
Either way it'll be [H]ard earned fun to try.
 
Some more evidence towards later chips doing better and more creedence to why AMD may have waited for ramping before releasing these...

Thanks for running the utility and keeping us updated!!

Seeing that microchannel overlay I'm wondering if the outer edges of the chiplets are impacted. Probably not as I believe the limit is the sillicon transfer side of things.

tangoseal approach is great it shows that a high thermal gradient benefits these chips to minimise silicon heat transfer issues.
On that line I want to look at if cooling these at daily sub ambient temps will further help.
When I do my next build ( looking q1 at this rate) I want to test chilled water on a spare chiller, then if promising mill a custom copper block with a 500w pelt or stacked pelts and water chiller and see what happens.. Ultimately would love to develop an embedded pelt that can do pure water or pelt+ water without much efficiency loss but the design will be very challenging... Whole different way to do a waterblock that I have not seen before.
Either way it'll be [H]ard earned fun to try.

LTT did heavy testing of some pretty high wattage peltiers and found that were abhorrent. I can't remember the video precisely though.
 
^ likes to wear ear pro by his rig, but I can appreciate Delta or sunon industrial fans on an open loop.
Lots of guys went that way back when peptides were a thing

I can't wait for any of the TR3 guys I used to work with that were on x299 + multi gpu + external square rads packed with 9x120mms to show me their next builds. Those video editors and vfx guys would benefit from higher density local nvme storage options. They'll generate a ton of heat but require quieter open loop solutions bc they're in the studio.

OP guess flipping an Asetek bracket isn't an option.
It would be a worthwhile exercise if you could get your hands on a 280mm or 360mm aio to see if there was an optimal can patch and water flow path.
If I had the same issues as you, I'd go grab one from Best Buy or similar local chain retailer with generous return policy and generate the data.

Thoughts on this "AIO" http://www.swiftech.com/drive-x3-aio.aspx

360_3B.jpg

I was looking at this back when I built my machine, but went all Corsair instead. A quick search doesn't yield any results of how well the apogee SKF LT waterblock does with Ryzen. The site lists AM4 and TR4 compatibility. I also question how much air is getting through the part of the radiator above the reservoir. I guess no push pull unless you drilled your own holes on the one side.
 
Last edited:
My 3900X seemed to respond nicely to some extra voltage (and cooling!)

Personally, the ROI got bad after 1.38V, squeezed another ~50Mhz stable after getting to 1.416V (@80C!!). Dipped my toe in 1.42+ and got nothing much out of it.




LLC was 'Turbo' in Aorus Master which is 3rd out of 5 levels. VRM was one step down from max setting.

For 24/7 dropped back to 1.375, gets me stable with CCX1/2/3/4 at 4.525, 4.450, 4.425 and 4.400
Testing showed just stable at 1.368, keeping a tiny more in it to be safe.

All tests with memory at 3800 @ 14-16-16-32, fclock 1900


Looking at the review here, https://www.techpowerup.com/review/swiftech-swiftech-h360x3-drive-x3-aio/

Seems pretty good, higher end 240s seem to still beat it on occasion. Pump or fan must play a bigger factor than 50% more radiator.
 
My 3900X seemed to respond nicely to some extra voltage (and cooling!)

Personally, the ROI got bad after 1.38V, squeezed another ~50Mhz stable after getting to 1.416V (@80C!!). Dipped my toe in 1.42+ and got nothing much out of it.




LLC was 'Turbo' in Aorus Master which is 3rd out of 5 levels. VRM was one step down from max setting.

For 24/7 dropped back to 1.375, gets me stable with CCX1/2/3/4 at 4.525, 4.450, 4.425 and 4.400
Testing showed just stable at 1.368, keeping a tiny more in it to be safe.

All tests with memory at 3800 @ 14-16-16-32, fclock 1900



Looking at the review here, https://www.techpowerup.com/review/swiftech-swiftech-h360x3-drive-x3-aio/

Seems pretty good, higher end 240s seem to still beat it on occasion. Pump or fan must play a bigger factor than 50% more radiator.


