AMD Ryzen 3000 Builders Thread

[coercitiv]

Does that mean that if I buy a Ryzen 3000 series CPU I should be buying JEDEC spec'd 3200 with real DDR4-3200 chips and not memory that is factory overclocked and only DDR4-3200 through XMP if I want to ensure stability? I was thinking about buying a Ryzen 3900x and I'm wondering if I should buy some real JEDEC spec'd DDR4-3200 despite the higher latencies? Am I overthinking this XMP/JEDEC situation?

You buy 3200 RAM to ensure compatibility. To ensure stability for 3200 - 3600Mhz operation you normally just enable XMP in BIOS and most of the time everything is configured properly, passing stability tests. However, it's still technically an overclock and it still relies on silicon quality and properties of the memory sticks (configuration, chip types etc).

What some refrain from saying here is Intel CPUs aren't guaranteed to hit the max clocks that modern memory offers either. Just because memory makers sold 4266+ RAM a year ago did not mean that every Coffee Lake CPU was capable of going over 3866Mhz with good timings. And that was ok since nobody really needed those memory clocks on CFL anyway, not normal consumers anyway.

And what's that got to do with Zen 2? Well... let's take a look:

See the 1-2% difference between 3000 CL16 and 3200 CL14 / 3600 CL16? Want that 1%? Then take the pain, read the forums, waste the money, waste the nights, and complain like a true Anandtech member!!!

 

[Snarf Snarf]

So peeps lucky enough to grab a 3900x, is it true one chiplet boost higher than the other thus better binned? Or is it due to workload?

In a sample size of 3 (my own and 2 friends, all pre-ordered together) all 3 show that 1 chiplet is "elite" bin and the second is either a "good" bin or a "meh" bin. My coworkers does 4.5 all core turbo but gets a little too hot for a daily driver at around 92C on a Dark Rock Pro 4 so his is a golden sample. Mine and my other friends both will do about 4.5 on one CCX and 4.4 on the second CCX of CCD1, second chiplet maxes at around 4.35.

 

[Markfw]

Does the XMP setting on that memory also add voltage to the IMC? I've come across a few memory kits in the past, including the one I'm currently using that add voltage to the IMC if using the XMP setting. Do Ryzen 3000's work fine with DDR4-3600 without adding additional voltage to the IMC?

Well, you could actually try the 1.35, I just set it to 1.375 to guarantee stability. They say it can go as high as 1.5, but I would never do that. Maybe 1.4 if need be. And 3733 (1866 IF) is as high as it does you any good. Beyond that it uses a divider, and is a lost cause.

 

[Interitus]

This set looks to be kinda amazing given the timings and the speeds.

https://www.newegg.com/g-skill-16gb...cription=neo&cm_re=neo-_-20-232-856-_-Product

That cheap set though, that cant be B-Die can it?

I don't know that GSkill ever said it was all B-Die or not. I can't imagine that bargain kit is. Im sure someone will grab one and report later. Im curious as to what the thought process is behind a lot of the SKUs GSkill has with the Neo line. Like some are obvious speed/latency improvements or gap fillers for Ryzen 3000, but some are just... why?

If your want b-die and you can wait a week or so it's available for backorder for $139.99 on Amazon currently. Shows eta of the 1st. 3rd party seller, but fulfilled by Amazon.

https://www.amazon.com/gp/offer-listing/B06XFT7DF9/ref=dp_olp_all_mbc?ie=UTF8&condition=all

I bought the same kit from newegg a couple of days before the 3xxx series launch for $129.99 and mine took to 3600 CL14 with a couple of mouse clicks. I did have to go back in and set all the sub-timings to shave a few ns off the latency.

Im more than likely going to grab the Neo 32GB 3600 16-16-16 kit when I see some US pricing. Everywhere I've seen Neo listed, the 16GB Kit of 3600 CL14 is more expensive than the 32GB 16-16-16 kit I'm after.

So the 32GB 3600 16-19-19 kit is on Newegg for $209. The 16GB 3600 CL14 is on Newegg for $280.

Pretty sure I can swallow 32GB of B-Die for $210-$280. If it lands at $250'ish, I'm a happy dude.

 

[Space Tyrant]

I have 2 of the G.Skill Flare X 3200CL14 16GB memory kits. One is dated April 2017, the other is from June 2019. I did all of my initial testing on the 3600X with the old kit. With an aggressive set of subtimings on top of the XMP profile, that kit would work at 1566 MHz. To do that successfully, I had to bump the memory voltage from the XMP 1.35 to 1.4v.

Today I swapped in the 2019 kit while leaving all of the other system settings unchanged. It will run at 1600 MHz at the stock 1.35v. Two small wins for the new kit.

