First AMD build in 20 years!

That's insane, congrats on those clocks. Very impressive. I think you might be right on silicone lottery for my CPU. Although, I've been trying this morning to do the opposite of what I've read with regards to CO in that I am giving my "bad" cores LESS negative offset and my "good" cores MORE negative offset. Seeing higher clocks, so apparently Makaveli@BETA was on to something. It's just so weird that everywhere else has said to do the opposite.

View attachment 349700

As for my L3, I am on the latest stable for the Dark Hero. Previously I was getting in the 400 gb/s range, L3 has gone up over each successive bios update. There is a beta, but I am going to wait till it goes stable to update. Ram is at 1.35 volts, so pretty happy with those sticks.
That is actually what you WANT to do... the negative curve lowers the voltage setting per clock. Weaker cores need MORE voltage to be stable and boost, while good cores will need less. I can't say I have messed with per-core boosting yet though or if it is even worth it; seems like it may be hard to truly stability test until you get a random boost and crash (or don't).

Those are some awesome DDR4 sticks to hit that speed with those timings at 1.35V! From what I have been reading, anything above IF1900 is still very much an extreme overclock in AMD's eyes and may never be stable. So congrats on 1900IF @ 3800Mhz.
 
Here's my Hynix @ 3733 and 1.4v. It hates 3800, but I was able to get 4000 with looser timings, go figure.

aida64.JPG
 
1619055445201.png



Coming from a 3570k, this is fun lol

Blu-ray movie encodes went from 6 hours to 45 minutes a film lol

My setting for 4K HDR will still take 9 hours a movie, but the 3570k said it wanted 3 days and that's when I was like shit I have to upgrade

I wanted to see what it could do stock before I played with clocks and memory - I'm really not gonna notice any difference in encode times really - I have a silver sample on air (noctua nh-d15 and 2x noctua nf-a20 and 2x noctua nf-a14 cooling it)

Edit: and oh so many 4 core 4gb virtual machines

1619057185076.png
 
Last edited:
So as I posted previously, as my RAM is only "rated" for 3200Mhz 1.35V, I managed to get it fully stable at 3666Mhz at 14-14-14-28 w/ 1.45V, but had to use a 2T command rate. Oddly, that did not hurt latency at all in all my testing.

This morning, I decided to try "Gear Down Mode"; because, why not? I guess its like 1.5T by using the rising edge of the next cycle... lol. Turns out, I'm stable at that mode at 3666Mhz 14-14-14-28 @ 1T (GDM). My bandwidth actually went up by 1000MB/s and my latency went down by 1ns.

Nifty little feature these AMDs have that Intel does not... Not sure how many people use it, but allows me to get every last drop out of my ram. :D
 
So as I posted previously, as my RAM is only "rated" for 3200Mhz 1.35V, I managed to get it fully stable at 3666Mhz at 14-14-14-28 w/ 1.45V, but had to use a 2T command rate. Oddly, that did not hurt latency at all in all my testing.

This morning, I decided to try "Gear Down Mode"; because, why not? I guess its like 1.5T by using the rising edge of the next cycle... lol. Turns out, I'm stable at that mode at 3666Mhz 14-14-14-28 @ 1T (GDM). My bandwidth actually went up by 1000MB/s and my latency went down by 1ns.

Nifty little feature these AMDs have that Intel does not... Not sure how many people use it, but allows me to get every last drop out of my ram. :D
You're not the only one who has had that result. My motherboard had gear down mode enabled to begin with so I disabled it. Stability went down and performance went down so I turned it back on and everything went back to "normal".
 
View attachment 349967


Coming from a 3570k, this is fun lol

Blu-ray movie encodes went from 6 hours to 45 minutes a film lol

My setting for 4K HDR will still take 9 hours a movie, but the 3570k said it wanted 3 days and that's when I was like shit I have to upgrade

I wanted to see what it could do stock before I played with clocks and memory - I'm really not gonna notice any difference in encode times really - I have a silver sample on air (noctua nh-d15 and 2x noctua nf-a20 and 2x noctua nf-a14 cooling it)

Edit: and oh so many 4 core 4gb virtual machines

View attachment 349977

Your config sounds very similar to mine (5950x, 64 GB RAM, NH-D15). Which board? I was overclocking with PBO and getting some pretty high clocks (4.8-4.9 on multiple cores), but I was having temp issues and I found the Fmax Enhancer setting on my Dark Hero was probably causing the issue and temporarily disabled it while I wait for another BIOS revision. Waiting for my 3080 to arrive and when I get it, I'm going to change some fans and might redo thermal paste too.
 
