Aorus X570 Extreme (new rev 1.+) + HyperX Predator HX446C19PB3K2/16 + 3950X

[JoeD]

Hiya everyone,

As mentioned in the title, this is the system I’m running.
—-(Processor Temps kept at around the 60c - 70c range with water cooling)
—-(System temps really low at about average of 35-40c )

I Just wanted to know if it’s actually possible to get my memory (4 sticks) HX446C19PB3K2/16 ( 32GB) to actually run at the 4600 xmp profile on my system.

So far I’ve only been able to get stable 4200MHz (4216.80Mhz) with the following timings: 20-26-26-45 tRC 97 tRFC 736

HOWEVER, the 4600MHz XMP profile is: 19-26-26-45 tRC 67 @ 1.5volts

My current bios settings are as follows:

((PROCESSOR Settings ))

CPU lock control : auto (100mhz)
Spread Spectrum: auto
CPU Ratio mode : All cores

CPU Vcore: Auto (1.200v)
Dynamic Vcore (DVID): auto 0.00000V

VCORE SOC: auto ( 1.2v)
Dynamic VCORE SOC (DVID) : auto 0.00000V

CPU VDD18 : auto 1.800V
CPU VDDP : auto

PM_CLDO12: auto 1.200V
PM_1VSOC: auto 1.000V
PM_1V8: auto 1.800V

CPU Vcore Loadline Calibration: auto
Vcore SOC Loadline Calibration: auto
CPU Vcore Protection: auto 350.0mV
CPU VCORE SOC Protection: auto 350mV
CPU Vcore Current Protection: auto
PWM Phase Control: auto

((( XFR Enhancement settings )))

FCLK frequency: 1800 MHz
SOC OVERCLOCK VID : 48
UCLK DIV1 MODE : auto
VDDP Voltage Control : manual
VDDP Voltage : 1150

VDDG Voltage Control: manual
VDDG Voltage: 1050

SOC/Uncore OC Mode : auto
LN2 Mode: auto

Precision Boost Overdrive: manual
PPT Limit: 300
TDC Limit: 230
EDC Limit: 230
Precision Boost Overdrive Scalar: Manual
customized Precision Boost Overdrive Scalar: 4x


((( MEMORY TIMINGS ))) ( both Channel A & Channel B)

CAS latency: auto 19
tRCDRD: auto 26
tRCDWR: auto 26
tRP: auto 26
tRAS: auto 45

tRC: auto 67
tWR: auto 35
tCWL: auto -

tRRD_S: auto 12
tRRD_L: auto 12

tWTR_S: auto 6
tWTR_L: auto 14

tRFC: auto 312
tRFC2: auto 192
tRFC4: auto 132

tRTP: auto 14
tFAW: auto 49
tRCPAGE: auto -
tRDWR: auto 11

tRDRDSC: auto 1
tRDRDSD: auto 4
tRDRDDD: auto 5
tRDRD_SCL: auto 5

tWRRD: auto 1
tWRWRSC: auto 1
tWRWRSD: auto 6
tWRWRDD: auto 7
tWRWR_SCL: auto 5

tCKE: auto 16

ProcODT: auto
Command Rate (tCMD): auto
Gear Down Mode: auto


If anyone could give me some insight and help it’d be awesome! I’m not sure if I need to change the infinity fabric frequency or what..

Thanks in advance!

 

[JoeD]

256 views and not a single reply? come on guys, anyone here able to help me out a bit?

 

[Caring1]

Ryzen DRAM Calculator by 1usmus

Hi guys, I created a calculator that will make it easier to overclock the memory. It works with all kinds of memory and all zen processors. Ryzen DRAM Calculator 1.5.1 Last update : May 13th 2019 >> Download latest version >AMD Ryzen Memory Tweaking & Overclocking Guide

www.techpowerup.com

use in conjunction with Typhoon burner and export settings.

 

[TheLostSwede]

Why do you want to run the RAM at a higher speed than 3800MHz?
Anything above 3800MHz on a Ryzen 3000-series CPU equals high memory latency.
Instead, focus on low memory timings on 3800MHz with FCLK at 1900MHz.

If it's just for shits and giggles, see above, but read up on how to use it first.

Also, remember that XMP is an Intel feature and the programmed XMP timings may or may not work on AMD.