Maybe it's just my samples or maybe the new kits are marginally better than the ones from 2 years ago.

Also, with both systems (2700x, 3600x) configured as close as I could reasonably get them (pstate 4GHz OC, DDR 3200, identical timings, mostly identical subtimings, same OS and kernel, SMT disabled) I ran Geekbench 4.3.4. The results are in my sig. Note that despite the 33% core advantage, the 2700X wins in multi-core by less than 6%. In single-core, the 3600X wins by 12.6%.

 

[Dave3000]

I have 2 of the G.Skill Flare X 3200CL14 16GB memory kits. One is dated April 2017, the other is from June 2019. I did all of my initial testing on the 3600X with the old kit. With an aggressive set of subtimings on top of the XMP profile, that kit would work at 1566 MHz. To do that successfully, I had to bump the memory voltage from the XMP 1.35 to 1.4v.

Today I swapped in the 2019 kit while leaving all of the other system settings unchanged. It will run at 1600 MHz at the stock 1.35v. Two small wins for the new kit.

Maybe it's just my samples or maybe the new kits are marginally better than the ones from 2 years ago.

Also, with both systems (2700x, 3600x) configured as close as I could reasonably get them (pstate 4GHz OC, DDR 3200, identical timings, mostly identical subtimings, same OS and kernel, SMT disabled) I ran Geekbench 4.3.4. The results are in my sig. Note that despite the 33% core advantage, the 2700X wins in multi-core by less than 6%. In single-core, the 3600X wins by 12.6%.

Does your June 2019 kit have faster memory chips than your April 2017 kit?

 

[Space Tyrant]

Does your June 2019 kit have faster memory chips than your April 2017 kit?

They are both Samsung B-die kits with identical timings and (G.Skill) part numbers (F4-3200C14D-16GFX). The only claimed difference is the date of manufacture, afaik.

 

[beginner99]

And what's that got to do with Zen 2? Well... let's take a look:

See the 1-2% difference between 3000 CL16 and 3200 CL14 / 3600 CL16? Want that 1%? Then take the pain, read the forums, waste the money, waste the nights, and complain like a true Anandtech member!!!

Take it with a huge grain of salt and why? The single channel results are telling.

 

[Shamrock]

For anyone with an ASRock Taichi X470, they have a new BIOS out, 3.50

https://www.asrock.com/mb/AMD/X470 Taichi/index.us.asp#BIOS

Update AMD AGESA Combo-AM4 1.0.0.3

 

[lopri]

I 100% agree and it is my main gripe with AMD. Intel doesn't seem to have the same memory issues.

This has got to be a philosophical thing. AMD's memory controller has been stuck at 1800 MHz since the Hammer days. Someone must REALLY love that number.

PS. this post was written using DDR4 4000 CL17 because I miss my blue Corsair DDR 400 dual channel kit and burning memtest on a CD-ROM for optical precision memory testing.

Is that Corsair BH-5 or TCCD?

IMO get 3600 cl16, use XMP on it. Set dram voltage at 1.375. Thats what I have on all my Ryzens

You are so boring.

 

[lopri]

That is my favorite heatsink as well. I always use down-blowing HSF for a long term after my initial testing with towers. Gotta cool those motherboard components. I do not want to spend that much $$ on it, though.

 

[Snarf Snarf]

Solid news everyone, in the final minutes of my night last night, after 2 weeks of tinkering I finally got my x570 Crosshair VIII Hero to boost to the correct clocks under stock settings.

After a lot of poking around on reddit there was a theory going around that on some x570 boards that were having boost issues there was a correlation to people running 3600 MHz RAM and max clocks of 4.3 GHz. This sounded a bit far fetched and went into testing with 0 expectations of this being true but after lowering my RAM to 3200 CL 14 and going full default BIOS settings for stock behavior with no DOCP, manual DRAM clock to 3200 and timings to 14, I saw my first ST boost to 4575.

Reproduced in both CPU-Z and Cinebench R20 ST test as long as the workload sits on Core 1 or 3 on my system the boosts stay high in the 4550-4575 range without dropping. I do notice occasionally the scheduler will move the thread to Core 2 which is one of my weak core and it kills the boost to 4.35-4.4. Previous highest ST boosts were 4300 with all core at around 4.1-4.0.

Will be doing more testing tonight at 3600 IF 1800 to confirm the difference and what kind of tweaking I can do with SOC voltage to get the stock boost back at the higher memory clocks.

 

[Elfear]

Solid news everyone, in the final minutes of my night last night, after 2 weeks of tinkering I finally got my x570 Crosshair VIII Hero to boost to the correct clocks under stock settings.