Your chip is a monster, fclk 1900 to boot. I haven't tried pushing that on this chip for anything other than default settings and it wouldn't post. Anyway, right now, temps are so high on this thing I can't do anything. I'm now spiking in the mid 80s on a Cinebench run at stock. Something just isn't right here. I checked out my fan curves and what not and nothing has changed. On second review of that AIO I mentioned above, I don't really like it. Comes prefilled with that opaque fluid, can't run push-pull, doesn't do any better than what I have most likely. Maybe a little quieter.
 
I yanked the CPU and checked it for flatness with a precision straight edge. It's a little concave. I checked the old CPU and it's about the same. Then, I put the old CPU back in and this time, just put a blob of kryonaut in the center of the chip. Below are some screenshots of HWINFO right after both chips completed 5 minutes of CB20:

Chip 1
Post CB 20 CPU 1.jpg


Chip 2
Post CB 20 CPU 2.jpg


The old chip with lower boost potential beats the newer chip in CB20 score because it runs cooler and therefore can boost higher. The new chip causes the system to run so much louder because of the temps it hits ramping up the fans. This is a huge difference. This new chip is maxing around 10 C higher than the old one. I know there are variations in chips, but this seems extreme to me. My graphite thermal pad arrives tomorrow. I may run the new chip with just a blob in the middle and see if it helps just to make sure I'm not overdoing the TIM, but I don't think it will matter. There shouldn't be this much variance.
 
I yanked the CPU and checked it for flatness with a precision straight edge. It's a little concave. I checked the old CPU and it's about the same. Then, I put the old CPU back in and this time, just put a blob of kryonaut in the center of the chip. Below are some screenshots of HWINFO right after both chips completed 5 minutes of CB20:

Chip 1
View attachment 206217

Chip 2
View attachment 206218

The old chip with lower boost potential beats the newer chip in CB20 score because it runs cooler and therefore can boost higher. The new chip causes the system to run so much louder because of the temps it hits ramping up the fans. This is a huge difference. This new chip is maxing around 10 C higher than the old one. I know there are variations in chips, but this seems extreme to me. My graphite thermal pad arrives tomorrow. I may run the new chip with just a blob in the middle and see if it helps just to make sure I'm not overdoing the TIM, but I don't think it will matter. There shouldn't be this much variance.

Considering it's clocking higher it shouldn't be worse sillicon (maybe higher leakage though...), but perhaps might be variation in indium solder application?
Very rare for people to have two chips..
 
Considering it's clocking higher it shouldn't be worse sillicon (maybe higher leakage though...), but perhaps might be variation in indium solder application?
Very rare for people to have two chips..

I don't think it's a leaky chip, it can do a higher all core overclock at a lower voltage than the first chip, the problem is there is no thermal headroom. If it's always bumping 85, it's just not going to be able to boost like it should and I'll just have a louder system that performs about the same in everyday tasks. I think you're right about the solder. This is a Malaysian chip. The first is Chinese. Maybe the assembly is slightly different. Maybe Der8auer would like me to send this to him for delidding :D
 
I notice your CCD 2 on Chip 2 appears to be inline with CCD 1 and 2 from Chip 1 for temperatures. Maybe CCD 1 on Chip 2 is bringing that chip down?

Those voltages seem in line with those speeds - in general. Definitely seems like you'd need further thermal headroom to really know what you've got.

If you didn't try it this way the first time, maybe do the per CCX OC again, but only OCing 1 at a time and leave the remainder on stock (or declocked) values. Voltage will still be up there, so it might not make a difference.

I'm using conductonaut for TIM, maybe something like that could help?

Also, maybe manually clock and volt both CPUs the exact same - same test, same duration. If the solder was questionable I think you'd see the majority of that temp difference carry over into this test.

How do you find the build week?
 
Last edited:
I don't think it's a leaky chip, it can do a higher all core overclock at a lower voltage than the first chip, the problem is there is no thermal headroom. If it's always bumping 85, it's just not going to be able to boost like it should and I'll just have a louder system that performs about the same in everyday tasks. I think you're right about the solder. This is a Malaysian chip. The first is Chinese. Maybe the assembly is slightly different. Maybe Der8auer would like me to send this to him for delidding :D
Fair call re:leakage. Interesting that they switched from CN to Malaysia!!!? First I've heard of this, I really wish a hardware review site would focus on little details like this. Maybe the trade imbalance microcode patches in USA are the cause ;D TBH I'm surprised they did anything in CN for assembly, pretty sure the Epyc2/TR2 was Taiwan as I remember them making a bit of a point to mention Germany + Taiwan + USA on the heatspreader in past.
Hmmmm... can you borrow another water block and or rad? Kyle had some up for sale a few weeks back, or maybe you could work something out with Dan_D ; he's been aching to get his hands on some of the 3000 series stuff....
 