 

[tabascosauz]

256 views and not a single reply? come on guys, anyone here able to help me out a bit?

Pointless to venture into RAM overclocking at all without knowing what your actual ICs are. Open up Thaiphoon Burner and find out the manufacturer and name of the ICs.

• When you're in Windows, use Zentimings to check all your timings + GDM/PDM in the same place. https://github.com/irusanov/ZenTimings/releases
• If you then need to check your Clkdrv and procODT in Windows, look in Ryzen Master.
• As has already been mentioned, just go down to 3733/3800 and at least as tight as tCL 16. At those speeds, DRAM Calc can give you a head start on suggestions for subtimings, you can tweak from there.
• As soon as your FCLK can't keep up 1:1 with your memory clock, you suffer an instant performance penalty. If you had tight timings on 4600MT/s, you might have a chance at beating 3600/3733 1:1 performance, but as you've said, you can't make it work, and you're definitely not getting anywhere near breaking even at 4200 20-26-26.
• If you can't even boot and believe that your starting timings are loose enough and voltages are reasonable, mess around with procODT until you can boot. Manually set command rate to 2T or set 1T and enable Gear Down Mode.

Keep in mind that 1.5V VDIMM is, thermally, no trivial matter, even if it's "XMP". You've mentioned that you have watercooling, which rules out any passive airflow. I'm leaning towards your RAM being Samsung B-die (maaaaaybe Rev.E, but doubtful, and also kinda temp sensitive), in which case it's highly temperature sensitive anyways in regards to stability under load, so if you'll want either a small fan pointed directly at the RAM, or a top-down 120mm/140mm air cooler (which you obviously won't be doing).

Making sure the DIMMs get the passive/active airflow they deserve is a big part of playing around above 4000MT/s. Conversely, if you can't do that, coming down to 3600/3733/3800 1:1 should allow you to run at a leisurely 1.3-1.4V VDIMM, which is a good chunk cooler than 1.5V.

3733/3800 1:1 also should easily be sustainable on no more than 1.125V VSOC. I'm still not entirely convinced that 1.2V VSOC doesn't cause stability problems on its own; 1.2V SOC is commonly understood to be the safe limit on these CPUs.

Lastly, "fully stable" is 8+ runs of Memtest86 from USB stick, in addition to 200% full capacity HCI Memtest and/or 1 hour Karhu and/or 1 hour of the custom TM5 configuration by a dude whose name I can't remember. If you don't ever use your computer for anything non-trivial, maybe you'll get away with something less than stable, but the data corruption risk is yours to take.

Ryzen DRAM Calculator by 1usmus

Hi guys, I created a calculator that will make it easier to overclock the memory. It works with all kinds of memory and all zen processors. Ryzen DRAM Calculator 1.5.1 Last update : May 13th 2019 >> Download latest version >AMD Ryzen Memory Tweaking & Overclocking Guide

www.techpowerup.com

use in conjunction with Typhoon burner and export settings.

DRAM Calc only gives suggestions up to 3866MT/s (1933MHz IF being a special achievement in itself), even though the dropdown contains up to 4600MT/s. Probably for good reason, it's a Zen calculator after all.

The Thaiphoon import feature isn't quite so buggy anymore, but really doesn't give any extra useful information beyond what the built-in suggestions already give you.

 

[TheLostSwede]

The Thaiphoon import feature isn't quite so buggy anymore, but really doesn't give any extra useful information beyond what the built-in suggestions already give you.

Version 1.7.3 of the calculator can't read the RAM data on its own, so the data needs to be imported.

 

[tabascosauz]

Version 1.7.3 of the calculator can't read the RAM data on its own, so the data needs to be imported.

I mean, you are correct that it replaces the R-XMP button, but manually specifying the CPU, IC, #ranks, #sticks, and freq in the top left is 100% consistent and results in the actual suggestions provided by 1usmus (which actually change and are refined throughout the program's versions), which is really what the Calc is for. The old R-XMP button really seemed to be nothing more than an extra obstacle from getting to the recommendations.

For all the ICs that I've used Calc for, the Safe preset is really conservative, so I haven't really seen the benefit in leveraging the Thaiphoon XMP data. Something tells me that OP's kit won't be so garbage that it fails even the Safe preset, lol.

