Overclock.net banner
1 - 20 of 35 Posts

papakilo

· Registered
Joined
·
26 Posts
Discussion starter · #1 ·
Hi all,

I just picked up a x870 Taichi, 9950x and this kit of memory T-Force Xtreem Overclocking 48GB DDR5-8200 which is on the QVL but it is not stable when running the XMP profile. Doing some googling and searching this forum, it sounds like I may have made a mistake buying this stuff and perhaps should have just gotten a good set of 6400 memory (I'm new to the OC world if you can't tell.)

So wanted to get some advice on whether or not to try and make this stuff work or return it and get something that you don't have to fight with for possibly dubious improvement.

Thanks in advance!
 
XMP profiles are essentially useless at anywhere near these speeds on AMD platforms. If you expect a mostly plug and play experience, you're going to want EXPO memory.

AMD and ASRock's newest firmware, which you absolutely want with a dual-CCD 9000 series CPU (AGESA 1.2.0.2 fixes the inter-CCX latency issue) also requires different settings for the same memory clocks which is also likely to break XMP/EXPO compatibility.

All that said, there is nothing wrong with the memory kit itself. I'm using the same stuff and can do 8200 on my much cheaper HDV + 9700X, with some work.

A cheaper 6000-6400 kit that you can drop in and run EXPO 1:1 is probably the easiest way to acceptable performance, but if you want to keep what you've got (which will be better than most cheaper kits, once tuned), I can provide some suggestions.
 
  • Rep+
Reactions: papakilo
Discussion starter · #3 ·
XMP profiles are essentially useless at anywhere near these speeds on AMD platforms. If you expect a mostly plug and play experience, you're going to want EXPO memory.

AMD and ASRock's newest firmware, which you absolutely want with a dual-CCD 9000 series CPU (AGESA 1.2.0.2 fixes the inter-CCX latency issue) also requires different settings for the same memory clocks which is also likely to break XMP/EXPO compatibility.

All that said, there is nothing wrong with the memory kit itself. I'm using the same stuff and can do 8200 on my much cheaper HDV + 9700X, with some work.

A cheaper 6000-6400 kit that you can drop in and run EXPO 1:1 is probably the easiest way to acceptable performance, but if you want to keep what you've got (which will be better than most cheaper kits, once tuned), I can provide some suggestions.
Hey thanks a bunch for the feedback! I'm definitely down to give it a shot. I already updated the bios to latest and it says it has the 1.2.0.2 fix, so should be good there. What should I do next?
 
You'll want a bootable memory test so you can make sure any settings that pass training aren't so flaky they'll corrupt your OS install. The free version of Memtest86 will work fine for this purpose; you'll need an empty usb drive to make the bootable media. Once you've created that, and before you safe and exit with any new UEFI settings, make sure that USB drive is the only boot device enabled, until you've got the settings you want to test further (in Windows) dialed in.

Settings wise (use the advanced view to make this easier), first thing is to get some baseline values that should POST/train correctly, which you can then clock up until you encounter issues.

For voltages, I'd start with: 1.3 VDDIO, 1.4 DRAM VDD, 1.4 DRAM VDDQ, 1.8 DRAM VPP, 1.15 VDD_SOC, 1.1 VDD_MISC, 0.85 on both VDDG CCD & IOD, and 1.1 VDDP. Set the DRAM Frequency to 8000MT/s, Infinity Fabric Frequency (FCLK) to 2000MHz, and UCLK DIV1 MODE to UCLK=MEMCLK/2. All of those settings should be in the main OC Tweaker page. Most of the stuff under "External voltage settings" should be fine at AUTO, though if you see any options to adjust VRM PWM frequencies, maxing these out, as long as you don't exceed 700kHz, may be prudent.

Under "DRAM Timings Configuration:, try matching the timings (not the drive strengths/termination resistances, nor clocks I have already specified above) that I'm using in this post, with the following exceptions: tRCD can be 48 (at 8000MT/s), tRFC2 376, tRFCsb 296, tRDWR 16, and tWRRD 2 (note I don't believe that tRFC2 and tRFCsb aren't actually used, yet, but I fill them in anyway) . GDM will almost certainly need to be enabled for 8000MT/s+ without more voltage. There is a submenu ("Data Bus Configuration") with the output driver impedances and terminations resistances in it. You can try leaving these AUTO for now, but if that doesn't work, manually setting RTT_NOM to Disabled, all the other DRAM ODT impedance values to 40ohm (/6), DRAM DQ drive strength to 48ohm (the highest value), ProcODT to 28.2ohm, and Processor DQ drive strength to 34ohm, and Processors CA drive strength to 30ohm.

