Build Advice Pete J's massive system overhaul discussion

Whelp, I've decided to follow @GeorgeStorm 's advice and stick to 64GB. Trying to achieve stability at 4800MHz was basically impossible, let alone any higher. In fact, the Corsair RAM (that I've returned today) was better performing in this regard, being able to hold its latency values from the 6600MHz profile.

I'm currently stress testing one set of RAM with Prime95. Unfortunately, its 6400MHz CL32 profile had one core fail after 50 minutes (even when trying the alternative EXPO profile, which actually failed pretty much immediately); the 6000MHz CL30 profile has lasted an hour. Going to swap out the other set and see how that performs.

 

The other set of RAM passed 2 hours or Prime95 testing. So, on to the next thing...

Whilst waiting for the stability test to do some decent time, I read a bit more of the Techpowerup article about RAM speed and latency performance I linked earlier. In some cases, it was better to have lower latencies than higher frequencies. Curious, I ran Cinebench (R23.2) a few times with the two profiles. Here are the results:

• 6400MHz 32-39-39-84
  • 23513 mean
• 6000MHz 30-36-36-77
  • 23565 mean
• Common factors
  • Standard deviation (sample) of 11 for both.
  • At least 5 tests done.
  • Test done until result stopped increasing.
  • Highest three results used.

Since the Cinebench performance correlates with gaming performance (admittedly making a fraction of a percent difference), I'm going to use the 6000 MHz profile. Also, I must update the OP with details later.

 

Nothing really's going to happen until the 5090 turns up. However, in the meantime, I'm fiddling with the AMD Curve Optimizer. So far I have it down to -10 and have noticed that clock speeds run a little better under extreme test conditions.

 

I found that way more fascinating than I ought to.

 

Still waiting for the 5090 to turn up, although starting to feel a tiny bit of buyer's remorse, what with all the bad news articles coming out: melting power connectors, missing ROPs, dropping 32 bit PhysX support etc.

Anyway, this weekend, I was intending to do long duration stress tests on the CPU. This has got off to a bad start as settings that were relatively stable a fortnight ago suddenly aren't . On top of that, even though I didn't have the wifi aerial plugged in, previously, the PC could still pick up wifi. Now I've had to plug the aerial in to make it work.

So, I reset Curve Optimiser to do a quick stability test (seems fine), installed all Windows updates, and am once again twiddling. I've done a test with all cores at -20 curve offset and that seems fine, so will work from there. Again. Next is -25, then I'll start tweaking the cores individually.

Just a bit puzzled as to how the system went from stable to unstable. A decade ago, when I was doing overclocking, once I had values reasonably dialed in, that was that.

 

No change on the GPU front, but am only 4 weeks in to the vaguely promised "2-12" week delivery estimate. Oh well, plenty of time for the super early adopters to find solutions to the ugly issues that have plagued the 5000 series launch. Although this GPU is intended to last a decade for me, it may be the case that I trade it for the 6000 series in...3 years time?

Anyway, I'm still playing around with the AMD Curve Optimiser, and I think I'm going to give up on individual core tweaking: it's proving to be a pain in the bum and the benefit over a CPU wide values is arguably insignificant. Funnily enough, one of the stability issues I encountered was simply leaving the PC on at the login screen, which crashed within 90 minutes or so. I put this down to the Curve Optimiser also affecting things at the bottom end of the CPU power state, Here's my overclocking testing methodology:

Stage 1: CPU Baseline Stability:

1. Increase negative offset.
2. Run CineBench R23 benchmark.
   • If benchmark passes, go back to 1.
   • If benchmark fails, go to 3.
3. Decrease negative offset.
4. Run small scale stability testing subroutine.
   • Small scale stability testing subroutine.
     1. Run Cinebench R23 benchmark.
     2. Run CPU-Z stress test for ≥10 minutes.
     3. Run Prime95 mixed blend stress test
        • ≥10 minutes required.
        • ≥25 successful passes per core required.
     4. Run CPU-Z stress test for ≥10 minutes.
     5. Run Cinebench R23 benchmark.
   • If subroutine fails, go back to 3.
   • If subroutine passes, go to Stage 2.

Stage 2: CPU Extended stability:

1. Run large scale stability testing subroutine.
   • Small scale stability testing subroutine.
     1. Run Cinebench R23 benchmark twice.
     2. Run CPU-Z stress test for ≥20 minutes.
     3. Run Prime95 mixed blend stress test
        • ≥12 hours required.
     4. Run CPU-Z stress test for ≥20 minutes.
     5. Run Cinebench R23 benchmark twice.
     6. Restart PC
     7. Leave at login screen for ≥12 hours.
   • If subroutine fails, decrease negative offset and go back to 1.
   • If subroutine passes, go to 2.
2. Decrease negative offset by 1.
3. System is considered stable.