 

[TheLostSwede]

I mean, you are correct that it replaces the R-XMP button, but manually specifying the CPU, IC, #ranks, #sticks, and freq in the top left is 100% consistent and results in the actual suggestions provided by 1usmus (which actually change and are refined throughout the program's versions), which is really what the Calc is for. The old R-XMP button really seemed to be nothing more than an extra obstacle from getting to the recommendations.

For all the ICs that I've used Calc for, the Safe preset is really conservative, so I haven't really seen the benefit in leveraging the Thaiphoon XMP data. Something tells me that OP's kit won't be so garbage that it fails even the Safe preset, lol.

Right, I guess if you select the correct PCB revision (not sure how one can find that out without Thaiphoon burner either, but hey), it spits out some numbers. In my case, they're nowhere near what my RAM can do, but I guess it's an acceptable starting point.
Then again, even the imported data gives me results that are on the high side, but I guess YMMV here.

 

[phill]

Just as a thought, wouldn't the OP @JoeD be better off lowering the speeds and getting tighter timings? I kinda thought that going so far isn't really going to help out with masses of extra performance?

I would have thought aiming about the 3600Mhz - 3800MHz marker would have given much better results?

 

[tabascosauz]

Right, I guess if you select the correct PCB revision (not sure how one can find that out without Thaiphoon burner either, but hey), it spits out some numbers. In my case, they're nowhere near what my RAM can do, but I guess it's an acceptable starting point.
Then again, even the imported data gives me results that are on the high side, but I guess YMMV here.

Originally, you could tell between A0 and A1/A2 if you peer between the PCB and heatspreader, because A0 had the chips towards the top of the card. But then so many revisions came along that you had to look closely at the location of specific SMD capacitors to guess what revision it was.

Maybe it's just me. After G.skill put a 4Gb E-die sticker on my other backup Trident Z sticks while programming the DIMMs to display as 4Gb D-die, I've been highly suspicious of what Thaiphoon "reads"; it's not always gospel.

Honestly, the A3/B3/B2 vs. A0/B0 is just a Good Bin vs. Bad Bin distinction, because the newer PCBs are in theory designed for better overclocking, but IC binning quality still rules the day. Start from the best and go down the list if it doesn't work, that kinda thing. Besides, the CJR/DJR A3/B3/B2 recommendations really aren't that spectacular either. My DJR is on a B1 PCB, Calc doesn't have a category for it, and doesn't seem to affect it in either direction.

 

[JoeD]

Hi everyone, thank you guys so much for the replies!

I've included multiple screenshots / photos of my current settings and the performance test results with the intent of using these as a reference point.

Going forward, after I try to apply the tweaks / suggestions from you guys, I'll retest and post the results to see if there is an increase of measured performance concerning "lower latency" versus "higher Mhz Ram."

Ryzen DRAM Calculator by 1usmus

Hi guys, I created a calculator that will make it easier to overclock the memory. It works with all kinds of memory and all zen processors. Ryzen DRAM Calculator 1.5.1 Last update : May 13th 2019 >> Download latest version >AMD Ryzen Memory Tweaking & Overclocking Guide

www.techpowerup.com

use in conjunction with Typhoon burner and export settings.

Hiya Caring1, I've included the Typhoon burner result / screenshot and I'm noticing that Hardware Info, Aida64, and Typhoon,.. ALL report a slightly different Part Number than what is physically written on the Ram Packaging, but also it differs from the actual stickers affixed to each stick of ram.

As we can see from the actual packaging and stickers above, they differ from the following below:

Why do you want to run the RAM at a higher speed than 3800MHz?
Anything above 3800MHz on a Ryzen 3000-series CPU equals high memory latency.
Instead, focus on low memory timings on 3800MHz with FCLK at 1900MHz.

If it's just for shits and giggles, see above, but read up on how to use it first.

Also, remember that XMP is an Intel feature and the programmed XMP timings may or may not work on AMD.

Hiya Swede!

I'm not exactly an overclocking master

The reason I was thinking faster Mhz on the RAM is that it would allow me to somehow brute force LOWER FRAME TIMES running VR applications (specifically anything to allow constant under 25ms frame times)

I'm trying to learn as much as possible to get the best performance outta the hardware I have for near-real-time simulations as well while having good stable memory for 3D rendering in apps like Blender, etc.

Pointless to venture into RAM overclocking at all without knowing what your actual ICs are. Open up Thaiphoon Burner and find out the manufacturer and name of the ICs.

