5930K vs 5960X

[PipJones]

I have recently replaced my 5930K with a 5960X.

Voltages aside, overclock settings are the same between each system - 44x CPU and 42x cache.

Q1: I have C-STATES disabled. I am using Windows power settings to have the CPU clock down to ~12x when idle (minimum processor @ 20% or something). At 40x cache, this works without issue. At 42x cache, the PC will hang just after dropping the multiplier. I've tried LLC Auto -> 9 and increasing DRAM voltage. It did not appear to make a difference. Can anyone suggest which things in BIOS should I focus on to try and fix this issue?

*NOTE. If I use high performance power plan, system does not hang.


Q2: I am using AIDA64 cache and memory benchmark tool to measure performance. I am getting the expected result for READ and COPY, however, WRITE remained the same - it did not increase inline with READ / COPY. Is this expected / normal?

5930K



5960X



Many thanks!

[Tonyboj]

I've never been a fan of using the Balanced power profile in Windows. But if you want to do it successfully you'll have to boost ring voltage/cpu input voltage/pwm phase and dram, then it will work fine. If you can live with the extra power draw (and temperature) of running at locked 100% then you'll get away with lower voltages (that to a some degree will mitigate the higher performance profile). As side note I had the cache at 30x and cpu at 44-45x. Couldn't find any instance except for synthetic benchmarks where it made a difference.

As for your second question Aida has not always been reliable during memory benchmarks. I can't remember if it was HardOCP or some other review site that did a big comparison of all high core count processors and found that Aida scores were all over the place. I've used Intels own memory tool and I think it gives more reliable numbers. Maybe the more experienced members have other suggestions.

Intel Memory Latency Checker can be found here https://software.intel.com/protected.../739797/493768

It has some useful tests. You can benchmark with read-write-ratios that more closely resemble some real-world scenarios.

In short, I wouldn't, stress too much over the write-score. I don't think it is measurable in any real-world testing scenario. Also I would not set the cache to higher than 3200. Preferably keep it at 3000.

[Arne Saknussemm]

Are you setting manual cache voltage for 42x...because the board can set voltage that will power a small town at higher ratios

[PipJones]

Both, thanks for taking the time to reply. Good to see that this forum is still active.

Before we begin, I'm on good cooling. I know the voltages are close to the acceptable limit. Temps are good.

Are you setting manual cache voltage for 42x...because the board can set voltage that will power a small town at higher ratios

For the CPU, I am Adaptive. Auto and 1.350 for Turbo. Was hitting around 1.360 on VCore0
For the cache, I am Offset. I got as high as +0.405. Was hitting > 1.380 on CPU Cache voltage. Adaptive would not boot.

I've never been a fan of using the Balanced power profile in Windows. But if you want to do it successfully you'll have to boost ring voltage/cpu input voltage/pwm phase and dram, then it will work fine. If you can live with the extra power draw (and temperature) of running at locked 100% then you'll get away with lower voltages (that to a some degree will mitigate the higher performance profile). As side note I had the cache at 30x and cpu at 44-45x. Couldn't find any instance except for synthetic benchmarks where it made a difference.

As for your second question Aida has not always been reliable during memory benchmarks. I can't remember if it was HardOCP or some other review site that did a big comparison of all high core count processors and found that Aida scores were all over the place. I've used Intels own memory tool and I think it gives more reliable numbers. Maybe the more experienced members have other suggestions.

Intel Memory Latency Checker can be found here https://software.intel.com/protected.../739797/493768

It has some useful tests. You can benchmark with read-write-ratios that more closely resemble some real-world scenarios.

In short, I wouldn't, stress too much over the write-score. I don't think it is measurable in any real-world testing scenario. Also I would not set the cache to higher than 3200. Preferably keep it at 3000.

"Balanced power profile in Windows": I'm not a fan of Windows either, but, needs must. I have customised the "High Performance" power plan, minimum processor only. I have to use this as I can not get the system stable with C-States. It is not locked at 100% and is dropping multiplier (and voltages) accordingly. See SIV LCD panel.

"boost ring voltage/cpu input voltage/pwm phase and dram": OK, thanks. Kinda my thoughts too. I've gone as high as I feel comfortable with cache (ring) / cpu. I will have a tinker with the others.

"Aida scores were all over the place": Yes, I agree. There is a lot of variance and unexpected results. I have never seen it with this magnitude. I've used AIDA for so long now, it has been my "performance" measuring stick. Time to renew license and post in their (his!) forum too!

"I wouldn't, stress too much over the write-score. I don't think it is measurable in any real-world testing scenario.": Noted, the only real gain is in benchmarks. It is fun trying tho! It's why I (we?) buy these boards and CPU's that are designed for overclocking.



