Ryzen Owners Zen Garden

[Mussels]

3100 x 2 is 6200

Fabric has no bearing on anything as long as it's 2000 minimum, there's no relationship since it's all out of sync

people did find the fabric mattered, there is a sync. Originally it was 2033 with 6000MT/s, but AMD fixed that internal math problem and 2000 became the faster/sync setting
That math error having it one setting off made early testing wrong on that hunt.

buildzoid has a few videos on it before the fix and i've had it confirmed by ir_cow (our RAM reviewer)

I'm just trying to find out what the best values for each platform are, for... reasons.

 

[AsRock]

Yeah i just assumed it wouldnt run that high
I look forward to getting my hands on some AM5 stuff and finding out the best values there... i've learned its not worth running out of sync, period.
Find that 1:1 sweetspot and live in it.


With 6100MT/s, you'd want 3100/2050/3100, no?

The math trips me out unless i double check, with it now being 1.5:1? instead of 1:1

Running it 6000\30 run 79ms.

 

[Mussels]

Running it 6000\30 run 79ms.

If you have the time fire up a few comparisons and the results, but don't stress if you can't be bothered.
The few people discussing it (Like buildzoid) was always as a side concern they never went into detail on - in his case, he just showed that above 2033 performance got worse and worse, if he didn't also increase the RAM speed with it
Oh and some MSI boards were running 2000/3000/850 which was absolutely trash tier for performance early on, that was fun.


I suppose the best way the math could be described is we went from 1:1 to 2:3 with the DDR5 change, and I assume the new gear is 1:2

1800/1800 (3600MT/s on DDR4)
2000/3000/2000? on DDR5

And that new gear would work like
2000/4000/2000

if that holds up, then any value between DDR5 6000 and 8000 would be slower, unless you tuned in UCLK and MCLK to their new ratios
1800/3600/1800 for DDR5 7200 should in theory have better latency values by being in sync again, but everyone goes by "2000 is best, run 2000" misunderstanding WHY we run 2000



This is where i'm burrowing deeper inception style because some softwares like aida64 can read the multipliers used, like this example with some madman reaching DDR5 9000
Overclocker hits DDR5-9058 memory OC on AMD AM5 platform with new firmware - VideoCardz.com

That math is pretty obvious for the rams speed (4525/9050 with a 1:44 ratio)
But then what ratios of UCLK/FCLK match up best with that ratio?


My AM4 system runs 58:3 and now i wanna learn how that all works out - when it was locked it was locked, but with freedom on AM5? We gotta do 'shperiments.

Found on reddit from someone who posted some info at least
7800 to 9000 is where the new 'sweet spot' is, with the second gear. 9000 being the equivalent to 6000, but if you match the U/F clocks you can stay in sync from 7200 and up

Theres a gap between 6000 and 7200 that you cant get good performance from no matter what

That said, latencies on the RAM itself will probably be much worse so it may take time for RAM kits at 9000 to equal 6000, in gaming performance.

 

[TheLostSwede]

Default XMP settings.

 

[Mussels]

NB (FCLK) is at 1866 in the AIDA screenshot - what's it like with the second image settings?

 

[TheLostSwede]

NB (FCLK) is at 1866 in the AIDA screenshot - what's it like with the second image settings?

Yeah, AIDA seems to do that, not sure why.

A couple of easy tweaks later and I won't mess around any more.

NB (FCLK) is at 1866 in the AIDA screenshot - what's it like with the second image settings?

Updated to the latest version, still the same.

 

[Mussels]

I wonder if thats the bug people reported with the values being off, if you feel up for it in the future raise the setting so AIDA's reading matches what you wanted it to be and see if the latency lowers?

you dropped a full ns there, but it still feels slow
i found a random thread on reddit searching for the aida64 bug, and with slower timings this person managed 64ns

This was 2133/6200
he found 2166 slower and less stable

a few people discussing the higher IF/UCLK - I wonder if its brand specific, as in not all of them have corrected the clocks being off?