Fair call re:leakage. Interesting that they switched from CN to Malaysia!!!? First I've heard of this, I really wish a hardware review site would focus on little details like this. Maybe the trade imbalance microcode patches in USA are the cause ;D TBH I'm surprised they did anything in CN for assembly, pretty sure the Epyc2/TR2 was Taiwan as I remember them making a bit of a point to mention Germany + Taiwan + USA on the heatspreader in past.
Hmmmm... can you borrow another water block and or rad? Kyle had some up for sale a few weeks back, or maybe you could work something out with Dan_D ; he's been aching to get his hands on some of the 3000 series stuff....

I've actually had nearly the entire 3000 series in my hands at some point. I do have a 3950X inbound as well. The only CPU's I haven't had the pleasure of working with are the 3rd generation Threadrippers.

As for reviewers not getting into some of those details, it's a matter of time. Before a launch, we barely have two weeks to gather our data, learn to overclock, deal with problems with the platforms, BIOS revisions etc. and all of that. I had just 10 days prior to the 10980XE launch. I spent a ton of that time trying to get a hold of comparison CPU's. The 3900X launch gave me about as long but that was more complicated. Some of the finer details don't get seen by us until we've had the CPU's for awhile, or unless we throw them in our personal machines. A factor in something like examining differences in thermals between Malaysian, China, etc. comes down to the fact that we get our allocation of CPU's prior to launch. We rarely get any after that. We'd have to try and source the various CPU's from retail and that costs money. Few sites make enough to just do that outside of isolated cases.

To investigate this, I'd really need to get ahold of a CPU that's the same and assembled in each location. Then I'd have to test, delid and retest. That's a crap ton more work than it might sound like to you guys.
 
I've actually had nearly the entire 3000 series in my hands at some point. I do have a 3950X inbound as well. The only CPU's I haven't had the pleasure of working with are the 3rd generation Threadrippers.

As for reviewers not getting into some of those details, it's a matter of time. Before a launch, we barely have two weeks to gather our data, learn to overclock, deal with problems with the platforms, BIOS revisions etc. and all of that. I had just 10 days prior to the 10980XE launch. I spent a ton of that time trying to get a hold of comparison CPU's. The 3900X launch gave me about as long but that was more complicated. Some of the finer details don't get seen by us until we've had the CPU's for awhile, or unless we throw them in our personal machines. A factor in something like examining differences in thermals between Malaysian, China, etc. comes down to the fact that we get our allocation of CPU's prior to launch. We rarely get any after that. We'd have to try and source the various CPU's from retail and that costs money. Few sites make enough to just do that outside of isolated cases.

To investigate this, I'd really need to get ahold of a CPU that's the same and assembled in each location. Then I'd have to test, delid and retest. That's a crap ton more work than it might sound like to you guys.

Hey Dan thanks for the detailed response! That was certainly not a dig - I get how much time pressure is on you review gurus to get performance data alone, let alone the fiddly stuff like this. Typically details like that take months to figure out as you have alluded to, I have not seen it discussed anywhere else (3950 specific) other than this thread so far as it's still early days.
I guess in this case you are stuck with user data, perhaps another way to do this is the usual user-submitted database approach to see if there is any correlation, get readers/viewers to help where possible.
Maybe kamikazi could be persuaded in the name of science if a loaner chip was available ;) The re-lidding may just fix the thermal difference on the second chip too.

I'm really looking forward to seeing your take on the TR3 stuff, I guess AMD is playing hard to get with the current stock numbers. If I was a millionaire I'd just send you one to review for all the hard work you have done over the years. Cheers
 
Hey Dan thanks for the detailed response! That was certainly not a dig - I get how much time pressure is on you review gurus to get performance data alone, let alone the fiddly stuff like this. Typically details like that take months to figure out as you have alluded to, I have not seen it discussed anywhere else (3950 specific) other than this thread so far as it's still early days.
I guess in this case you are stuck with user data, perhaps another way to do this is the usual user-submitted database approach to see if there is any correlation, get readers/viewers to help where possible.
Maybe kamikazi could be persuaded in the name of science if a loaner chip was available ;) The re-lidding may just fix the thermal difference on the second chip too.

