Memory overclocking... am I on the right track here?

[PovGRide742]

This is going to be a long post. I've already started the overclocking journey and my results (further down) and any questions I have are in bold, but first...

Details and Goals:

• Platform: 11900K (ABT enabled, MCE disabled, no overclock), ASUS STRIX Z590-E
• Memory: F4-3600C14-32GVK (2x16GB DDR4-3600 14-15-15-35 1.45V)
• Personally Set Limits:
  
  Stay under 1.5V DRAM voltage
  Stay in Gear 1​
  I know the minor penalty for going Gear 2 can easily be made up for with higher frequencies, but being that I mainly game and the gaming benchmarks out there so far show performance falling off after ~4000MHz, so I would rather the low CAS latency and stay in Gear 1​
  Retain CL14​
• Guide used:
  
  https://github.com/integralfx/MemTestHelper/blob/master/DDR4 OC Guide.md
  ​
• Stress Tests used:
  
  MemTest86 (Right from UEFI) - To make sure it'll still boot and function in Windows
  Karhu Ram Test - To test for stability​
• Benchmark used:
  
  AIDA64 Extreme
  ​

First Steps:

• Find my limit in Gear 1:
  
  XMP II turned on and manual settings of 100:133, 1:1, and 3733 but DRAM voltage left at 1.45V boots but errors almost instantly in MemTest86
  XMP II turned on and manual settings of 100:133, 1:1, 3733, and 1.5V DRAM boots and passes at least one pass of MemTest86
  XMP II turned on and manual settings of 100:133, 1:1, 3866, and 1.5V DRAM fails to boot (even with VCCIO (Memory) and VCCSA set to Auto which are already pretty high (close to if not higher than 1.5V)
  Being that the real world latency of 3866 CL15 is higher than that of 3733 CL14, I don't bother trying at CL15 and settle for 3733 CL14 as my Gear 1 limit​
• Determine baseline primary timings:
  
  Basically, I went on G.SKILL's website and found the lowest CL14 primary timings (14-14-14-34)
  Enabled XMP II, manually set 100:133, 1:1, 3733, 1.5V DRAM, 14 tRCD, 14 tRP, and 34 tRAS
  After passing one pass of MemTest86, determined 14-14-14-34 as my baseline primary timings​

Hiccup:

• XMP I and II set tWR to 'Auto' and the BIOS doesn't give a readout, so I don't have a baseline tWR

Following Steps:

• Using the guide linked above, I first set the suggested 'tight' timings for tRRDS, tRRDL, tFAW, and tWR as 4, 6, 16, and 12 respectively
  
  Passes one pass of MemTest86
  ​
• Set tRRDS, tRRDL, tFAW, and tWR to the suggested 'extreme' timings of 4, 4, 16, and 10 respectively
  
  Passes one pass of MemTest86
  ​
• Seeing that the next step is setting tRFC, which appears to be a lot of trial and error, I decide to do a 6400% coverage stress test using RAM Test:
  
  0 errors at 6415% (2:32:36, 26154mb tested)
  ​

Confusion / Questions for Next Steps:

• After each change, I run AIDA64 3-5 times (SEE ATTACHED TABLE)
  
  If you don't count my 3rd set of settings (which somehow had a freakishly low latency one run compared to the others), my 4th set of settings have the best average 'Read', 'Write', and 'Latency'.
  Changing from the suggested 'tight' timings of tRRDS, tRRDL, tFAW, and tWR to the suggested 'extreme' timings, performance drops off in the 5th set of settings (0 errors at 6415% coverage). Does this seem odd? Can going too low, even if it's stabile, cause a performance loss? Should I revert back to my 4th set of settings before diving into tRFC?​
• Dependent on the questions directly above, can tRRDS, tRRDL, tFAW, and tWR be set even lower than 4, 4, 16, and 10 respectively? And if so, should I even bother?
• The guide linked in the beginning describes how to tune tRFC
  
  How do you go about tuning tRFC2 and tRFC4?
  ​
• Not that I'm at this point in the process per the guide linked above yet, but can tRCD/tRP be set lower than the CAS latency (tCL)? (Assuming no, and probably the stupidest question I'll ask)
• After I'm done with tRFC, tRFC2, tRFC4, the tertiary timings mentioned in the guide linked above, revisiting the primary timings (or just tRAS if tRCD/tRP can't be lower than tCL), and seeing if 1T works (leaving it at 2T if it doesn't), should I start dropping the DRAM voltage until I'm no longer stabile or just leave it at 1.5V? If so, in what increments do you recommend?
• VCCIO (Memory) and VCCSA are on Auto and read high (close to if not above 1.5V). I read that 1.25V to 1.45V should be sufficient up to 4000 or 4400MHz. After my timings are finalized and confirmed stabile, should I manually set these to 1.45V and start dropping until I'm stabile? If so, in what increments and before or after dropping DRAM voltage until stabile?

Again, very long post, but I try to be as thorough as possible to answer any questions that might come up ahead of time. Thanks ahead of time to anyone who takes the time to read this and/or provides insight!

 

[PovGRide742]

Found some answers to my own questions...

Does this seem odd? Can going too low, even if it's stabile, cause a performance loss? Should I revert back to my 4th set of settings before diving into tRFC?

Apparently this is normal and tRRDL should be slightly looser than tRRDS.

Can tRRDS, tRRDL, tFAW, and tWR be set even lower than 4, 4, 16, and 10 respectively? And if so, should I even bother?