That final posters images are:

Aida certainly doesn't read his 'NB clock' accurately either


seeing a spread of 59ns to 67ns has me wondering WTAF is going on - that final fast result shows an NB clock of 934, which certainly makes no sense at all

Things like this just drive me nuts, without covering what board/bios the issue was on

There is a bug in current BIOS that forces your RAM to run at 6100 when you set 6000-2033. Because it runs at higher speed, you need more voltages. Set to 2066-2133 if you don’t want to deal with this.

Oh! the memory integrity security option in windows (needs SVM enabled in the BIOS) can screw with the readings, if you have that on it could be related

 

[tabascosauz]

I wonder if thats the bug people reported with the values being off, if you feel up for it in the future raise the setting so AIDA's reading matches what you wanted it to be and see if the latency lowers?

you dropped a full ns there, but it still feels slow
i found a random thread on reddit searching for the aida64 bug, and with slower timings this person managed 64ns

This was 2133/6200
he found 2166 slower and less stable


a few people discussing the higher IF/UCLK - I wonder if its brand specific, as in not all of them have corrected the clocks being off?

That final posters images are:

Aida certainly doesn't read his 'NB clock' accurately either


seeing a spread of 59ns to 67ns has me wondering WTAF is going on - that final fast result shows an NB clock of 934, which certainly makes no sense at all

Oh! the memory integrity security option in windows (needs SVM enabled in the BIOS) can screw with the readings, if you have that on it could be related

Pretty sure it's less voodoo Fabric math and more that unstable Fabric begins error correcting and hurting performance long before it shows any visible signs of instability, bz even had an oscilloscope video for it

If TSME is enabled (which appears to be possible on almost all Ryzen despite AMD trying to segment its PRO line) and Memory Integrity enabled in Windows, both will hurt memory performance primarily in latency penalty

 

[TheLostSwede]

I wonder if thats the bug people reported with the values being off, if you feel up for it in the future raise the setting so AIDA's reading matches what you wanted it to be and see if the latency lowers?

you dropped a full ns there, but it still feels slow
i found a random thread on reddit searching for the aida64 bug, and with slower timings this person managed 64ns
View attachment 311411

This was 2133/6200
he found 2166 slower and less stable
View attachment 311412


a few people discussing the higher IF/UCLK - I wonder if its brand specific, as in not all of them have corrected the clocks being off?
View attachment 311413


That final posters images are:
View attachment 311414View attachment 311415
Aida certainly doesn't read his 'NB clock' accurately either


seeing a spread of 59ns to 67ns has me wondering WTAF is going on - that final fast result shows an NB clock of 934, which certainly makes no sense at all

Things like this just drive me nuts, without covering what board/bios the issue was on


Oh! the memory integrity security option in windows (needs SVM enabled in the BIOS) can screw with the readings, if you have that on it could be related

Now you made me fiddle more...

 

[AsRock]

If you have the time fire up a few comparisons and the results, but don't stress if you can't be bothered.
The few people discussing it (Like buildzoid) was always as a side concern they never went into detail on - in his case, he just showed that above 2033 performance got worse and worse, if he didn't also increase the RAM speed with it
Oh and some MSI boards were running 2000/3000/850 which was absolutely trash tier for performance early on, that was fun.


I suppose the best way the math could be described is we went from 1:1 to 2:3 with the DDR5 change, and I assume the new gear is 1:2

1800/1800 (3600MT/s on DDR4)
2000/3000/2000? on DDR5

And that new gear would work like
2000/4000/2000

if that holds up, then any value between DDR5 6000 and 8000 would be slower, unless you tuned in UCLK and MCLK to their new ratios
1800/3600/1800 for DDR5 7200 should in theory have better latency values by being in sync again, but everyone goes by "2000 is best, run 2000" misunderstanding WHY we run 2000



This is where i'm burrowing deeper inception style because some softwares like aida64 can read the multipliers used, like this example with some madman reaching DDR5 9000
Overclocker hits DDR5-9058 memory OC on AMD AM5 platform with new firmware - VideoCardz.com

View attachment 311378

That math is pretty obvious for the rams speed (4525/9050 with a 1:44 ratio)
But then what ratios of UCLK/FCLK match up best with that ratio?


