Ryzen: Strictly technical

[Markfw]

Is it possible to have the timings so tight that it hurts performance ? My default timings@3466 on my threadripper number 4 @3.8 are beating (per thread) my 2700x with supposedly tight timings @4.1. That should not be possible on the same memory, the same OS, and the same software, but running FIXED speed 300 mhz slower.

 

[stAbb]

Some of the timings are now looser then on auto setting while others are indeed to tight.
Gear down mode is a great way to get memory stable at higher frequencies without taking too much of a performance hit.
You now loosened part of your primary timings to compensate for the lack of gear down mode. This is not always the best course of action.

If you'd like to follow my advice, just reset to default and try the timings I wrote down in a previous post. They are ballpark numbers, but they should work.

Only the primary timings and tRDRDSCL/tWRWRSCL are tighter then what I use for my Hynix AFR memory.

Your primary timings work so no need to change them.
A tRDRDSCL/tWRWRSCL setting of 2 is stable according to others so you can try that first.
If those timings don't work, I would loosen the tRDRDSCL/tWRWRSCL timings to 3 before touching anything else.

I will take another look at my previous post as well and make sure I put the timings in there properly.

And remember:
Go for the low hanging fruit first, you might end up so happy that getting more performance isn't worth the effort.
Don't fix what isn't broken unless you got loads of free time.

With kind regards,
stAbb

Edit: Settings in my previous post should be more readable now.

 

[Markfw]

Some of the timings are now looser then on auto setting while others are indeed to tight.
Gear down mode is a great way to get memory stable at higher frequencies without taking too much of a performance hit.
You now loosened part of your primary timings to compensate for the lack of gear down mode. This is not always the best course of action.

If you'd like to follow my advice, just reset to default and try the timings I wrote down in a previous post. They are ballpark numbers, but they should work.

Only the primary timings and tRDRDSCL/tWRWRSCL are tighter then what I use for my Hynix AFR memory.

Your primary timings work so no need to change them.
A tRDRDSCL/tWRWRSCL setting of 2 is stable according to others so you can try that first.
If those timings don't work, I would loosen the tRDRDSCL/tWRWRSCL timings to 3 before touching anything else.

I will take another look at my previous post as well and make sure I put the timings in there properly.

And remember:
Go for the low hanging fruit first, you might end up so happy that getting more performance isn't worth the effort.
Don't fix what isn't broken unless you got loads of free time.

With kind regards,
stAbb

Edit: Settings in my previous post should be more readable now.

Thanks, reset to those, testing.

 

[formulav8]

As stated, there are some timings worse now. Did you do an Aida latency/bandwidth check before and after? Some persons get better results with the memory clock speed a little lower but timings tighter.

 

[Markfw]

As stated, there are some timings worse now. Did you do an Aida latency/bandwidth check before and after? Some persons get better results with the memory clock speed a little lower but timings tighter.

It locks up. The conputer works fine on all other apps, but for some reason that test will not run

 

[Markfw]

OK, I suspended BOINC, and now AIDA runs, so here is the result, what do you think of the timings and latency ?

 

[formulav8]

OK, I suspended BOINC, and now AIDA runs, so here is the result, what do you think of the timings and latency ?

Latency of 61ns looks to be very nice for a Ryzen setup. See if you can get tRDRDSCL and tWRWRSCL to 2. Or at least 3

 

[stAbb]

Great start Markfw,
If only our boards would auto to timings like these.


If you feel like tweaking some more:

You could try tRFC at 42, that usually works.
If you want to tighten tRDDL/tFAW, set them to 8 and 32.
You could indeed try tRDRDSCL and tWRWRSCL at 3.


tRTP at 13 looks weird to me with tCWL at 14.
I would lower tRTP to 12 and tRDWR to 6 and see what happens.
(I run 16/12/6 for tCWL, tRTP, tRDWR on Hynix AFR memory @ 3200Mhz)


If you feel really brave, you could lower the tWR to 10 and tRDRDSCL and tWRWRSCL to 2, but I'd be surprised if you still get gains without adding stability issues.
At this point you'd be running tRDRDSCL and tWRWRSCL below JEDEC spec at 2133Mhz and tWR at the lowest possible setting on most boards.
Would love to hear if that both works and still improves performance for your memory.

Have fun!
stAbb

Edit, took out mention of tRFC2/4. See The Stilts post.

 

[Markfw]

I may stay here for a while. My BOINC client went from 3 hours 15, to 2 hours 30 minutes, on tasks which are usually very consistent.

DEFINITELY want to thanks you all for the timing advice. I spent $280 on 16 gig of some of the best money could buy, and the best motherboard, and now its paying off.

 

[IRobot23]

This, https://www.newegg.com/Product/Product.aspx?Item=N82E16820232306
3600 cl15 currently running 3466 cl 14

Voltages?
I think :
- 1.15-1.225V SOC
- 1.45-1.5V ram
- 60ohms

would give you a chance for 3600MT/s

 

[Markfw]

Voltages?
I think :
- 1.15-1.225V SOC
- 1.45-1.5V ram
- 60ohms

would give you a chance for 3600MT/s

Not sure why I would need to up the voltage, working fine at 1.35 ram, and 1.05 SOC. And the 60 ohms ? And where would that "3600MT/s show up, versus what do I have now ?