Next place for memory timings is under the "Advanced\AMD Overclocking\DDR and Infiity Fabric" menu. You're going to need to enable DDR5 Nitro Mode and want to set RX Data to 1, TX Data to 3, and Control Line to 1. You can probably get away with disabling robust training and keeping the RX/TX Bust Lengths both at 1x. There is another submenu here called "DDR PMU Training", set the values there like this, except for ARdPtrInitVal, which should be "2".

Save and exist with those settings, with the boot device as the MemTest86 USB drive. It will likely take a significant amount of time to train these settings initially, then another few minutes for MemTest86 to start up. You can manually set it to run two passes of tests 6, 7, 8, and 9. If it passes, boot into Windows, grab a ZenTimings screen shot and post it here, then start using Windows tests (TM5, y-cruncher FFTv3/4, and VT3, etc) to look for instabilities.
 
I suggest never volunteer as a beta tester.
 
I suggest never volunteer as a beta tester.
And to think that everyone expects doing those speeds easy peasy..even with the guidance above, I'd still say its still silicon lottery..and after all that work on getting there is the performance satisfactory? I have an 8600G doing 8k+ on memory but the performance is crap (personally)..
 
@papakilo an example of the 8100MT/s settings I'm running with this kit on my HDV:

I wouldn't recommend copying those settings exactly, as our boards likely have different drive strength and ODT requirements and you'd need to validate the higher FCLK on your CPU sample first.

And to think that everyone expects doing those speeds easy peasy..even with the guidance above, I'd still say its still silicon lottery..and after all that work on getting there is the performance satisfactory? I have an 8600G doing 8k+ on memory but the performance is crap (personally)..
8000-8200 should not be that hard for most Granite Ridge samples on a high-end board with 8200 rated memory. Definitely took me a bit of trial and error on my $130 board, but that's as much firmware teething issues as anything. AM5 has generally been able to do 8000MT/s+ on decent boards with the 1:2 UCLK:MCLK ratio since AGESA 1.0.0.7b that added Nitro settings back in July 2023.

It is certainly true that the performance gains from 1:2 are often marginal, but a dual CCX Raphael or Granite Ridge part can often meaningfully benefit as the FCLK bottleneck is less of an issue. With my single CCX parts I'm on the fence between super tight 6200-6400 1:1 settings or 8100 1:2, but I'd probably choose the latter on a 9950X.
 
  • Love
Reactions: papakilo
Discussion starter · #9 ·
You'll want a bootable memory test so you can make sure any settings that pass training aren't so flaky they'll corrupt your OS install. The free version of Memtest86 will work fine for this purpose; you'll need an empty usb drive to make the bootable media. Once you've created that, and before you safe and exit with any new UEFI settings, make sure that USB drive is the only boot device enabled, until you've got the settings you want to test further (in Windows) dialed in.

Settings wise (use the advanced view to make this easier), first thing is to get some baseline values that should POST/train correctly, which you can then clock up until you encounter issues.

For voltages, I'd start with: 1.3 VDDIO, 1.4 DRAM VDD, 1.4 DRAM VDDQ, 1.8 DRAM VPP, 1.15 VDD_SOC, 1.1 VDD_MISC, 0.85 on both VDDG CCD & IOD, and 1.1 VDDP. Set the DRAM Frequency to 8000MT/s, Infinity Fabric Frequency (FCLK) to 2000MHz, and UCLK DIV1 MODE to UCLK=MEMCLK/2. All of those settings should be in the main OC Tweaker page. Most of the stuff under "External voltage settings" should be fine at AUTO, though if you see any options to adjust VRM PWM frequencies, maxing these out, as long as you don't exceed 700kHz, may be prudent.

Under "DRAM Timings Configuration:, try matching the timings (not the drive strengths/termination resistances, nor clocks I have already specified above) that I'm using in this post, with the following exceptions: tRCD can be 48 (at 8000MT/s), tRFC2 376, tRFCsb 296, tRDWR 16, and tWRRD 2 (note I don't believe that tRFC2 and tRFCsb aren't actually used, yet, but I fill them in anyway) . GDM will almost certainly need to be enabled for 8000MT/s+ without more voltage. There is a submenu ("Data Bus Configuration") with the output driver impedances and terminations resistances in it. You can try leaving these AUTO for now, but if that doesn't work, manually setting RTT_NOM to Disabled, all the other DRAM ODT impedance values to 40ohm (/6), DRAM DQ drive strength to 48ohm (the highest value), ProcODT to 28.2ohm, and Processor DQ drive strength to 34ohm, and Processors CA drive strength to 30ohm.

