Soldato
X99 Overclocking Guide
Firstly, make sure you have a look here. Thanks to ASUS ROG and Raja@ASUS there is plenty of 'get started' information. http://rog.asus.com/365052014/overclocking/rog-overclocking-guide-core-for-5960x-5930k-5820k/
Taking note of the following results found.
CPU | Frequency | Voltage
5960X | 4.6GHz | 1.30V = Good Result
5960X | 4.5GHz | 1.30V = Average Result
5960X | 4.4GHz | 1.30V = Fair Result
CPU Straps (125)
What strap you will be working with largely depends on what memory frequency you are going to be working with. 2666 and 3000 largely work best on 125 strap, where as 3200 is far happier on 100 strap. This may change as and when Intel update the microcode.
Certain frequencies such as 2800 requires the divider to step up BCLK to 127. This in turn steps PCI-E frequency up. The PCI-E and DMI cannot operate above 110 which is why the 1.25 divider is useful. Applying a strap downclocks the PCI-E /DMI. However it's only real benifit on this platform is that it allows you to use alternate memory frequencies, where as before on X79 for example increasing the base clock frequency (BCLK) actually helped stability due to the decreased modulation through PLL. There are no such benifits here anymore, you should be able to reach your optimum core frequency on both 100 and 125 strap.
Stress testing (core / uncore)
This is a difficult subject, and depending who you ask you may get a different answer. Firstly I would start by saying you do not want to be pushing tests that invoke AVX 2.0 and FPU related stressing. This is because Haswell-E unlike EP does not throttle CPU in intense load situations, and because of this, the CPU package can pull in excess of 400w - which according to ASUS representatives unsurprisingly is degrading territory. There are very few situations or in fact benchmarks where such high levels of this type of floating point calculation is even used, namely in some scientific application. So I would stay away from it. Prime95 v28 and above, and AIDA64 have FPU stressing, however AIDA allows for you to disable FPU.
I would recommend 2 to 4 hours depending on your workload of AIDA64 with FPU disabled, or Prime95 27.9. 4 hours for video editing / encoding and 2 hours for a gaming machine for example. Or if you prefer, just use the machine and see how it goes. There is no precise method for this. It's entirely workload dependant.
Getting Started
Firstly, apply your memory XMP profile.
I would start by applying 1.2v vcore in manual and trying for 4.2. If it boots into windows, raise the multi to 4.3 and so on. Leave cache at stock for now. For example I'm working with a fairly good sample at 4.375Ghz at 1.2 vcore. You want to be keeping core voltage below 1.4v on a decent water setup. For AIO coolers keep below 1.35v, and air 1.3v as a guideline.
VCINN (Input voltage)
Before you go anywhere I recommend increasing Input voltage to 1.9v. Having input at stock normally causes problems even potentially with low frequency overclocks. LLC (Loadline calibration) directly affects this voltage on this platform, not Vcore. You want to be using level 6 or 7 on reasonable overclocks. Do not exceed 2v VCINN.
Memory stability
For the sake of testing we will be working at 3000mhz working at 125 BCLK at XMP. I would recommend for this guide that you only adjust VCCSA (System Agent Voltage) if you are encountering memory instability.
Testing memory stability on Haswell-E and the 5960x processor can be fairy cantankerous due to both the number of threads and System Agent voltage. Unconditional memory stability (or close to) is something that is overlooked when testing memory on this platform.
I recommend HCI Memtest pro, which costs 5$. In order to correctly load up the memory bus you will need to run one instance per thread, not core. Loading up one per core is O.K, but you are not loading up the memory entirely. Loading up 16 instances is a brutal test of memory stability, but it's up to you. Once more it is workload dependant! You can feel free to use the machine for your daily tasks, it's personal preference.
You want to be using 85% of memory leaving enough for the OS and overheads. Should look something like below. Leave for a minimum of 200%. 400-600% optimally and 1000% is classed as the 'golden standard'. 400% should be more than adequate. This test can also flag any cache instability.
Example
System Agent voltage is a difficult one to cover, it is not a linear voltage. For example my sample likes around 0.9v for 2400 but needs around 1.016 to 1.048 for 3000mhz. You may find, especially when attempting to tighten down if you are planning on, that this voltage requires a bit of patience. Some samples may be fine in auto, but be warned board dependant you may find yourself above 1.15v. Using this much system agent normally isn't required, and some samples tend to be unstable above this voltage. Worst case do not exceed 1.2v
With the above in mind, I would work to the memory's XMP programming firstly. Then you can work on memory timings. This is more in-depth than I'd like to go into within the first post so this will come later. All I will say is do not be overly obsessed with DDR4s higher latency, due to far better ICs and 3rd tier timings, lowering latency doesn't have an overall massive impact to system performance. Memory write times are also very much uncore frequency dependant.
Uncore
Firstly the below is using an ASUS X99 Deluxe. Non 'OC Socket' motherboards will struggle to exceed 3.8 cache frequency for daily use.
For clocking Cache (uncore) much like when pushing core frequency it's CPU dependant. Some samples can achieve 4.6.-4.8. Cache is some what unpredictable. You may find it to be stable only to experience a random reboot within a week, maybe even a fortnight. So for this part I recommend you see what works best for you. When going above 4Ghz, stability may become tricky. If you are really struggingly with overclocking your cache I would recommend sticking with a manual voltage. When above 4.2 the cache really needs the voltage. If you are using offset ok, stick with it, only switch to manual voltage if you are finding cache stability difficult. You may find the system just shutting itself off when first experiementing with cache. This is Haswell-E's internal voltage regulator basically saying "no." There isn't anything wrong with your machine, it's just your cache isn't stable.
