My budget 3950x build

[Michael Sherwin]

For my Ryzen 3950x, in the BIOS I set Vcore=1.20 and Multiplier to 40 and don't enable PBO. This gives an all-core frequency of 4.0 Ghz and when running Wasp with 32 threads the total CPU package power is about 132 watts. With the Noctua UH-12S
air cooler the cpu temp stays under 70C. I really like the 3950x.

John

Thank you, that seems very close to optimal! A sight variance may exist due to motherboard and silicon lottery.

Procedure to achieve best all thread overclock in shortest amount of time.

STEP 1
Start with memory overclocking. My motherboards flash bios button does not work if the memory fails so do memory overclocking with an open case. I was able to overclock my memory from 3600 MHz to 3733 MHz and the infinity fabric from 1800 to 1867 with about an 800 kn/sec improvement.

After the memory overclock is finalized makesure you have AMD Ryzen Master downloaded.

STEP 2
Disable PBO and find the cpu vcore voltage that does not exceed 100% PPT, TDC and EDC. On my setup that was 1.1375V.

STEP 3
Then find the highest stable all core multiplier. On my setup it is 40.25.

STEP 4
IMO AMD uses a higher bin chiplet for the first n cores. Not sure if n is 4 or 8. Anyway, switch to assigning frequencies by 4 core groups. Assign 40.25 (in my case) to all 4 groups. Then you can try increasing the multiplier for the first group. Currently I'm testing 41.00 but I interrupted the test to make this post as it needs to test for several hours to be sure.

[Michael Sherwin]

FINAL REPORT
To get the best manual overclock follow the 4 steps above. The 3950x has two 8 core chiplets. The cpu core multipliers must be set separately to get the best manual overclock. The first chiplet is definitely a higher bin chip. At 1.1375V vcore setting I'm now testing:

CCD0 CCX0 @ 4200 MHz
CCD0 CCX1 @ 4200 MHz
CCD1 CCX0 @ 4000 MHz
CCD1 CCX1 @ 4000 MHz

And everything is running fine at the moment. Even as I type this. CPU temperature topped out at 71c. PPT 97%. TDC 100%. EDC 100%.
SF 10 44109kn/sec.

Thanks for following and I hope this has helped someone!

[jstanback]

That's a very nice overclock! The Ryzen's do something called "clock-stretching" where they may slow the clock down to prevent crashing, or something like that. I found that with Vcore below about 1.175V my CPU seemed stable at 4.1Ghz, but my benchmark speeds were slower, or at least no faster. So I upped to the voltage to 1.20 and reduced the speed to 4.0Ghz and seemed to get about the same benchmark speed (it's hard to tell with MT variability). I haven't tried overclocking the memory beyond my default DOCP value of 3200.

John

[Michael Sherwin]

That's a very nice overclock! The Ryzen's do something called "clock-stretching" where they may slow the clock down to prevent crashing, or something like that. I found that with Vcore below about 1.175V my CPU seemed stable at 4.1Ghz, but my benchmark speeds were slower, or at least no faster. So I upped to the voltage to 1.20 and reduced the speed to 4.0Ghz and seemed to get about the same benchmark speed (it's hard to tell with MT variability). I haven't tried overclocking the memory beyond my default DOCP value of 3200.

John

Thanks again! I decided to try 1.2v but it was exceeding PPT wattage and TDC amperage by a little too much for comfort. So I set a goal for no more than 115% for either. So 1.1850v satisfied that goal. This seems to be the sweet spot for my setup. PPT was 113%. EDC was 115%. And this is after about 2 hours of non stop searching. The temperature stayed below 80c at 79c and some change. And the overclock you might not believe!

ccd0 ccx0 4300 MHz
ccd0 ccx1 4300 MHz
ccd1 ccx0 4100 MHz
ccd1 ccx1 4100MHz

SF 45,679 kn/sec

I'm finally satisfied with this result. If it fails after more testing I'll just back off all cores .25 and be satisfied with that, lol.

Thanks again for your input.

[MikeB]

That's a very nice overclock! The Ryzen's do something called "clock-stretching" where they may slow the clock down to prevent crashing, or something like that. I found that with Vcore below about 1.175V my CPU seemed stable at 4.1Ghz, but my benchmark speeds were slower, or at least no faster. So I upped to the voltage to 1.20 and reduced the speed to 4.0Ghz and seemed to get about the same benchmark speed (it's hard to tell with MT variability). I haven't tried overclocking the memory beyond my default DOCP value of 3200.

John

Thanks again! I decided to try 1.2v but it was exceeding PPT wattage and TDC amperage by a little too much for comfort. So I set a goal for no more than 115% for either. So 1.1850v satisfied that goal. This seems to be the sweet spot for my setup. PPT was 113%. EDC was 115%. And this is after about 2 hours of non stop searching. The temperature stayed below 80c at 79c and some change. And the overclock you might not believe!

ccd0 ccx0 4300 MHz
ccd0 ccx1 4300 MHz
ccd1 ccx0 4100 MHz
ccd1 ccx1 4100MHz

SF 45,679 kn/sec

I'm finally satisfied with this result. If it fails after more testing I'll just back off all cores .25 and be satisfied with that, lol.