Next place for memory timings is under the "Advanced\AMD Overclocking\DDR and Infiity Fabric" menu. You're going to need to enable DDR5 Nitro Mode and want to set RX Data to 1, TX Data to 3, and Control Line to 1. You can probably get away with disabling robust training and keeping the RX/TX Bust Lengths both at 1x. There is another submenu here called "DDR PMU Training", set the values there like this, except for ARdPtrInitVal, which should be "2".

Save and exist with those settings, with the boot device as the MemTest86 USB drive. It will likely take a significant amount of time to train these settings initially, then another few minutes for MemTest86 to start up. You can manually set it to run two passes of tests 6, 7, 8, and 9. If it passes, boot into Windows, grab a ZenTimings screen shot and post it here, then start using Windows tests (TM5, y-cruncher FFTv3/4, and VT3, etc) to look for instabilities.
Wow, thanks so much for the amazing write up!! I'll give it a shot today and report back.
 
Discussion starter · #10 ·
@papakilo an example of the 8100MT/s settings I'm running with this kit on my HDV:

I wouldn't recommend copying those settings exactly, as our boards likely have different drive strength and ODT requirements and you'd need to validate the higher FCLK on your CPU sample first.



8000-8200 should not be that hard for most Granite Ridge samples on a high-end board with 8200 rated memory. Definitely took me a bit of trial and error on my $130 board, but that's as much firmware teething issues as anything. AM5 has generally been able to do 8000MT/s+ on decent boards with the 1:2 UCLK:MCLK ratio since AGESA 1.0.0.7b that added Nitro settings back in July 2023.

It is certainly true that the performance gains from 1:2 are often marginal, but a dual CCX Raphael or Granite Ridge part can often meaningfully benefit as the FCLK bottleneck is less of an issue. With my single CCX parts I'm on the fence between super tight 6200-6400 1:1 settings or 8100 1:2, but I'd probably choose the latter on a 9950X.
Great info thanks again!! Yeah I keep seeing manufacturers like MSI and Gigabyte touting their new 870 boards support 8400 and 8600 speeds and I just found this review where it shows that of the four boards tested, the one I picked is the one that doesn't work just selecting a preset profile: AMD X870 Flagship Roundup: Zen 5 Motherboards Are Here
 
Hi all,

I just picked up a x870 Taichi, 9950x and this kit of memory T-Force Xtreem Overclocking 48GB DDR5-8200 which is on the QVL but it is not stable when running the XMP profile. Doing some googling and searching this forum, it sounds like I may have made a mistake buying this stuff and perhaps should have just gotten a good set of 6400 memory (I'm new to the OC world if you can't tell.)

So wanted to get some advice on whether or not to try and make this stuff work or return it and get something that you don't have to fight with for possibly dubious improvement.

Thanks in advance!
...I had ordered the T-Force Xtreem 48 GB DDR5 8200 from Amazon in the spring - at the time, that kit was advertised as 'A-die', which it is not according to folks who have it and took the heat spreader off (it is the newer 24 GB M-die, which is good stuff for 8000/+). The final leg of the delivery went horribly wrong, so Amazon sent me a refund. I saved a bunch by getting the GSkill DDR5 8400 1.4V 48GB kit instead from Newegg (spoiler)
Product Font Screenshot Technology Engineering

The T-Force Xtreem 48 GB DDR5 8200 kit is a great kit from what I have seen and read, and do believe that the chips in the GSkill 48 GB DDR5 8400 and the T-Force Xtreem 48 GB DDR5 8200 are very close if not the same. While I run a 7950X3D / FCLK 2200 on an Aorus 670E Master, the 8000/8200 Zen timings could still serve as a good overall 'general map' for you as the IO die is the same, noting that the 8200 is done via bclk...as is the 8600 (which was just for fun and certainly is not stable).

As others pointed out, there will be a lot of stability testing to do. With a new mobo / CPU / RAM combo, I typically tip-toe in at first and make a shortlist of potential RAM profiles by running AIDA64E / System Stability Test ('Stress system memory' only) for ~ 20 min or so. BTW, make sure that the DDR5 kit is very well cooled; once temps rise towards the high 40 C to mid 50 C, results and stability will be affected. My personal air-cooled DDR5 VDD voltage limit is 1.55 V, but I usually stay well below that. For 8000/8200/+, my VDDIO is always at 1.34 V or less and my VDDQ varies with VDD settings - I usually have VDDQ below VDD by ~0 .1 V

Once I have the shortlist, I run longer tests with TM5, Y-Cruncher VT3 and 50 laps of Memtest64 (from TechPowerUp). There is also some proprietary RAM test software I use for longer runs as well. On a final note, all my 16c/32t AMD setups over four gens (TR 2950X - 7950X3D) run with Intel XMP only (no Expo). This is a bit unusual but makes sense for me as I have a host of machines for work and play (including lots of Intel), so having one type of RAM that runs on both AMD and Intel without issue helps....never had a problem with XMP on AMD. So for DDR5, my kits are from the GSkill TZ5RK range (7200, 8400) and I can also mix them easily enough by adopting the '(s)lowest common parameter values'. The 4 dimm CL30 6400 / 96 GB below is an example.

