Which is better? tCL 28 with 25K tREFI vs tCL 30 with 50K tREFI?

[rcarlos243]

I have two sets of 64GB (2x32GB) Dual Rank Hynix A-Die on X670E-E with 7800X3D.

 

kit #1 can do tCL 28 with 25K tREFI and kit #2 can do tCL 30 with 50K tREFI.

 

The speed and all other sub-timings are identical, which one is better?

[RONOTHAN##]

Assuming the same frequency, tREFI at 50k is significantly better. tCL doesn't affect performance all that much, while tREFI does. 

[rcarlos243]

21 minutes ago, RONOTHAN## said:

Assuming the same frequency, tREFI at 50k is significantly better. tCL doesn't affect performance all that much, while tREFI does. 

 

 

 

This is a Dual Rank Hynix A Die, Any subtiming worth lowering?

[RONOTHAN##]

3 minutes ago, rcarlos243 said:

 

 

 

This is a Dual Rank Hynix A Die, Any subtiming worth lowering?

tREFI and tRFC are the most worthwhile, though admittedly with the X3D chips memory performance is not that big a deal.

 

What timings are you currently running? Most of the ones from that screenshot should just work on most A die kits (with the exception of tREFI and tRFC, those are really tight). 

[rcarlos243]

10 minutes ago, RONOTHAN## said:

tREFI and tRFC are the most worthwhile, though admittedly with the X3D chips memory performance is not that big a deal.

 

What timings are you currently running? Most of the ones from that screenshot should just work on most A die kits (with the exception of tREFI and tRFC, those are really tight). 

 

This is the kit #2. CL30 and 50K tREFI was the stock 1.4v voltage.

 

I increased the DRAM voltage to 1.45v and is currently stress testing the max tREFI with TM5 anta777 extreme (it is currently running for about 1hr 30min). Will also test with Y-Cruncher VST, N63 and VT3 for 8 hours.

 

I tried a lower tRFC (383) with stock 1.4v, but the computer did not POST.

If the max tREFI is stable at 1.45v, I will revisit tRFC again to see if it can go lower with 1.45v

[RONOTHAN##]

3 minutes ago, rcarlos243 said:

 

This is the kit #2. CL30 and 50K tREFI was the stock 1.4v voltage.

 

I increased the DRAM voltage to 1.45v and is currently stress testing the max tREFI with TM5 anta777 extreme (it is currently running for about 1hr 30min).

 

I tried a lower tRFC (383) with stock 1.4v, but the computer did not POST.

If the max tREFI is stable at 1.45v, I will revisit tRFC again to see if lower it at 1.45v

tREFI is not usually a timing that scales with voltage, in fact in my experience it's more likely to do the opposite as it more scales with temps, and higher voltage means higher temps. tRFC does scale with both, though going from 400 to 383 is IMO a waste of time if the tREFI timing is maxed out since there's at most a 0.026% performance uplift to be had. If it does 65k at 1.45V, I'd be shocked if it didn't do 65k at 1.4V.

[rcarlos243]

11 minutes ago, RONOTHAN## said:

tREFI is not usually a timing that scales with voltage, in fact in my experience it's more likely to do the opposite as it more scales with temps, and higher voltage means higher temps. tRFC does scale with both, though going from 400 to 383 is IMO a waste of time if the tREFI timing is maxed out since there's at most a 0.026% performance uplift to be had. If it does 65k at 1.45V, I'd be shocked if it didn't do 65k at 1.4V.

 

increasing the voltage to 1.45v definitely helped with tREFI to 65K.

I tried 65K tREFI with 1.4v and it BSOD my PC after running TM5 for about 20mins.

 

The other identical Dual Rank Hynix A-Die kit I have cannot go above 25K tREFI at 1.4v (haven't tried with higher voltage yet) without throwing errors in TM5.

 

Also, I currently have no active cooling for the RAM and the maximum RAM temp so far is 74c and stress test still no errors or crashes.