• When you're in Windows, use Zentimings to check all your timings + GDM/PDM in the same place. https://github.com/irusanov/ZenTimings/releases
• If you then need to check your Clkdrv and procODT in Windows, look in Ryzen Master.
• As has already been mentioned, just go down to 3733/3800 and at least as tight as tCL 16. At those speeds, DRAM Calc can give you a head start on suggestions for subtimings, you can tweak from there.
• As soon as your FCLK can't keep up 1:1 with your memory clock, you suffer an instant performance penalty. If you had tight timings on 4600MT/s, you might have a chance at beating 3600/3733 1:1 performance, but as you've said, you can't make it work, and you're definitely not getting anywhere near breaking even at 4200 20-26-26.
• If you can't even boot and believe that your starting timings are loose enough and voltages are reasonable, mess around with procODT until you can boot. Manually set command rate to 2T or set 1T and enable Gear Down Mode.

Keep in mind that 1.5V VDIMM is, thermally, no trivial matter, even if it's "XMP". You've mentioned that you have watercooling, which rules out any passive airflow. I'm leaning towards your RAM being Samsung B-die (maaaaaybe Rev.E, but doubtful, and also kinda temp sensitive), in which case it's highly temperature sensitive anyways in regards to stability under load, so if you'll want either a small fan pointed directly at the RAM, or a top-down 120mm/140mm air cooler (which you obviously won't be doing).

Making sure the DIMMs get the passive/active airflow they deserve is a big part of playing around above 4000MT/s. Conversely, if you can't do that, coming down to 3600/3733/3800 1:1 should allow you to run at a leisurely 1.3-1.4V VDIMM, which is a good chunk cooler than 1.5V.

3733/3800 1:1 also should easily be sustainable on no more than 1.125V VSOC. I'm still not entirely convinced that 1.2V VSOC doesn't cause stability problems on its own; 1.2V SOC is commonly understood to be the safe limit on these CPUs.

Lastly, "fully stable" is 8+ runs of Memtest86 from USB stick, in addition to 200% full capacity HCI Memtest and/or 1 hour Karhu and/or 1 hour of the custom TM5 configuration by a dude whose name I can't remember. If you don't ever use your computer for anything non-trivial, maybe you'll get away with something less than stable, but the data corruption risk is yours to take.



DRAM Calc only gives suggestions up to 3866MT/s (1933MHz IF being a special achievement in itself), even though the dropdown contains up to 4600MT/s. Probably for good reason, it's a Zen calculator after all.

The Thaiphoon import feature isn't quite so buggy anymore, but really doesn't give any extra useful information beyond what the built-in suggestions already give you.

Hiya tabascosauz, !!

Thank you so much for the extremely detailed and informative reply.

With my current setup, here is the Specifics as pulled from Thaiphoon (but as mentioned above in reply to Caring1, it's strange how the part number is slightly different from what is listed on the official packaging / stickers on each ram stick)


Manufacturing Description

​

Module Manufacturer:	Kingston
Module Part Number:	KHX4600C19D4/8GX
Module Series:	Undefined HyperX Series
DRAM Manufacturer:	Samsung
DRAM Components:	K4A8G085WB-BCRC
DRAM Die Revision / Process Node:	B / 20 nm
Module Manufacturing Date:	Week 02, 2020
Manufacturing Date Decoded:	January 6-10, 2020
Module Manufacturing Location:	Keelung, Taiwan
Module Serial Number:	D15AD18Dh
Manufacturing Identification Number:	0000008664422
Module PCB Revision:	00h

Physical & Logical Attributes

​

Fundamental Memory Class:	DDR4 SDRAM
Module Speed Grade:	DDR4-2400T downbin
Base Module Type:	UDIMM (133.35 mm)
Module Capacity:	8 GB
Reference Raw Card:	A2 (8 layers)
JEDEC Raw Card Designer:	SK hynix
Module Nominal Height:	31 < H <= 32 mm
Module Thickness Maximum, Front:	1 < T <= 2 mm
Module Thickness Maximum, Back:	T <= 1 mm
Number of DIMM Ranks:	1
Address Mapping from Edge Connector to DRAM:	Standard
DRAM Device Package:	Standard Monolithic
DRAM Device Package Type:	78-ball FBGA
DRAM Device Die Count:	Single die
Signal Loading:	Not specified
Number of Column Addresses:	10 bits
Number of Row Addresses:	16 bits
Number of Bank Addresses:	2 bits (4 banks)
Bank Group Addressing:	2 bits (4 groups)
DRAM Device Width:	8 bits
Programmed DRAM Density:	8 Gb
Calculated DRAM Density:	8 Gb
Number of DRAM components:	8
DRAM Page Size:	1 KB
Primary Memory Bus Width:	64 bits
Memory Bus Width Extension:	0 bits
DRAM Post Package Repair:	Supported
Soft Post Package Repair:	Supported