SIV LCD Panel showing stable system at idle with Windows power plan active.

[PipJones]

BIOS, see if anything leaps out ...

https://photos.app.goo.gl/r5anm5KZhhB41VyG6

[PipJones]

Results from MLC using stable x44 CPU and x40 Cache

Microsoft Windows [Version 10.0.17763.437]
(c) 2018 Microsoft Corporation. All rights reserved.

C:\WINDOWS\system32>d:

D:\>cd D:\5960x\mlc_v3.6\Windows

D:\5960x\mlc_v3.6\Windows>mlc
Intel(R) Memory Latency Checker - v3.6
Measuring idle latencies (in ns)...
Numa node
Numa node 0
0 47.9

Measuring Peak Injection Memory Bandwidths for the system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using traffic with the following read-write ratios
ALL Reads : 74724.5
3:1 Reads-Writes : 72510.1
2:1 Reads-Writes : 72834.1
1:1 Reads-Writes : 71387.4
Stream-triad like: 67965.6

Measuring Memory Bandwidths between nodes within system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Numa node
Numa node 0
0 72737.4

Measuring Loaded Latencies for the system
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Inject Latency Bandwidth
Delay (ns) MB/sec
==========================
00000 126.89 75110.0
00002 128.89 75120.0
00008 125.47 75470.2
00015 126.40 75533.8
00050 119.63 75468.6
00100 88.02 65519.0
00200 70.30 37739.9
00300 64.85 26143.8
00400 60.77 20149.2
00500 59.60 16441.3
00700 57.82 12162.9
01000 56.54 8911.2
01300 55.74 7146.7
01700 55.17 5752.9
02500 53.09 4336.6
03500 53.15 3443.7
05000 51.86 2803.8
09000 51.63 2112.4
20000 51.31 1640.7

Measuring cache-to-cache transfer latency (in ns)...
Using small pages for allocating buffers
Local Socket L2->L2 HIT latency 22.5
Local Socket L2->L2 HITM latency 25.3

D:\5960x\mlc_v3.6\Windows>



AIDA64

[Tonyboj]

BIOS, see if anything leaps out ...

https://photos.app.goo.gl/r5anm5KZhhB41VyG6

Looks good, only thing I would make sure is that DRAM is fixed at a minimum of 1.35-1.38. The bios reading is showing ~1.2. Also, if I remember correctly, wasn't SVID support supposed to be enabled.

[Arne Saknussemm]

SVID should be disabled for both CPU and DRAM if you are setting manual OCs and voltages

Adaptive cache was never fixed by Intel. 1.38v for cache is pretty ridiculous for anything but short LN2 runs

[PipJones]

SVID should be disabled for both CPU and DRAM if you are setting manual OCs and voltages

Adaptive cache was never fixed by Intel. 1.38v for cache is pretty ridiculous for anything but short LN2 runs

Thanks, yes, SVID is disabled for both CPU and DRAM.

I was using this guide for reference with the Cache overclock:

https://rog.asus.com/articles/overcl...x-5930k-5820k/

"Good processor samples can achieve 4.6GHz Uncore frequency in tandem with CPU core frequency at the same value. Depending upon the processor sample, this may require cache voltage in the region of 1.35~1.45V."

"We may need to apply a small boost to Vcore and VCCSA when the cache frequency is overclocked. This is because a faster cache ratio will increase the amount of data over the bus thanks to faster L3 cache access times."

"VCCSA helps stabilize the processor's memory controller. For the most part the ROG UEFI BIOS left to 'Auto' will scale fine for most CPU/memory combinations. Some CPUs may need manual adjustment. and the maximum we have needed to use is 1.15V. However, some CPUs do not respond well to anything over 1.05V*; such processors usually overclock memory fine at stock settings - even though they prefer lower voltage. Start low (default is 0.80V approx.) and work upwards. The relationship between voltage increases to VCCSA and stability are non-linear. There are certain ranges of voltage which may exhibit worse stability, once you go past them (higher or lower) stability improves once again. Try 1.02V for memory speeds above DDR4-2900 as a starting point and work up or down from there."



I really was not comfortable going above 1.35 for both CPU and Cache. So I guess I am maxed out there. Time to start focusing on VCCSA. If I don't get anywhere with that I will have to accept that the Cache on this chip has a 4Ghz (100 x 40) limit.

[Vlada011]

I think write-copy-read depend of memory modules as well.
My Dominator Platinum 2666 are all on 50-52MB/s and grow on 61-62.000MB/s with increased Cache, i7-5820K.
Copy-Write-Read, all.
With Dual sided 32GB 4x 8GB Dominator Platinum 3200MHz I would have around 65-70.000MB/s

It's weird how here stay perfectly same Write Score.