My AM4 system runs 58:3 and now i wanna learn how that all works out - when it was locked it was locked, but with freedom on AM5? We gotta do 'shperiments.
View attachment 311379

Found on reddit from someone who posted some info at least
7800 to 9000 is where the new 'sweet spot' is, with the second gear. 9000 being the equivalent to 6000, but if you match the U/F clocks you can stay in sync from 7200 and up

Theres a gap between 6000 and 7200 that you cant get good performance from no matter what

That said, latencies on the RAM itself will probably be much worse so it may take time for RAM kits at 9000 to equal 6000, in gaming performance.

All i can tell you by only changing the DDR speed from 6000 to 6200 improved the ns by about 5ns. i left all the other stuff default ( auto ) and was just fiddling to see if it actually improved it for my self and it did a little.

But by what @tabascosauz was saying it am not going get much or much better due to the ram needing more voltage so.

EDIT: and CB23 was turning out better numbers by a little constantly with it set to 6200 over 6000.

 

[tabascosauz]

All i can tell you by only changing the DDR speed from 6000 to 6200 improved the ns by about 5ns. i left all the other stuff default ( auto ) and was just fiddling to see if it actually improved it for my self and it did a little.

But by what @tabascosauz was saying it am not going get much or much better due to the ram needing more voltage so.

EDIT: and CB23 was turning out better numbers by a little constantly with it set to 6200 over 6000.

Did you run TM5 to verify at all? Unstable memory also causes wack AIDA numbers especially if you only ran AIDA once each time and took the numbers off that.

If you are just chasing latency (or any specific result), you can double click the latency box just to run that part again. Usually gets better.

 

[freeagent]

If you are just chasing latency (or any specific result), you can double click the latency box just to run that part again. Usually gets better.

I was wondering how people did that

 

[tabascosauz]

I was wondering how people did that

good lord how much time have you wasted running all the tests over again lol

 

[freeagent]

good lord how much time have you wasted running all the tests over again lol

Waaay too much!

I honestly thought guys were just taking a screenshot as the test was running

 

[AsRock]

Did you run TM5 to verify at all? Unstable memory also causes wack AIDA numbers especially if you only ran AIDA once each time and took the numbers off that.

If you are just chasing latency (or any specific result), you can double click the latency box just to run that part again. Usually gets better.

No TM5, although did OCCT on system and 95% of the ram for 15 minutes. How ever RDO crashed after 6 hours green screen reboot which i have not seen for months so changed it back now.

 

[freeagent]

Asus X570 is fat and lazy. Gets wheezy.

Asus upper end B550 is sporty and trim.

There, I said it

At least in the world of Aida and SuperPi..

 

[Mussels]

Pretty sure it's less voodoo Fabric math and more that unstable Fabric begins error correcting and hurting performance long before it shows any visible signs of instability, bz even had an oscilloscope video for it

If TSME is enabled (which appears to be possible on almost all Ryzen despite AMD trying to segment its PRO line) and Memory Integrity enabled in Windows, both will hurt memory performance primarily in latency penalty

True true, error correction lowering performance is a very good point. I'll have to remember that.

Asus X570 is fat and lazy. Gets wheezy.

Asus upper end B550 is sporty and trim.

There, I said it

At least in the world of Aida and SuperPi..

My b550 has far lower memory latency as well, I wonder if the newer chipsets just have better optimised defaults?

 

[freeagent]

True true, error correction lowering performance is a very good point. I'll have to remember that.


My b550 has far lower memory latency as well, I wonder if the newer chipsets just have better optimised defaults?