Apparently the lowest value for tRRDS supported by the memory controller is 4 (therefore tFAW can't be less than 16) and 10 is the minimum for tWR.

How do you go about tuning tRFC2 and tRFC4?

tRFC2 = tRFC/1.346
tRFC4 = tRFC2/1.625

 

[Woomack]

There are huge differences between some motherboards as I see. I had no time to dig into some settings but I can get 62GB+ read/write/copy and 52ns running XMP HyperX DDR4-4600 with Gear 2 without any adjustments. The same kit at 3600 16-19-19/Gear 1 and all subs at auto has ~48/50/47GB/s and 49ns in AIDA64.
DDR4-4400/4600 can be set on every above-average single-rank Hynix D, Samsung B, Micron E/B, and some more so I find rocket lake a platform which finally can use high-frequency memory. Problems are only with motherboards. Maybe it's a matter of early BIOS or something else.
For me, results in most tests were better at a higher memory clock. The only result which was better at DDR4-3600 was the Tomb Raider benchmark at 1080p. Something like Assassins Creed Valhalla at 2560x1440 and max details had ~2FPS better results at ~DDR4-4600 or 5000 than at ~DDR4-3600.

There is no guide that covers all platforms. There is no guide that covers all IC. There is also no guide that covers every motherboard and other variation of components. Results are different for every generation and we get a new chipset/CPU almost every year (AMD or Intel). I don't really have time to write memory guides recently and too often changes cause that every guide would be more or less a waste of time.

Another thing is that you can play with various sub-timings, but most have no strict rules and most don't affect performance much (out of synthetic memory benchmarks). Unless it's a specific application then results are +/- 1%. Some games react better but it's mostly because of significantly higher bandwidth or lower latency. Nothing that is really significant because of memory tuning.

Btw. if you are using Samsung B IC then you can go down to ~280-320 tRFC. Then tRFC2/4 can leave at auto or set something like 180/220. I've never seen any special rules about this timing. It's more like 30-50% lower for each next tRFC that still works well. This is one of the most important timings but as I said above, it doesn't matter much anyway. Still good to see that 2-3GB/s higher bandwidth because of a single setting.

 

[PovGRide742]

There are huge differences between some motherboards as I see. I had no time to dig into some settings but I can get 62GB+ read/write/copy and 52ns running XMP HyperX DDR4-4600 with Gear 2 without any adjustments. The same kit at 3600 16-19-19/Gear 1 and all subs at auto has ~48/50/47GB/s and 49ns in AIDA64.
DDR4-4400/4600 can be set on every above-average single-rank Hynix D, Samsung B, Micron E/B, and some more so I find rocket lake a platform which finally can use high-frequency memory. Problems are only with motherboards. Maybe it's a matter of early BIOS or something else.
For me, results in most tests were better at a higher memory clock. The only result which was better at DDR4-3600 was the Tomb Raider benchmark at 1080p. Something like Assassins Creed Valhalla at 2560x1440 and max details had ~2FPS better results at ~DDR4-4600 or 5000 than at ~DDR4-3600.

There is no guide that covers all platforms. There is no guide that covers all IC. There is also no guide that covers every motherboard and other variation of components. Results are different for every generation and we get a new chipset/CPU almost every year (AMD or Intel). I don't really have time to write memory guides recently and too often changes cause that every guide would be more or less a waste of time.

Another thing is that you can play with various sub-timings, but most have no strict rules and most don't affect performance much (out of synthetic memory benchmarks). Unless it's a specific application then results are +/- 1%. Some games react better but it's mostly because of significantly higher bandwidth or lower latency. Nothing that is really significant because of memory tuning.

Btw. if you are using Samsung B IC then you can go down to ~280-320 tRFC. Then tRFC2/4 can leave at auto or set something like 180/220. I've never seen any special rules about this timing. It's more like 30-50% lower for each next tRFC that still works well. This is one of the most important timings but as I said above, it doesn't matter much anyway. Still good to see that 2-3GB/s higher bandwidth because of a single setting.

Very well said.

After benching games and only noticing 1-2 difference in FPS after hours and hours of timing adjustments I decided just to set XMP and call it a day. Maybe I'll try again after a few more BIOS releases but looking at your review of the same kit (but in NEO) where you only experienced a few FPS as well, it's just not worth my time.

Thanks for taking time to reply!

 

[Woomack]

If you want to play some more then try a higher memory clock. Your RAM should handle 4800+ stable at something like CL18-18-18 or 18-20-20. I'm testing some motherboards now but will focus more on RAM in a couple of days. So far this is what I could make on relaxed timings on ASRock Z590 Extreme. Not much playing with timings because of the limited time but it's scaling well up to 5200.

 

[PovGRide742]

If you want to play some more then try a higher memory clock. Your RAM should handle 4800+ stable at something like CL18-18-18 or 18-20-20. I'm testing some motherboards now but will focus more on RAM in a couple of days. So far this is what I could make on relaxed timings on ASRock Z590 Extreme. Not much playing with timings because of the limited time but it's scaling well up to 5200.

View attachment 213512

Nice uplift! I think I'm going to wait for UEFI revisions before I start digging more into RAM overclocking but when I do, out will come the Excel again and see how far I can take the frequencies while not having to go too far up in CL, using the general rule of thumb that 200MHz to make up for every 1CL increase..