 

I ordered a RAM fan from AliExpress, but it looks like it won't arrive till 2nd week of January

[WereCat]

41 minutes ago, rcarlos243 said:

 

 

 

This is a Dual Rank Hynix A Die, Any subtiming worth lowering?

tRRDL should be 4

 

Others seem okay. There's not really much you can gain for daily use. 

You'll get more performance by just disabling TSME which is a memory encryption. Not very relevant if you're not using Virtual Machines as its intended to prevent memory peeping from VM to VM to host. 

[rcarlos243]

1 minute ago, WereCat said:

tRRDL should be 4

 

Others seem okay. There's not really much you can gain for daily use. 

You'll get more performance by just disabling TSME which is an memory encryption. Not very relevant if you're not using Virtual Machines as its intended to prevent memory peeping from VM to VM to host. 

 

I tried tRRDL 4 and it won't even POST.

I also tried tRRDL 6, but it throws errors when stress-tested.

I settled with tRRDL 8 and it passed 8 hours of Y-Cruncher VST, N63, and VT3, about 2hrs 30min of TM5 Anta777 extreme config and 1hr OCCT CPU stress test

 

I might revisit tRRDL if it can do better at 1.45v after I verified tREFI at 65k

[RONOTHAN##]

29 minutes ago, WereCat said:

tRRDL should be 4

 

That doesn't work on most A die unless you're slamming it with insane voltage values (I.E. 1.6V+). The bad will bottom out at 8, while the good will bottom out at 7 or 6 with very few doing 4. 

 

Not that it really matters that much though, the tRRDL timing is rarely used and because DDR5 bumped the read burst length from 4 to 8, going below 8 in general makes very little difference. 

 

 

[WereCat]

31 minutes ago, rcarlos243 said:

 

I tried tRRDL 4 and it won't even POST.

I also tried tRRDL 6, but it throws errors when stress-tested.

I settled with tRRDL 8 and it passed 8 hours of Y-Cruncher VST, N63, and VT3, about 2hrs 30min of TM5 Anta777 extreme config and 1hr OCCT CPU stress test

 

I might revisit tRRDL if it can do better at 1.45v after I verified tREFI at 65k

That's interesting. To be fair I haven't dabbled with DDR5 so I don't know how significant some changes to timings are but on DDR4 you can almost universally set tRRDS 4, tRRDL 4 (6 for dual rank sticks) and tFAW 16 and you are almost certainly stable and able to boot. It's also one of the most significant and safest way to increase memory performance. 

 

"Universally" for decent sticks I mean. 

 

Just disable TSME then. That's like 2ns to 3ns improvement if you measure with Aida64. (also basing this on DDR4 knowledge) 

[WereCat]

4 minutes ago, RONOTHAN## said:

That doesn't work on most A die unless you're slamming it with insane voltage values (I.E. 1.6V+). The bad will bottom out at 8, while the good will bottom out at 7 or 6 with very few doing 4. 

 

Not that it really matters that much though, the tRRDL timing is rarely used and because DDR5 bumped the read burst length from 4 to 8, going below 8 in general makes very little difference. 

 

 

I've seen people not doing 4,4,16 on DDR5 and usually see 4,8,20 or 4,4,20 and I wasn't sure why so thanks for info.

 

(and previous reply to OP) 

[RONOTHAN##]

Just now, WereCat said:

I've seen people not doing 4,4,16 on DDR5 and usually see 4,8,20 or 4,4,20 and I wasn't sure why so thanks for info.

 

(and previous reply to OP) 

The reason for the 20 is because AM5 has that as the lower limit for tFAW for whatever reason (Intel can still do 16), and the older Hynix 16Gb M die did do 4/4/16 more reliably, though it still had issues on occasion. 

 

When you get to the 24GB sticks it gets even weirder because they will not do 8 on tRRDL, so things like 8/12/20 is seen on the bad kits while 4/9/16 is seen on the good ones.