DRAM Timing Parameters

​

Fine Timebase:	0.001 ns
Medium Timebase:	0.125 ns
CAS Latencies Supported:	10T, 11T, 12T, 13T, 14T, 15T, 16T, 17T, 18T
Minimum Clock Cycle Time (tCK min):	0.833 ns (1200.48 MHz)
Maximum Clock Cycle Time (tCK max):	1.600 ns (625.00 MHz)
CAS# Latency Time (tAA min):	13.750 ns
RAS# to CAS# Delay Time (tRCD min):	13.750 ns
Row Precharge Delay Time (tRP min):	13.750 ns
Active to Precharge Delay Time (tRAS min):	32.000 ns
Act to Act/Refresh Delay Time (tRC min):	45.750 ns
Normal Refresh Recovery Delay Time (tRFC1 min):	350.000 ns
2x mode Refresh Recovery Delay Time (tRFC2 min):	260.000 ns
4x mode Refresh Recovery Delay Time (tRFC4 min):	160.000 ns
Short Row Active to Row Active Delay (tRRD_S min):	3.300 ns
Long Row Active to Row Active Delay (tRRD_L min):	4.900 ns
Write Recovery Time (tWR min):	15.000 ns
Short Write to Read Command Delay (tWTR_S min):	2.500 ns
Long Write to Read Command Delay (tWTR_L min):	7.500 ns
Long CAS to CAS Delay Time (tCCD_L min):	5.000 ns
Four Active Windows Delay (tFAW min):	21.000 ns
Maximum Active Window (tMAW):	8192*tREFI
Maximum Activate Count (MAC):	Unlimited MAC
DRAM VDD 1.20 V operable/endurant:	Yes/Yes

Thermal Parameters

​

Module Thermal Sensor:

Not Incorporated

SPD Protocol

​

SPD Revision:	1.1
SPD Bytes Total:	512
SPD Bytes Used:	384
SPD Checksum (Bytes 00h-7Dh):	BD97h (OK)
SPD Checksum (Bytes 80h-FDh):	53D8h (OK)

Part number details

​

JEDEC DIMM Label:

8GB 1Rx8 PC4-2400T-UA2-11

​

Frequency	CAS	RCD	RP	RAS	RC	RRDS	RRDL	WR	WTRS	WTRL	FAW
1200 MHz	18	17	17	39	55	4	6	18	3	9	26
1200 MHz	17	17	17	39	55	4	6	18	3	9	26
1067 MHz	16	15	15	35	49	4	6	16	3	8	23
1067 MHz	15	15	15	35	49	4	6	16	3	8	23
933 MHz	14	13	13	30	43	4	5	14	3	7	20
933 MHz	13	13	13	30	43	4	5	14	3	7	20
800 MHz	12	11	11	26	37	3	4	12	2	6	17
800 MHz	11	11	11	26	37	3	4	12	2	6	17
667 MHz	10	10	10	22	31	3	4	10	2	5	14