Using the exact same settings and she is maxed out If I go tighter she plays the don’t wanna post game

 

[Mussels]

I posted it elsewhere, i'm not sure i have the screenshots around.

The tuned 2xDR 64GB in mine manages 64-65ns, while the B550 with dirth cheap 2xSR for 32GB manages <60ns off XMP alone

I'll have to put the effort in with a RAM swap and see if its the RAM itself, the board (shorter traces?), or BIOS settings one day

 

[Space Lynx]

Does anyone know what node Ryzen 8800x3d is using? Because roadmap shows 2024 release, but Apple owns ALL of the 3nm TSMC node for 2024... so this is going to be a 5nm or 7nm chip?

 

[AusWolf]

Does anyone know what node Ryzen 8800x3d is using? Because roadmap shows 2024 release, but Apple owns ALL of the 3nm TSMC node for 2024... so this is going to be a 5nm or 7nm chip?

I wouldn't count on anyone "knowing" anything before an official announcement is made.

 

[tabascosauz]

Does anyone know what node Ryzen 8800x3d is using? Because roadmap shows 2024 release, but Apple owns ALL of the 3nm TSMC node for 2024... so this is going to be a 5nm or 7nm chip?

You buy CPUs based on......what node they're made on?

Probably more productive to hope that whatever node they use is appropriate and mature enough to meet their clock expectations consistently in all CPU samples, with reasonable thermal and power efficiency in that range.

Kinda odd to expect AMD to regress to N7 for whatever reason. When it's already a N5+N6 product currently.

 

[TheLostSwede]

Does anyone know what node Ryzen 8800x3d is using? Because roadmap shows 2024 release, but Apple owns ALL of the 3nm TSMC node for 2024... so this is going to be a 5nm or 7nm chip?

Most likely some N4 variant, which is an optimised version of N5.

TSMC N3, And Challenges Ahead

A Look At TSMC N3 Process

fuse.wikichip.org

 

[Space Lynx]

You buy CPUs based on......what node they're made on?

Probably more productive to hope that whatever node they use is appropriate and mature enough to meet their clock expectations consistently in all CPU samples, with reasonable thermal and power efficiency in that range.

Kinda odd to expect AMD to regress to N7 for whatever reason. When it's already a N5+N6 product currently.

wouldn't the same machines that make N3 node be the ones that made n4 and n5/n6? they don't build a new factory for new nodes very often to my knowledge. I don't know how these things work, and no I don't actually care what node it is on, I am simply curious cause this is an interesting hobby, especially with Apple making power moves/flexing an entire node for a year solid.

Most likely some N4 variant, which is an optimised version of N5.

View attachment 311832

TSMC N3, And Challenges Ahead

A Look At TSMC N3 Process

fuse.wikichip.org

Interesting chart, thanks. They may have problems, but they won't disappoint their trillion dollar overlord Apple, 3nm will be in mass supply next year I am certain. So 9900x3d will be the 3nm node is logical.

 

[tabascosauz]

wouldn't the same machines that make N3 node be the ones that made n4 and n5/n6? they don't build a new factory for new nodes very often to my knowledge. I don't know how these things work, and no I don't actually care what node it is on, I am simply curious cause this is an interesting hobby, especially with Apple making power moves/flexing an entire node for a year solid.

Interesting chart, thanks. They may have problems, but they won't disappoint their trillion dollar overlord Apple, 3nm will be in mass supply next year I am certain. So 9900x3d will be the 3nm node is logical.

Vcache might still be on an old node, if that's what you were getting at. Totally slipped my mind that 7000X3D no longer stacks like-on-like and continues to use N7 for Vcache. If cache not shrinking created a big enough challenge to disaggregate RDNA into chiplets, dunno how much longer this 7nm Vcache is going to last. There was a die shot somewhere that already showed the cache die not fitting neatly into the N5 CCD anymore and the design changes they had to make to get around that.

It's better all-around for efficiency and consistency if they stick to optimizing a similar node as last gen, keeping the same clock envelope and going after bigger cores instead.