kaliz

looks like a voltage issue somewhere, if you bluescreen btw! try running TM5 Usmus and look what the errors do by lowering or upping the voltage (any voltage really)

you never want to use someone elses RTT's, ProcODT and drivestrenght. As it is board dependent and CPU dependant. Leave all those at auto specially procODT and RTTnom. 

it is because you have 2x32GB M-die and nother CPU and mobo, that my settings in that section are different.

the other timings are setpretty conservative, cant believe it wont work for you

btw guys do not bother with Hydra Pro, that is really a bad software. It tests with Cinebench single core ... and does a per core Curve Optimizer. So if it is finished after a long long time, you get extreme curves which will not work 100%

What i do to test my Curves is use this program : SMUDebugtool 1.3.4 , found it somewhere on overclocknet. It lets you play with PBO in windows. Then to test it, i run Ycruncher (something really heavy on the CPU and RAM) and lower the curve alltogether.

then start tweaking core by core. After it is done set the whole curve 2 or 4 points higher just to make sure. This isnt waterproof but 10x faster and way more reliable!

 just rebooted and set 102.5 bclk in bios for this!

6 minutes ago, ENTERPRISE said:

Thanks for your input, super helpful. I have 2133 FCLK stable and will test performance with RAM at 6400 to see how that plays out against performance I get now with 6600. 

 

Currently stability testing 2167 FCLK. I am able to set 2200 and do benchmarks but GSAT was unstable. I did increase SOC voltage to 1.273 but to no avail but may need to revise the VDDG IOD and CCD a little higher to see if that can help get 2200 FCLK stable.

So if you have 2167 stable, what you can do is:

set 6400 : 2133 ratio and play with bclk!

On my cheapo mobo i can boot 102.5 , and run it up to 103,25 in windows with it! For 6600 2200 fclk in 1:1 mode. 

On 23/10/2023 at 09:04, Hurricane28 said:

Je i saw that, amazing. 

 

It won't matter on my chip tho so i don't bother. On single CCX i belive 6400MHz is the max buildzoid said. 

 

Sadly the lower timing didn't work on the 6400 timings. Today i try again. 

I did a 8400C34-45-40-42-72-504 with GDM disabled and all tight timings on a Ryzen 5 7500F single CCX for topscore in Pyprime 2b for non X3D models, it really depends on the RAM and mobo. The AM5 CPU's themselves are capable of around DDR5-8800!! With a 7800X3D managed to run DDR5-8720 and thats about the max this 4 DIMM slot mobo can handle! Imagine a GENE!

1 hour ago, ENTERPRISE said:

So I can confirm the latest AGESA 1.0.0.8 has enabled me to step up FCLK to 2133 all stable. Going to try for 2167 on the off chance. 

 

What parameters would make 1:1 MEMCLK/UCLK more possible, if any ? I might scale back to 6400 to see if I can get it working, 6600 it won't work but I think im asking too much at 6600.

I think 6400 with 2133 fclk in 1:1 mode is best for your setup!. That way it is all synced; 3:1:3 , eg.: 6400:2133:6400 in 1:1 mode. If your CPU is very capable, the next best setup will be 6600 with 2200 fclk, so its again synced for 3:1:3. 

So if your CPU is only capable of running 2133 fclk, i wouldnt even go 6600, but stay at 6400 to have that synchronization. You can test it for yourself ofcourse for speeds, but a AMD engineer said on social media last year that this is the way to do it..

The next option is to run in 1:2 mode, and to negate the latency penalty, you have to run above DDR5-7600. My preference is DDR5-8000 with either 2000 fclk (for 2:1:1) or 8000 with 2200 fclk for max throughput.

you can set VDDG IOD and CCD to 1,050v to runhigher fclk. max will be 1,15v for those 

11 hours ago, J7SC_Orion said:

Per below, I tried tRRDS 8 > 6 and tFAW 32 >24 on the 7950X3D with all else the same...as expected, it did not make any difference at all outside normal run-to-run variance (I used the TimeSpy CPU test as control). IMO, the memory controller just ignores tRRDS below 8 with DDR5 (unlike DDR4)...