Good luck in your chase down the many rabbit holes which await :p

Computer Font Screenshot Technology Software
 
With AGESA 1.2.0.2 I find that I need more VDDQ and a higher DRAM DQS output impedance than with prior firmware, as well as a much lower ProcODT.
 
  • Rep+
Reactions: papakilo
It is certainly true that the performance gains from 1:2 are often marginal, but a dual CCX Raphael or Granite Ridge part can often meaningfully benefit as the FCLK bottleneck is less of an issue. With my single CCX parts I'm on the fence between super tight 6200-6400 1:1 settings or 8100 1:2, but I'd probably choose the latter on a 9950X.
I am still yet to get a 9700x sample to try out but based on what I am reading everywhere the single CCD chips highly benefits 1:1 sync ratio, not entirely sure about the dual CCD chips, if you have tried one and was able to get 8200 stable at what fabric clock was this? 2200? I do rarely see chips doing 2200 on Fabric, yours might be binned or you simply won lottery.
 
I am still yet to get a 9700x sample to try out but based on what I am reading everywhere the single CCD chips highly benefits 1:1 sync ratio, not entirely sure about the dual CCD chips, if you have tried one and was able to get 8200 stable at what fabric clock was this? 2200? I do rarely see chips doing 2200 on Fabric, yours might be binned or you simply won lottery.
Granite Ridge CCDs seem to tolerate higher FCLKs in general and they likely have IODs binned to match. Many Granite Ridge setups will default to 2100 FCLK when using EXPO and most Granite Ridge samples can do 2200 FCLK.

I tried four 9700Xes (three could do 2200, one topped out at 2167) before being given a 2233 capable retail sample (that is still only unconditionally stable at 2233 when capped to ~130W...so I run 2200 because after I delidded it I found I could cool ~180W on air before hitting temperature limits).

Out of these five, four of them could do 8200 in an ASRock B650M-HDV/M.2, three at 2200 FCLK. One sample had a dud IMC and the other was not stable at 2200 FCLK at all.

Prior to AGESA 1.2.0.2 I was running 6200MT/s with tight timings (and Nitro 1-2-0) as that is what was performing best. However, whatever they did with AGESA 1.2.0.2 has changed that (latency improved slightly all-round, but I lost a fair bit of bandwidth in the 1:1 ratios), so I'm finding that 8100 now performs best (8200 needs to be so much looser on this board that the performance is worse than 8000, let alone 8100).

8000+ has generally had a small edge with dual-CCX parts, at least where both CCXes were heavily utilized, because two Fabric links can fully saturate 6400MT/s memory.
 
Granite Ridge CCDs seem to tolerate higher FCLKs in general and they likely have IODs binned to match. Many Granite Ridge setups will default to 2100 FCLK when using EXPO and most Granite Ridge samples can do 2200 FCLK.

I tried four 9700Xes (three could do 2200, one topped out at 2167) before being given a 2233 capable retail sample (that is still only unconditionally stable at 2233 when capped to ~130W...so I run 2200 because after I delidded it I found I could cool ~180W on air before hitting temperature limits).

Out of these five, four of them could do 8200 in an ASRock B650M-HDV/M.2, three at 2200 FCLK. One sample had a dud IMC and the other was not stable at 2200 FCLK at all.

Prior to AGESA 1.2.0.2 I was running 6200MT/s with tight timings (and Nitro 1-2-0) as that is what was performing best. However, whatever they did with AGESA 1.2.0.2 has changed that (latency improved slightly all-round, but I lost a fair bit of bandwidth in the 1:1 ratios), so I'm finding that 8100 now performs best (8200 needs to be so much looser on this board that the performance is worse than 8000, let alone 8100).

8000+ has generally had a small edge with dual-CCX parts, at least where both CCXes were heavily utilized, because two Fabric links can fully saturate 6400MT/s memory.
will see when I get my "not so good silicon" sample, for all I know, from my region all bins has been trash, only the APU's had success on those speeds.
 