Intel Extreme Memory Profiles

​

XMP Parameter	Profile 1	Profile 2
Profiles Revision: 2.0
Profile 1 (Certified) Enables: Yes
Profile 2 (Extreme) Enables: Yes
Profile 1 Channel Config: 2 DIMM/channel
Profile 2 Channel Config: 2 DIMM/channel
Speed Grade:	DDR4-4608	DDR4-4000
DRAM Clock Frequency:	2304 MHz	2000 MHz
Module VDD Voltage Level:	1.50 V	1.35 V
Minimum DRAM Cycle Time (tCK):	0.434 ns	0.500 ns
CAS Latencies Supported:	21T,20T,19T,18T, 17T,16T,15T,14T, 13T,12T,11T,10T	21T,20T,19T,18T, 17T,16T,15T,14T, 13T,12T,11T,10T
CAS Latency Time (tAA):	19T	19T
RAS# to CAS# Delay Time (tRCD):	26T	21T
Row Precharge Delay Time (tRP):	26T	21T
Active to Precharge Delay Time (tRAS):	45T	42T
Active to Active/Refresh Delay Time (tRC):	67T	94T
Four Activate Window Delay Time (tFAW):	49T	42T
Short Activate to Activate Delay Time (tRRD_S):	12T	9T
Long Activate to Activate Delay Time (tRRD_L):	13T	10T
Normal Refresh Recovery Delay Time (tRFC1):	807T	700T
2x mode Refresh Recovery Delay Time (tRFC2):	600T	520T
4x mode Refresh Recovery Delay Time (tRFC4):	369T	320T
Show delays in nanoseconds

Additionally, I'm actually running my system in the following picture (Fans and all directly at the ram [which never get hot it seems]) :

As far as the current performance, here's my Aida64 General Memory Tests results with the ram at 4200Mhz:

READ Speeds

WRITE Speeds

COPY Speeds

LATENCY Speeds

Just as a thought, wouldn't the OP @JoeD be better off lowering the speeds and getting tighter timings? I kinda thought that going so far isn't really going to help out with masses of extra performance?

I would have thought aiming about the 3600Mhz - 3800MHz marker would have given much better results?

Hiya Phill!,

Again, thanks for the insight that appears to be the consensus on Ryzen /Zen 2 and trying to optimize for the Infinity Fabric to Memory ratio 1:1 issue or the whole 1800mhz versus 1900mhz

My goal (which I must admit, appears to have been going about it the wrong way so far) was to simply achieve the performance numbers that the RAM was stated to be able to reach per the manufacturer.

Here is the actual spec sheet from HyperX:

Page 2:

And to all,

If I can actually get BETTER performance by placing emphasis on tweaking for lower latency (AND) if this lower latency will reduce my VR Frame-Times then, most definitely, looking forward to lowering the ram Mhz while getting the memory timings tighter.

I'll let you guys know after I configure the settings more.

Thanks again!!

-JoeD

 

[tabascosauz]

@JoeD okay, so it's B-die. Plug in Zen 2 - B-die - A3/A2/B2 PCB - single rank - 3600 - 2 sticks into DRAM Calc and try booting using the Safe profile. Keep in mind the recommended Vdimm and VSoC ranges, and be ready to change procODT to whatever the sticks seem to like. Start at 3600 CAS 16 ~1.35V and work your way up in frequency and down in timings from there.

Like I said, if you want even a fighting chance at overcoming the latency hit with pure frequency, you need to be much faster/tighter than you can currently achieve. Even then, 3800 16-16-16 or tighter will knock almost everything else clean out of the park, which should be more than 13ns faster than your 4200 CAS 20.

Given that your DIMMs appear to be completely lack any kind of temperature sensor , just forget about your 1.5V experiments.

The only reason they decoupled FCLK from MCLK on Matisse was to allow the illusion that pro overclockers can, just like on Intel platforms, shoot for 5000MT/s+ records. That they can and have done, but actual performance-wise you almost need to get to 5000MT/s to be competitive without 1:1. Paradoxically, many of those high overclocks:
- aren't at safe 24/7 voltage
- require substantial airflow
- tax the CPU memory controller *heavily* (even at just 4400MT/s)
- simply can't be made stable outside of benchmark validation

You can verify, using DRAM Calc's built in Membench, that even with reasonable 1:1 timings and freq, synthetic results like AIDA don't always tell the whole story. Some timings values should run faster on paper, but actual impose performance penalties in practice. Same story with 4000+ on Ryzen.

At this point, I'm sure you can see why it's best for the health of your CPU, RAM, board, and yourself to stay at 1:1.

 

[oxrufiioxo]

Here's one of the few Ryzen Memory scaling articles That shows ram speed above 4000...... Also shows how much different 3800 CL16 vs CL18 can be.

Ryzen 3000 Memory Benchmark & Best RAM for Ryzen (fClock, uClock, & mClock) | GamersNexus

stub Memory speed on Ryzen has always been a hot subject, with AMD’s 1000 and 2000 series CPUs responding favorably to fast memory while at the same time having difficulty getting past 3200MHz in Gen1. The new Ryzen 3000 chips officially support memory speeds up to 3200MHz and can reliably run...

www.gamersnexus.net

Although even when 4200 CL16 does well it just matches 3800 CL16 so as others have said there really isn't any point.