I'm really looking forward to seeing your take on the TR3 stuff, I guess AMD is playing hard to get with the current stock numbers. If I was a millionaire I'd just send you one to review for all the hard work you have done over the years. Cheers

I appreciate that. I don't know that I'll get to review TR3. It's very expensive and we are just starting out. I can get the TR40X motherboards of course, but AMD wouldn't sample us on the CPU's. I'm debating buying one for myself, but that's a hard sell. I don't really need it, even though I want one. It's of more limited value given that the launch window passed. The longer you wait, the less relevant that stuff becomes. So, it might be one of those things we simply missed the train on. I'm just glad that we got the 10980XE sample from Intel so we had something on that day. The upside was I got to be a bit more thorough with it than I would have been if I had all four CPU's to contend with at once.
 
Going through a similar process with my 3970x, but using stock voltage.
R20 loop for 300 seconds, usually 82-85C at the end with stock voltage 1.28749 in RM, RM says power usage is usually 265W, using a noctua SP3/TR cooler with push/pull.

Bios optimized defaults usually scores 17100-17500 in R20.

all core 4149 is the max it passes the 5min cinebench test.
going through each CCD at 4250 with the rest set at 4100, only CCD2 passed.
With CCD0: 4150, CCD1: 4150, CCD2: 4250, CCD3: 4150 it passes the CB 5 min test and scores 17,916 at stock voltage.

Next up is testing the individual CCX's.

I like how even with all core OC's it puts the cores not being used to sleep and sips very little power when idle.
 
Max 5 min loop R20 stable settings with stock vcore on a 3970x. I think the 3970x has no shitlet lol, they are all pretty close. I'd say it's a pretty good OC for stock voltages, won't be increasing those until it's water cooled.
 

Attachments

  • cinebench-r20-stock-vcore-oc.jpg
    cinebench-r20-stock-vcore-oc.jpg
    249.6 KB · Views: 0
I notice your CCD 2 on Chip 2 appears to be inline with CCD 1 and 2 from Chip 1 for temperatures. Maybe CCD 1 on Chip 2 is bringing that chip down?

Those voltages seem in line with those speeds - in general. Definitely seems like you'd need further thermal headroom to really know what you've got.

If you didn't try it this way the first time, maybe do the per CCX OC again, but only OCing 1 at a time and leave the remainder on stock (or declocked) values. Voltage will still be up there, so it might not make a difference.

I'm using conductonaut for TIM, maybe something like that could help?

Also, maybe manually clock and volt both CPUs the exact same - same test, same duration. If the solder was questionable I think you'd see the majority of that temp difference carry over into this test.

How do you find the build week?

I'll try exactly some of that testing this weekend. The build week is right on the chip as the second line of text I believe. My newer one says BF 1936PGS. Older one says BF 1926SUT.
 
I've actually had nearly the entire 3000 series in my hands at some point. I do have a 3950X inbound as well. The only CPU's I haven't had the pleasure of working with are the 3rd generation Threadrippers.

As for reviewers not getting into some of those details, it's a matter of time. Before a launch, we barely have two weeks to gather our data, learn to overclock, deal with problems with the platforms, BIOS revisions etc. and all of that. I had just 10 days prior to the 10980XE launch. I spent a ton of that time trying to get a hold of comparison CPU's. The 3900X launch gave me about as long but that was more complicated. Some of the finer details don't get seen by us until we've had the CPU's for awhile, or unless we throw them in our personal machines. A factor in something like examining differences in thermals between Malaysian, China, etc. comes down to the fact that we get our allocation of CPU's prior to launch. We rarely get any after that. We'd have to try and source the various CPU's from retail and that costs money. Few sites make enough to just do that outside of isolated cases.

To investigate this, I'd really need to get ahold of a CPU that's the same and assembled in each location. Then I'd have to test, delid and retest. That's a crap ton more work than it might sound like to you guys.

I may be convinced to donate 2 chips for testing if I knew I would get them back in good working order so that I could keep the best one and sell the other. I'm thinking that one of the chiplets is not making good contact with the heatspreader on my newest chip.
 