Discussion starter · #16 ·
...I had ordered the T-Force Xtreem 48 GB DDR5 8200 from Amazon in the spring - at the time, that kit was advertised as 'A-die', which it is not according to folks who have it and took the heat spreader off (it is the newer 24 GB M-die, which is good stuff for 8000/+). The final leg of the delivery went horribly wrong, so Amazon sent me a refund. I saved a bunch by getting the GSkill DDR5 8400 1.4V 48GB kit instead from Newegg (spoiler)
The T-Force Xtreem 48 GB DDR5 8200 kit is a great kit from what I have seen and read, and do believe that the chips in the GSkill 48 GB DDR5 8400 and the T-Force Xtreem 48 GB DDR5 8200 are very close if not the same. While I run a 7950X3D / FCLK 2200 on an Aorus 670E Master, the 8000/8200 Zen timings could still serve as a good overall 'general map' for you as the IO die is the same, noting that the 8200 is done via bclk...as is the 8600 (which was just for fun and certainly is not stable).

As others pointed out, there will be a lot of stability testing to do. With a new mobo / CPU / RAM combo, I typically tip-toe in at first and make a shortlist of potential RAM profiles by running AIDA64E / System Stability Test ('Stress system memory' only) for ~ 20 min or so. BTW, make sure that the DDR5 kit is very well cooled; once temps rise towards the high 40 C to mid 50 C, results and stability will be affected. My personal air-cooled DDR5 VDD voltage limit is 1.55 V, but I usually stay well below that. For 8000/8200/+, my VDDIO is always at 1.34 V or less and my VDDQ varies with VDD settings - I usually have VDDQ below VDD by ~0 .1 V

Once I have the shortlist, I run longer tests with TM5, Y-Cruncher VT3 and 50 laps of Memtest64 (from TechPowerUp). There is also some proprietary RAM test software I use for longer runs as well. On a final note, all my 16c/32t AMD setups over four gens (TR 2950X - 7950X3D) run with Intel XMP only (no Expo). This is a bit unusual but makes sense for me as I have a host of machines for work and play (including lots of Intel), so having one type of RAM that runs on both AMD and Intel without issue helps....never had a problem with XMP on AMD. So for DDR5, my kits are from the GSkill TZ5RK range (7200, 8400) and I can also mix them easily enough by adopting the '(s)lowest common parameter values'. The 4 dimm CL30 6400 / 96 GB below is an example.

Good luck in your chase down the many rabbit holes which await :p

View attachment 2675878
Thanks a ton for the great info!! I just saw Der Bauer's latest video on Nitropath (turns out to not be just a gimmick) and so am going to switch over to the Hero. I'm wondering if I should switch to that 8400 kit you have as well just to remove one more unknown.
 
  • Rep+
Reactions: J7SC
Thanks a ton for the great info!! I just saw Der Bauer's latest video on Nitropath (turns out to not be just a gimmick) and so am going to switch over to the Hero. I'm wondering if I should switch to that 8400 kit you have as well just to remove one more unknown.
Personally, I think those two kits are close as to their use of top tier SK Hynix ICs - I am just more comfortable with GSkill for other reasons (including running 96 GB by joining pre-existing GSkill kits). So I do not think it is necessary to switch kits, but if you have the option to do it w/o dollar penalty, why not.
 
Personally, I think those two kits are close as to their use of top tier SK Hynix ICs - I am just more comfortable with GSkill for other reasons (including running 96 GB by joining pre-existing GSkill kits). So I do not think it is necessary to switch kits, but if you have the option to do it w/o dollar penalty, why not.
Question: When it's all said and done what CPU will you be using?
 
Question: When it's all said and done what CPU will you be using?
I'm more than happy with my 7950X3D (great oc, IMC and fclk) for now but certainly will take a long hard look at the upcoming 9950X3D. That said, my next bigger 'full' upgrade might actually be a Zen5 based Threadripper, given the need for more PCIe 5(4) x16 slots for multiple GPUs on work projects.
 
I'm more than happy with my 7950X3D (great oc, IMC and fclk) for now but certainly will take a long hard look at the upcoming 9950X3D. That said, my next bigger 'full' upgrade might actually be a Zen5 based Threadripper, given the need for more PCIe 5(4) x16 slots for multiple GPUs on work projects.
I jumped the gun om 7950 as the 7950X3D wasn't in the offering at that time. I'm setting here now with a new X870E Tiachi and some 8200 memory and no place to go. I suppose that when my back surgery heals up and I get out of this brace and can lift, I can get get every thing up and going with the 7950 until the X3D becomes available down the line. You're thinking Threadripper. probably work related, if so, which one might that be?
 
1 - 20 of 35 Posts