Your config sounds very similar to mine (5950x, 64 GB RAM, NH-D15). Which board? I was overclocking with PBO and getting some pretty high clocks (4.8-4.9 on multiple cores), but I was having temp issues and I found the Fmax Enhancer setting on my Dark Hero was probably causing the issue and temporarily disabled it while I wait for another BIOS revision. Waiting for my 3080 to arrive and when I get it, I'm going to change some fans and might redo thermal paste too.

Asus TUF X570 Pro Wifi 6 - so I won't expect miracles from that either - haven't gotten around to playing with it for OC yet since built and made sure working, just too much other stuff going on atm
 
So this is the absolute fastest memory speeds I can get stable on my sticks... considering I bought them for my x99 platform rated for 3200Mhz CAS 14 and they are 4 single rank sticks total; I can live with this... lol. Have 1.44V on them right now; could probably throw more voltage at it and tighten timings more if I wanted too, but this took enough work as it is for now. When I try 3800Mhz the mobo does not even post and I have to hit the CMOS reset button (best button ever invented BTW). This is of course coupled with my PBO Curve modified with -20 offset all core and +200Mhz FMax (curve power modified at: 200, 200, 150).

I do enjoy doing this kind of stuff on PCs, but to put people to rest, the benchmark and gaming differences between 3200Mhz CAS 14 and 3733Mhz CAS 16 on Zen 3 were mot massive... I gained maybe ~5% tops in the FPS arena (and at the 4k resolution I game at, even less)... but it's nice to have the IF (which connects everything) running as fast as possible, so I'd say boost your speed if you can without sacrificing latency. Probably helped I had a Samsung B-Die kit from the Haswell-e/Broadwell-e era. I am able to get sub 60ns all the time now; which on AMD is quite good from my understanding, trash on Intel tho (I was sub 52ns on my x99 platform)... lol

cachemem_5950x_3733_2_CstatesOFF.png
 
So this is the absolute fastest memory speeds I can get stable on my sticks... considering I bought them for my x99 platform rated for 3200Mhz CAS 14 and they are 4 single rank sticks total; I can live with this... lol. Have 1.44V on them right now; could probably throw more voltage at it and tighten timings more if I wanted too, but this took enough work as it is for now. When I try 3800Mhz the mobo does not even post and I have to hit the CMOS reset button (best button ever invented BTW). This is of course coupled with my PBO Curve modified with -20 offset all core and +200Mhz FMax (curve power modified at: 200, 200, 150).

I do enjoy doing this kind of stuff on PCs, but to put people to rest, the benchmark and gaming differences between 3200Mhz CAS 14 and 3733Mhz CAS 16 on Zen 3 were mot massive... I gained maybe ~5% tops in the FPS arena (and at the 4k resolution I game at, even less)... but it's nice to have the IF (which connects everything) running as fast as possible, so I'd say boost your speed if you can without sacrificing latency. Probably helped I had a Samsung B-Die kit from the Haswell-e/Broadwell-e era. I am able to get sub 60ns all the time now; which on AMD is quite good from my understanding, trash on Intel tho (I was sub 52ns on my x99 platform)... lol

View attachment 353081

i've kept my B-die kit at 3200 with tight timings over going up in speed, while the synethic benchmarks look great application performance doesn't get that large of a boost. And right now on my Rig Stability is the number 1 priority since I'm working from home on it.
 
So this is the absolute fastest memory speeds I can get stable on my sticks... considering I bought them for my x99 platform rated for 3200Mhz CAS 14 and they are 4 single rank sticks total; I can live with this... lol. Have 1.44V on them right now; could probably throw more voltage at it and tighten timings more if I wanted too, but this took enough work as it is for now. When I try 3800Mhz the mobo does not even post and I have to hit the CMOS reset button (best button ever invented BTW). This is of course coupled with my PBO Curve modified with -20 offset all core and +200Mhz FMax (curve power modified at: 200, 200, 150).