 

with my Hynix A-die kits, at DDR5 8000 and beyond, RRD, FAW and WTR can go as low as 8 4 16 16 6 , and for lower freq, 6 4 16 14 4, and if you want to be sure, you can set 8 8 32 24 6.

It does matter in latency and throughput. You can test this with Pyprime , YC, 7zip and those kind of benchmarks. With Hynix M-die you can set 4 4 16 14 4. RRD on A-die needs a bit more. 6 is really good, 7 good, 8 normal.

E-cores are so dumb just cant believe they added more this gen. Benchmarkboosters Buildzoid called them, Hurricane,  dont listen to BZ too much. tFAW 20 is a joke of him .  16 or 32 , or if RRD-s is 6, 24. Rest of the timings look good and i see you get <60ns latency. If you really try, you can get 49ns, 48ns even lower, good luck  

will there be non-K overclocking with the later dual core / quad core models?

The non X3D chips are a bit hotter, 89C is to be expected. What CPU clocks and voltageyou run?

MSI has new beta bios out with option for GDM disabled : this is a very nice OC example 

now i know my 3rd core is bad

its for 4 cores 4 threads only so you have to disabled threads!

just ran 6600 C26-36-36-26-58-390 for the competition

really nice, now you can tighten up the primary timings to 30-38-38-50-88     4-8-16-4-16-(120ns)-384   you can get around up to 70000 read with this CPU and RAM combo!! and about 60ns latency!

to find the best Curve for your CPU this program lets you use Curve Optimizer (PBO) in Windows : SMUdebugTool  

Release v1.3.2: add PBO tab for testing purposes, limited to 16 cores · irusanov/SMUDebugTool

GITHUB.COM

---

add PBO tab for testing purposes, limited to 16 cores and family 19h update core dll to v 1.6.3

If you run a hammering benchmark, like Ycruncher, you can find your limits pretty easily. Also other bencmarks like 7zip are very usefull to find the max curve and optimal RAM frequency/fclk frequency .

those 6400C3239G16G kits are M-die nowadays, for a brief period they put the best Hynix A-die in them from Nov 22 until Mar 23... yours might be newer.   And barcode says it all if last letter is M. 

Try this:

6400 : 2133 : 6400

Very nice review!  The Copper heatsinks look really good. I wish G.Skill had heatspreaders that could be taken apart with a screwdriver.

. it is a struggle to get DDR5 stable, now im wondering , you set tRFC pretty high at around 150ns, what memory dies are in these modules? I cannot tell from the Zentimings if its Samsung b-die, Hynix M or A-die? For 2x32GB 6400C28 is a very good speed , but you forgot to set MEMCLK/UCLK in 1:1 mode. So performance will be skewed a lot.

i have to apologize i think Round trip latency is a Intel thing and AMD does not name it like that! i have found it on my B760M bios but not in B650E nor B650 boards , only training options in Nitro settings menu. You can try 4 taps for a similar effect and tighten down those nitro settings to 1 or 0

Tested a different CPU + Mobo combo and this setup was stable 5 cycles in TM5 for what its worth , its pretty tight, could do 26-36-36 but it gets hot eventually.  you need a fan on top of the ram to keep it <45C VDD 1,69 VDDQ 1,62 anything lower gave errors, also MISC at 1,2 was a test its ok at 1,1v also i left VDDP auto. This is pretty much the settings the guys with AM5 on LN2 bench with @6400 on hwbot, maybe bit looser subs. If you want to use lower voltages, every primary timing needs at least 2 ticks more, but the second and third you can leave

tried again yesterday but it was half second slower  this was with chilled water and the cheapest B650 mobo , too bad it didnt give points on hwbot this was pretty decent for a Ryzen!