I may be convinced to donate 2 chips for testing if I knew I would get them back in good working order so that I could keep the best one and sell the other. I'm thinking that one of the chiplets is not making good contact with the heatspreader on my newest chip.

That chiplet might just be awful. You normally get one good chiplet and one shitlet.
 
That chiplet might just be awful. You normally get one good chiplet and one shitlet.

The thing is, the high temp chiplet is CCD1, the good chiplet. It's the one that boosts the highest, but it's also very hot maxing at 85 on a 5 minute CB20 run at stock. CCD2 hits 80. I wouldn't think I should hit 85 on 5 minutes of CB20 at stock with my AIO with agressive fan curves. The older chip actually scores higher on the same test and maxes at 73.5, even though it seems to need more voltage for the same clocks as the new chip. It just seems out of whack.
 
I need to do a big writeup on this as I have now collected tons of data. However, the write up will focus on the original chip. The second chip is trash. I honestly don't think it would even be viable with the stock cooler. It will reboot the machine due to over temp warnings as though it can't even throttle quickly enough. I've reseated it with new TIM 4 or 5 times and I've tried an IC graphite pad. Doesn't matter. It just runs too hot. I've gotten the first chip to run very cool with a new TIM application and pull off some very low voltage runs with no clock stretching. It may not be a good clocker, but it seems very efficient. I don't think I have the stomach for a 3rd chip.
 
  • Like
Reactions: mikeo
like this
It's been a while since I updated this. Here goes.

Chip #2 was a turd, although it boosted higher using that Ryzen Boost checker script, with any real load, it ran about 10 C hotter than the original chip with all settings equal. It would reboot the machine due to going over temperature if you tried to run Prime 95. I reseated 4 to 5 times and tried and IC graphite pad that ran 4 to 5 C higher than using kryonaut. There was no point in doing anything more with that chip. The IC graphite pad was also 4 to 5 C higher on my 3900x chip #1 as well. Therefore, there was no point doing any more testing with the graphite pad either. I didn't try the thermal grizzly pads.

I had one magical reseating of chip #1 with kryonaut before I tried out the graphite pad. For some reason with that one seating, I got lower temps and never got the temperature spikes that I got with every other seating. This allowed the chip to run faster on lower voltages. I kick myself for ever letting that go for further testing.

Pictures of the magical seating:
20191215_000833.jpg
20191215_000749.jpg
With every other seating of the cooler on the cpu, when running CB20, I can see temperatures hovering say in the mid 60s consistently, but then the max temp will suddenly show a spike to the mid 70s that you can never see in real time. With this seating, that never happened. For all testing, I was running fclk at 1800 and memory at 3600 with 16-16-16-36 timings. SOC was on auto which stayed right around 1.1v. The temperature that i used with the Tctl/Tdie temperature as monitored in HWInfo.
I monitored ambietn temps with a cheap ACURITE digital thermometer in the room. I monitored in Fahrenheit and converted to celcius, so not the best, but you can see the trends. Temps varied during testing from 23.9C to 26.1C and I've used those temps to calculate the temperature deltas. I tested by running CB20 for 300 seconds. The cooler is a 240mm Corsair H100i Platinum SE in push pull with LL120s pushing and Noctua NF A12x25 120 mm fans in pull.

The neatest result I got with this seating was being able complete a CB20 run at 4.25 GHz with voltage of 1.19375v. The score was 7458. I was never able to get anywhere near that low of a voltage stable any other time. When I did that, max temp was 61.4C at ambient 25C for a delta of 36.4C Here's a chart for all of the voltages I ran with corresponding max temps. CB20 scores ranged from 7437 to 7480 across all tests which I consider margin of error. For the record, this chip's best single core score is 519.

4.25 Temp Delta.JPG

With other seatings, the lowest voltage that would complete the same CB20 run at 4.25 GHz was 1.2125v.

I was also able to get stable results at 4.3 GHz with this seating, which I had trouble with before for some reason. Scores ranged from 7545 to 7576, which I consider margin of error:

4.3 Temp Delta.JPG

So, my conclusion, and this is a real shocker with Ryzen ;-) is that the lower the temperature you can keep the chip, the lower the voltage required to run at a given speed. Also, for this chip 4.25 GHz is the sweet spot for efficiency with an all core overclock. Temps and required voltages for stability start to really increase at 4.3 GHz.
 
Back
Top