After a lot of poking around on reddit there was a theory going around that on some x570 boards that were having boost issues there was a correlation to people running 3600 MHz RAM and max clocks of 4.3 GHz. This sounded a bit far fetched and went into testing with 0 expectations of this being true but after lowering my RAM to 3200 CL 14 and going full default BIOS settings for stock behavior with no DOCP, manual DRAM clock to 3200 and timings to 14, I saw my first ST boost to 4575.

Reproduced in both CPU-Z and Cinebench R20 ST test as long as the workload sits on Core 1 or 3 on my system the boosts stay high in the 4550-4575 range without dropping. I do notice occasionally the scheduler will move the thread to Core 2 which is one of my weak core and it kills the boost to 4.35-4.4. Previous highest ST boosts were 4300 with all core at around 4.1-4.0.

Will be doing more testing tonight at 3600 IF 1800 to confirm the difference and what kind of tweaking I can do with SOC voltage to get the stock boost back at the higher memory clocks.

Very interesting. That's the exact ram speeds I'm running. I'll have to go do some boost testing of my own now.

 

[misuspita]

For anyone with an ASRock Taichi X470, they have a new BIOS out, 3.50

https://www.asrock.com/mb/AMD/X470 Taichi/index.us.asp#BIOS

Update AMD AGESA Combo-AM4 1.0.0.3

I just checked my B450m and it has same update, so they updated the whole line. Nice

 

[zinfamous]

I mostly agree with your stance on the advertised boost speed being an issue right now, but this is being too pedantic. I've seen these half or 3/4 frequencies many times on my first gen and it's not unreasonable to round those up for marketing.

Ha anyone successfully sued WD or Samsung or crucial for selling them a 500gb hard drive that is actually 488 gb or, well, every other number just below 500?

 

[UsandThem]

Ha anyone successfully sued WD or Samsung or crucial for selling them a 500gb hard drive that is actually 488 gb or, well, every other number just below 500?

Probably.

But I remember reading articles decades ago used over the formatted or usable space, so every drive since then has a footnote about the available space being less than the size of the drive.

 

[PotatoWithEarsOnSide]

From what I gather, those experiencing issues are using RAM clocked at 3600MHz or higher. Below that, with everything at stock, there are plenty of reports of people hitting the advertised clocks.

 

[DrMrLordX]

Hmm. Wonder why the higher RAM speeds are interfering with boost clocks?

 

[PotatoWithEarsOnSide]

Hmm. Wonder why the higher RAM speeds are interfering with boost clocks?

My guess is the IF causing heat buildup when clocked higher.

 

[guachi]

The Zen of Zen 2:

The less you try, the better the results.

Don't manually overclock, just put good cooling in your system and let it boost on its own.
Don't put DDR4 3600 in your system, just put DDR4 3200, the stated maximum supported speed, in your system. Leave your BIOS alone.

 

[Markfw]

The Zen of Zen 2:

The less you try, the better the results.

Don't manually overclock, just put good cooling in your system and let it boost on its own.
Don't put DDR4 2600 in your system, just put DDR4 3200, the stated maximum supported speed, in your system. Leave your BIOS alone.

close. But 3566 is the fastest if you want max performance. 3600 or more, and you won't get max boost, but you will get better memory speed than 3200. and low CAS as well, like 15

 

[Pandora's Box]

On my Asrock X470 Taichi with 3.46 beta bios, I am still getting 4575MHz boost clocks with my ram at DDR4-3600 and the infinity fabric at 1:1 with ram. *shrug*, I'm not messing with my setup as it appears to be running smoothly.

 

[Kenmitch]

My guess is the IF causing heat buildup when clocked higher.

I'd try it on my 3600, but @3600 memory it'll boost to 4200 on single core CB R20 already.

 

[Snarf Snarf]

close. But 3566 is the fastest if you want max performance. 3600 or more, and you won't get max boost, but you will get better memory speed than 3200. and low CAS as well, like 15

I'll be trying 3566 with CL 14 tonight and compare that to my previous scores to see how it stacks up. As much as I loved the all core speed at 4.4 manual, seeing the voltage go down to sub .9v at desktop made me a lot happier.

Ninja edit: As an amendment to my previous post about boost clocks, this applies specifically to the 3900x on x570 with RAM 3600 MHz or faster. Very specific and small problem but I finally can stop slamming my head into the wall over stock boosts.

 

[Kenmitch]

On my Asrock X470 Taichi with 3.46 beta bios, I am still getting 4575MHz boost clocks with my ram at DDR4-3600 and the infinity fabric at 1:1 with ram. *shrug*, I'm not messing with my setup as it appears to be running smoothly.

Why settle? That last 25MHz might increase your fps by .6 or even .9 depending on the game. Might even get another couple fps if your a 720P gamer! /s