I do enjoy doing this kind of stuff on PCs, but to put people to rest, the benchmark and gaming differences between 3200Mhz CAS 14 and 3733Mhz CAS 16 on Zen 3 were mot massive... I gained maybe ~5% tops in the FPS arena (and at the 4k resolution I game at, even less)... but it's nice to have the IF (which connects everything) running as fast as possible, so I'd say boost your speed if you can without sacrificing latency. Probably helped I had a Samsung B-Die kit from the Haswell-e/Broadwell-e era. I am able to get sub 60ns all the time now; which on AMD is quite good from my understanding, trash on Intel tho (I was sub 52ns on my x99 platform)... lol

View attachment 353081

So I got bored and decided to tweak the 3rd level timings like I did on my old system... gained a good amount of bandwidth already by doing so! 😁

cachemem_5950x_3733_4_CstatesOFF.png
 
Yes, I am un-necroing my old thread as an update... lol.

So I got bored and did a thing again, pushing ever closer to the absolute limits... lol.

Sadly, anything above 3800Mhz with a 1:1:1 FCLK=MEMCLK=UCLK is a no go... my FCLK simply will not go higher without Whea 19 errors.

I decided (against all these hardcore AMD OC'ers) to uncouple my UCLK from my FCLK to UCLK=MEMCLK/2. I knew this would add a latency penalty, but through some tweaking (and tight timings), I have so far lowered that penalty to only 5~6ms vs ~10ms. I think once I get to tweaking the 2nd and 3rd level timings, I may get that penalty even lower! It will be a methodical process, but fun for me.

So far, now have these puppies at 4200Mhz @ 18-18-18-36 @1T with only 1.45V.

Initial testing shows mixed results. Cinebench R23, and 3dmark graphics scores are up by a good 3~5%! Superposition is about the same, and 3dmark CPU score was down slightly. Will be fun to bench some games and see if the increase in memory speed and bandwidth actually improves things even with a minor latency bump (which is usually not preferred).

cachemem_5950x_4200_2_CstatesON_Win11.png
 
Yes, I am un-necroing my old thread as an update... lol.

So I got bored and did a thing again, pushing ever closer to the absolute limits... lol.

Sadly, anything above 3800Mhz with a 1:1:1 FCLK=MEMCLK=UCLK is a no go... my FCLK simply will not go higher without Whea 19 errors.

I decided (against all these hardcore AMD OC'ers) to uncouple my UCLK from my FCLK to UCLK=MEMCLK/2. I knew this would add a latency penalty, but through some tweaking (and tight timings), I have so far lowered that penalty to only 5~6ms vs ~10ms. I think once I get to tweaking the 2nd and 3rd level timings, I may get that penalty even lower! It will be a methodical process, but fun for me.

So far, now have these puppies at 4200Mhz @ 18-18-18-36 @1T with only 1.45V.

Initial testing shows mixed results. Cinebench R23, and 3dmark graphics scores are up by a good 3~5%! Superposition is about the same, and 3dmark CPU score was down slightly. Will be fun to bench some games and see if the increase in memory speed and bandwidth actually improves things even with a minor latency bump (which is usually not preferred).

View attachment 521760
Wow congrats for making those 4 sticks to run at that 4200 freq, and with 1.45v too.
Those B-Dies are just too awesome (I have two different b-die kits).
You may tighten those primary timings but will need more than 1.52v and additional fan directly to the ram(s).
 
Last edited:
Wow congrats for making those 4 sticks to run at that 4200 freq, and with 1.45v too.
Those B-Dies are just two awesome (I have two different b-die kits).
You may tighten those primary timings but will need more than 1.52v and additional fan directly to the ram(s).
Yeah, this was an amazing Quad Channel B-Die set I had from my x99 days. All singe rank sticks, so 4 sticks run perfectly on this setup.

I am going for mostly secondary and third level timings right now to see what can be gained back on latency with the voltage i have now. Testing for stability is going to take a lot of time as I am only changing a setting or two at a time. I have airflow on the sticks, but under full Prim95 Blend Test load, it starts to get a little warm at 1.45V as the sticks get just above 40C. Not sure I am willing to push 1.5V on them to get CAS 16.

Mostly looking to compare what are the gains/losses in real world applications and games once I prove full stability at tighter timings. It's just fun for me, not going to make a hell of a difference as I game at 4K... lol.
 
My current 5800x build is my first AMD since moving from Bulldozer to Sandy Bridge. I have zero AMD related complaints. I'm not digging MSI BIOS but I've learned to live it with it.
 
Back
Top