Thanks again for your input.

Also you can run Windows Memory Diagnostic tool which will stress the memory and if it fails , you will know you're mem oc will eventually give you a BSOD, so just back it down. No need to run the long one, just run the sort one, one time.

Normally , it's used for checking problems after a BSOD, but you can also run it after OC'ing memory, as often, the OC
memory will boot and appear to be fine , but will fail the stress test.

[MikeB]

That's a very nice overclock! The Ryzen's do something called "clock-stretching" where they may slow the clock down to prevent crashing, or something like that. I found that with Vcore below about 1.175V my CPU seemed stable at 4.1Ghz, but my benchmark speeds were slower, or at least no faster. So I upped to the voltage to 1.20 and reduced the speed to 4.0Ghz and seemed to get about the same benchmark speed (it's hard to tell with MT variability). I haven't tried overclocking the memory beyond my default DOCP value of 3200.

John

Thanks again! I decided to try 1.2v but it was exceeding PPT wattage and TDC amperage by a little too much for comfort. So I set a goal for no more than 115% for either. So 1.1850v satisfied that goal. This seems to be the sweet spot for my setup. PPT was 113%. EDC was 115%. And this is after about 2 hours of non stop searching. The temperature stayed below 80c at 79c and some change. And the overclock you might not believe!

ccd0 ccx0 4300 MHz
ccd0 ccx1 4300 MHz
ccd1 ccx0 4100 MHz
ccd1 ccx1 4100MHz

SF 45,679 kn/sec

I'm finally satisfied with this result. If it fails after more testing I'll just back off all cores .25 and be satisfied with that, lol.

Thanks again for your input.

Excellent nps/$ - about one of the best I believe.

[Dann Corbit]

Re: "Excellent nps/$ - about one of the best I believe."
I paid less than $800 for my HP Proliant DL585 G7 that got 47 million NPS on the Asmfish bench. Now we factor in the soundproof box at another $800 and that gives $1600 for 47M NPS. That gives about 30,000 nodes per dollar, which would be hard to beat (though the original price of the machine was astronomical).

On the other hand, if you calculate the thousands of dollars I have spent on electricity, and you will definitely have laid me low.

[MikeB]

Re: "Excellent nps/$ - about one of the best I believe."
I paid less than $800 for my HP Proliant DL585 G7 that got 47 million NPS on the Asmfish bench. Now we factor in the soundproof box at another $800 and that gives $1600 for 47M NPS. That gives about 30,000 nodes per dollar, which would be hard to beat (though the original price of the machine was astronomical).

On the other hand, if you calculate the thousands of dollars I have spent on electricity, and you will definitely have laid me low.

Yes, the $800 for 47M was also excellent , even with the sound proofing. Werewolf would be better off building a cluster with 10 of these 3950 builds - nearly 500 M nodes for $11200 and change.

[Zenmastur]

Re: "Excellent nps/$ - about one of the best I believe."
I paid less than $800 for my HP Proliant DL585 G7 that got 47 million NPS on the Asmfish bench. Now we factor in the soundproof box at another $800 and that gives $1600 for 47M NPS. That gives about 30,000 nodes per dollar, which would be hard to beat (though the original price of the machine was astronomical).

On the other hand, if you calculate the thousands of dollars I have spent on electricity, and you will definitely have laid me low.

Yes, the $800 for 47M was also excellent , even with the sound proofing. Werewolf would be better off building a cluster with 10 of these 3950 builds - nearly 500 M nodes for $11200 and change.

If you all you want is 500M nps processing power it's possible to get it for less than $8000 using Ryzen 9 3900X's. It's cheaper to do it this way because of the prices of the 3900X vice 3950X and the fact that the 3900X comes with a CPU cooler. To keep costs down use Micro ATX MB's and ATX/Micro-ATX case, standard PS and 16GB of DDR4-3200 RAM. The price includes 2 x 8-port KVM switches, keyboard, mouse and cables, and a monitor. You can find name brand Micro ATX B450 MBs on sale for $65 or less. If you want more RAM of course the price per box goes up about $70 per 16GB. If you want RTX video cards the price goes up at least $260 per box depending on which RTX card you want.

Regards,

Zenmastur

[Michael Sherwin]

Hey guys, Just got done with a 24 hour SF "stress test". It passed! It failed the previous overnighter so for the 24 hour test things were done differently. It turns out that it is best to enable PBO. That is because there are settings that PBO controls very well other than core multipliers and core voltages. The 4300 MHz cores 1 - 8 and 4100 MHz cores 9 - 16 was kept. Voltage was increased to 1.2375v though.It ran a bit hotter at 87c. In 24 hours it made it to depth 54 with a node rate of 45,785 kn/sec! Dann made me curious what Asmfish will do on this machine. That report will be next.