 

[ERazer]

that latency

 

[JoeD]

@JoeD okay, so it's B-die. Plug in Zen 2 - B-die - A3/A2/B2 PCB - single rank - 3600 - 2 sticks into DRAM Calc and try booting using the Safe profile. Keep in mind the recommended Vdimm and VSoC ranges, and be ready to change procODT to whatever the sticks seem to like. Start at 3600 CAS 16 ~1.35V and work your way up in frequency and down in timings from there.

Like I said, if you want even a fighting chance at overcoming the latency hit with pure frequency, you need to be much faster/tighter than you can currently achieve. Even then, 3800 16-16-16 or tighter will knock almost everything else clean out of the park, which should be more than 13ns faster than your 4200 CAS 20.

Given that your DIMMs appear to be completely lack any kind of temperature sensor , just forget about your 1.5V experiments.

The only reason they decoupled FCLK from MCLK on Matisse was to allow the illusion that pro overclockers can, just like on Intel platforms, shoot for 5000MT/s+ records. That they can and have done, but actual performance-wise you almost need to get to 5000MT/s to be competitive without 1:1. Paradoxically, many of those high overclocks:
- aren't at safe 24/7 voltage
- require substantial airflow
- tax the CPU memory controller *heavily* (even at just 4400MT/s)
- simply can't be made stable outside of benchmark validation

You can verify, using DRAM Calc's built in Membench, that even with reasonable 1:1 timings and freq, synthetic results like AIDA don't always tell the whole story. Some timings values should run faster on paper, but actual impose performance penalties in practice. Same story with 4000+ on Ryzen.

At this point, I'm sure you can see why it's best for the health of your CPU, RAM, board, and yourself to stay at 1:1.

@tabascosauz

Hiya! Thanks again for all the detailed and extremely informative support!

I've gone ahead and did some preliminary testing which got me down to 3600Mhz (but) with a CL of 14!

The Ram is rated for 1.5V at the XMP 4600Mhz CL19 profile, but I'm only running it at 1.38 as shown below:

Here are the New Timings for my 4 stick kit of HX446C19PB3K2/16 (or KHX4600C19D4/8GX as it's being reported as in all software..)

Aida64 Cache & Memory Performance Benchmark

Previous Read = 56527 MB/s
Previous Write = 56508 MB/s
Previous Copy = 59711 MB/s
Previous Latency = 81.2ns


CineBench Benchmark

3D Mark (TimeSpy EXTREME)

Previous score with 4200Mhz @ CL20 was 6364




VR Mark (Blue Room)

Previous score with 4200Mhz @ CL20 was 4706

Here's one of the few Ryzen Memory scaling articles That shows ram speed above 4000...... Also shows how much different 3800 CL16 vs CL18 can be.

Ryzen 3000 Memory Benchmark & Best RAM for Ryzen (fClock, uClock, & mClock) | GamersNexus

stub Memory speed on Ryzen has always been a hot subject, with AMD’s 1000 and 2000 series CPUs responding favorably to fast memory while at the same time having difficulty getting past 3200MHz in Gen1. The new Ryzen 3000 chips officially support memory speeds up to 3200MHz and can reliably run...

www.gamersnexus.net

Although even when 4200 CL16 does well it just matches 3800 CL16 so as others have said there really isn't any point.

@oxrufiioxo Thank you! That was also a great read to build more knowledge on the subject

that latency

@ERazer And now?


Do you guys feel that it is possible to keep the same (or lower) latency with an increase from 3600 to 3800 with Infinity Fabric at 1900?

 

[tabascosauz]

@JoeD nice! That's a big step up (ironic, isn't it).

67.3ns is a little bit on the high side for just 2 single rank sticks. I'd give you about 1ns leeway (3900X and 3950X have slightly higher latency for 2x the read performance compared to 3600/3700X/3800X), which still leaves you about where I am (66.4-66.9ns)...except I run DJR with inferior tRCD and tRP at 3733 16-19-19.

Remember that higher freq also lowers latency. 3600 vs. 3733, 3733 vs. 3800 while keeping the exact same timings will yield somewhere around ~0.5ns lower latency for 3733.