 

[stAbb]

Glad to hear it's working out for you Markfw!

Your voltages seem very reasonable and your memory is already at 60 Ohm ProcODT.
With my memory I had to lower it to 53 Ohm to get it stable.

If your system becomes unstable you might want try lowering the SOC voltage instead.
Lowering it a bit has helped me stabilize my ram OC. With my Hynix AFR modules at 3200mhz I only need a SOC voltage of 1.0125 volts.
Simply adding more voltage isn't always the answer.

With kind regards,
stAbb

Edit: As The Stilt states, I remembered the max SOC voltage wrong. Nothing to worry about Irobot23.

 

[The Stilt]

I can only see the tRFC.
Did you set tRFC2 and TRFC4 accordingly?
tRFC2=tRFC/1.345 = 280/1.345 = 208
tRFC4=tRFC/2.187 = 280/2.187 = 128

Have fun!
stAbb

tRFC2 / tRFC4 are never used and therefore there is no need to change them.
tRFC2 / tRFC4 determines tRFC during 2x and 4x refresh mode.

The normal refresh mode is 1x and currently you cannot even change that.

A SOC voltage above 1.1 volt will potentially kill your chip over time. 1.15-1225V sounds way to high unless you are working with LN2.

It won't.
Up to 1.200V is perfectly fine, however generally higher than 1.100V makes no difference to the maximum memory clock.
At >= 2800MHz MEMCLK Threadripper CPUs default to >= 1.100V SoC out of the box.

 

[stAbb]

tRFC2 / tRFC4 are never used and therefore there is no need to change them.
tRFC2 / tRFC4 determines tRFC during 2x and 4x refresh mode.

The normal refresh mode is 1x and currently you cannot even change that.



It won't.
Up to 1.200V is perfectly fine, however generally higher than 1.100V makes no difference to the maximum memory clock.
At >= 2800MHz MEMCLK Threadripper CPUs default to >= 1.100V SoC out of the box.

Thank you for clearing that up!

With kind regards,
stAbb

 

[formulav8]

DEFINITELY want to thanks you all for the timing advice. I spent $280 on 16 gig of some of the best money could buy, and the best motherboard, and now its paying off.

No probs. You definitely have yourself a nice setup.

 

[coercitiv]

Hardware Unboxed: Ryzen gaming performance with... The Stilt's timings

 

[CatMerc]

Hoping that with Zen 2 they greatly improve their DDR4 controller. Memory performance is the single greatest factor elevating Skylake above Zen in gaming. Zen+ low latency caches and refinements to the controller already elevate it quite a bit. And if I'm right about the DDR5 in Zen 2, I hope it will be a useful test vehicle to get consumer DDR5 right the first time.

 

[IRobot23]

Hardware Unboxed: Ryzen gaming performance with... The Stilt's timings

Some results are weir, but okay.

 

[epsilon84]

Some results are weir, but okay.

Which results?

Seems like pretty good scaling with tight timings, similar to what computerbase.de achieved

Would have liked to see Intel tested with the better timings too for a fair comparison.

 

[.vodka]

Hardware Unboxed: Ryzen gaming performance with... The Stilt's timings

I think mindblanktech did a better job than Steve did and a few days before, too... not to mention he was the first to get this knowledge out of the forums and into media like youtube last year on Ryzen 1xxx, kind of a pioneer. Only recently have bigger channels like GN and HWU covered the thing.

 

[coercitiv]

I think mindblanktech did a better job than Steve did and a few days before, too... not to mention he was the first to get this knowledge out of the forums and into media like youtube last year on Ryzen 1xxx, kind of a pioneer. Only recently have bigger channels like GN and HWU covered the thing.

Better job indeed, especially by adding low latency results for the Intel side as well. Hopefully (thorough) memory speed & latency testing makes a comeback in the review media.

 

[CrushCrush]

thank you

 

[wahdangun]

So I'm keep seeing that you can overclock using bclk and can make the cpu turbo to 4,5 ghz, so anyone can verify it ?

 

[IRobot23]

Which results?

Seems like pretty good scaling with tight timings, similar to what computerbase.de achieved

Would have liked to see Intel tested with the better timings too for a fair comparison.

What mindblank and thestilt posted... here is like 3-5%.

 

[LightningZ71]

So I'm keep seeing that you can overclock using bclk and can make the cpu turbo to 4,5 ghz, so anyone can verify it ?

It appears that you can use bclk overclocking to manage to overclock the processor while still maintaining the processor's own automatic overclocking abilities. Depending on the silicon lottery, and the power delivery of your motherboard, single core boost can take an excursion over 4.5 GHz.

There are problems with it though. Past about 103 MHz, the PCI-E bus drops back to 2.x speeds and some NVMe drives begin to have controller issues. The processor PB/PB2 tends to stop a few MHz past that from what I've read as well.

I'm actually quite interested in seeing how a 2400g behaves with Bclk overclocking. I want to think that the overall SOC overclock will prove to be a nice improvement to system performance. With little need for add in cards, PCIE behavior may be different and not impactful. Bclk OC also seems to be helpful in memory overclocking as well.