1.38V is good, as long as its properly stable (again, 8x memtest86 + 30min P95 Large FFT + 200% HCI with all available memory capacity or 1hr Karhu). I'm also running 1.38V for my 3733/16, easily coollable with
some case airflow, and mine are dual rank.
See if you can get RRDS/RRDL/FAW to 4-6-16, WTRS/WTRL/WR to 4-12-12 without much increase in voltage.

tRFC should be significantly lower. This is B-die after all. Refer to Reous' chart, just keep in mind that the suggested ranges might be a little...optimistic. Still, tRFC is IMO a simple boot or no boot affair, no harm in exploring.



Disabling Gear Down Mode is tricky; so far I've had a little bit of success with changing Clkdrv settings to make it stable enough to at least do Membench. It costs me 0.5ns AIDA latency and ~2sec Membench but the added stability of having it On is absolute. YMMV and your call to make

As long as you can get the important subtimings to a healthy spot and 3600 CL14 is stable, I'd say you're good to stay put and enjoy, or shoot for 3733/3800 CL14, at which point you will probably have reached (more or less) the end of the road for RAM performance on your current platform.

 

[oxrufiioxo]

I would shoot for this actually..... Good luck though they used a 2x8 version of my 4000 CL15 kit but your 4600 kit might have a chance if your cpu can do 1900 on the IF

 

[JoeD]

@JoeD nice! That's a big step up (ironic, isn't it).

67.3ns is a little bit on the high side for just 2 single rank sticks. I'd give you about 1ns leeway (3900X and 3950X have slightly higher latency for 2x the read performance compared to 3600/3700X/3800X), which still leaves you about where I am (66.4-66.9ns)...except I run DJR with inferior tRCD and tRP at 3733 16-19-19.

Remember that higher freq also lowers latency. 3600 vs. 3733, 3733 vs. 3800 while keeping the exact same timings will yield somewhere around ~0.5ns lower latency for 3733.

1.38V is good, as long as its properly stable (again, 8x memtest86 + 30min P95 Large FFT + 200% HCI with all available memory capacity or 1hr Karhu). I'm also running 1.38V for my 3733/16, easily coollable with
some case airflow, and mine are dual rank.
See if you can get RRDS/RRDL/FAW to 4-6-16, WTRS/WTRL/WR to 4-12-12 without much increase in voltage.

tRFC should be significantly lower. This is B-die after all. Refer to Reous' chart, just keep in mind that the suggested ranges might be a little...optimistic. Still, tRFC is IMO a simple boot or no boot affair, no harm in exploring.

View attachment 163150

Disabling Gear Down Mode is tricky; so far I've had a little bit of success with changing Clkdrv settings to make it stable enough to at least do Membench. It costs me 0.5ns AIDA latency and ~2sec Membench but the added stability of having it On is absolute. YMMV and your call to make

As long as you can get the important subtimings to a healthy spot and 3600 CL14 is stable, I'd say you're good to stay put and enjoy, or shoot for 3733/3800 CL14, at which point you will probably have reached (more or less) the end of the road for RAM performance on your current platform.

@tabascosauz @oxrufiioxo

Hiya! Apologizes for the delayed reply guys!

I've been doing quite a bit of testing and the results are ODD (if) latency is truly "King"...

Let me explain

I've actually been able to set and hold 1900Mhz Infinity Fabric without issue with CL 16 which gives me lower latency than my current settings

However, I actually get better performance (lower frame times in milliseconds) (higher Frames Per Second) with a HIGHER latency setup with my current 1800 mhz Infinity Fabric settings + Cl16 and a reduction in tRFC down to 285 which was previously 304.

Perhaps someone can explain why this is occurring to me...


Here are my current Timings for my (4) Sticks of Ram :

Aida64 Numbers:

Time Spy:

Time Spy Extreme:

CineBench R20

Will be doing more testing with CL14 with IF 1900mhz, but as always any suggestions / lessons learned definitely welcome!

 

[oxrufiioxo]

its really strange sometimes and I prefer to test in a cpu limited gaming scenario vs using aida because I've noticed lower latency doesn't always equal higher FPS..... I used gears 5 at 720p to test gains personally mostly because it has one of the best in game benchmarks.


You do need to manually select 1900 it seems for the IF looking at your aida benchmark your IF is running 2:1