There seems to be a very steep climb with uncore, some samples may be able to do 4ghz with as little as 1.1v. For example my sample seems to like around 1.14v for 4Ghz. Above 4.2 I require 1.3v to eradicate cold boot instability, as I say you may find the system appears stable when it is not. Example one can stress test at 1.275v and appear stable but experience cold boot freezes without more voltage. 1.3v seemed to remove this for my sample.
Others may require in excess of 1.25-1.3v. All I will say is don't be afraid to feed the cache voltage. Up to 1.4v with adequate cooling is fine. I wouldn't exceed 1.45v on this voltage for daily use. AIDA64 is fairly cache intensive so I would recommend the latest stable build to test cache initially, minus FPU tests.
Passing POST
This can sometimes be more difficult than getting your system OS stable. Post stability issues are normally down to either memory or cache related instability. If you are an ASUS user you may find yourself running into the Q-Code 'BF' or bf continually. This code isn't assigned as of yet to the table, but you want to be looking at your cache voltage/frequency or SA voltage. These can be very sensitive to cold booting. At higher memory frequency you may find there is enough drift within memory training that your system is failing POST on occasion. You can disable memory training once you are sure the system is stable, however I would recommend re-running memory tests to insure that the system is stable as it will re-train one last time after disabling.
This is really an intermediate start out than a guide, but I'm welcome to answer any questions. There is an awful lot to cover with X99 which is still in it's infancy!
Links
Memtest Pro http://hcidesign.com/memtest/purchase.html
http://www.tweaktown.com/guides/689....it&utm_medium=twitter&utm_campaign=tweaktown
Firstly, make sure you have a look here. Thanks to ASUS ROG and Raja@ASUS there is plenty of 'get started' information. http://rog.asus.com/365052014/overclocking/rog-overclocking-guide-core-for-5960x-5930k-5820k/
Taking note of the following results found.
CPU | Frequency | Voltage
5960X | 4.6GHz | 1.30V = Good Result
5960X | 4.5GHz | 1.30V = Average Result
5960X | 4.4GHz | 1.30V = Fair Result
CPU Straps (125)
What strap you will be working with largely depends on what memory frequency you are going to be working with. 2666 and 3000 largely work best on 125 strap, where as 3200 is far happier on 100 strap. This may change as and when Intel update the microcode.
Certain frequencies such as 2800 requires the divider to step up BCLK to 127. This in turn steps PCI-E frequency up. The PCI-E and DMI cannot operate above 110 which is why the 1.25 divider is useful. Applying a strap downclocks the PCI-E /DMI. However it's only real benifit on this platform is that it allows you to use alternate memory frequencies, where as before on X79 for example increasing the base clock frequency (BCLK) actually helped stability due to the decreased modulation through PLL. There are no such benifits here anymore, you should be able to reach your optimum core frequency on both 100 and 125 strap.
Stress testing (core / uncore)
This is a difficult subject, and depending who you ask you may get a different answer. Firstly I would start by saying you do not want to be pushing tests that invoke AVX 2.0 and FPU related stressing. This is because Haswell-E unlike EP does not throttle CPU in intense load situations, and because of this, the CPU package can pull in excess of 400w - which according to ASUS representatives unsurprisingly is degrading territory. There are very few situations or in fact benchmarks where such high levels of this type of floating point calculation is even used, namely in some scientific application. So I would stay away from it. Prime95 v28 and above, and AIDA64 have FPU stressing, however AIDA allows for you to disable FPU.
I would recommend 2 to 4 hours depending on your workload of AIDA64 with FPU disabled, or Prime95 27.9. 4 hours for video editing / encoding and 2 hours for a gaming machine for example. Or if you prefer, just use the machine and see how it goes. There is no precise method for this. It's entirely workload dependant.
Getting Started
Firstly, apply your memory XMP profile.
I would start by applying 1.2v vcore in manual and trying for 4.2. If it boots into windows, raise the multi to 4.3 and so on. Leave cache at stock for now. For example I'm working with a fairly good sample at 4.375Ghz at 1.2 vcore. You want to be keeping core voltage below 1.4v on a decent water setup. For AIO coolers keep below 1.35v, and air 1.3v as a guideline.
VCINN (Input voltage)
Before you go anywhere I recommend increasing Input voltage to 1.9v. Having input at stock normally causes problems even potentially with low frequency overclocks. LLC (Loadline calibration) directly affects this voltage on this platform, not Vcore. You want to be using level 6 or 7 on reasonable overclocks. Do not exceed 2v VCINN.
Memory stability
For the sake of testing we will be working at 3000mhz working at 125 BCLK at XMP. I would recommend for this guide that you only adjust VCCSA (System Agent Voltage) if you are encountering memory instability.
Testing memory stability on Haswell-E and the 5960x processor can be fairy cantankerous due to both the number of threads and System Agent voltage. Unconditional memory stability (or close to) is something that is overlooked when testing memory on this platform.
I recommend HCI Memtest pro, which costs 5$. In order to correctly load up the memory bus you will need to run one instance per thread, not core. Loading up one per core is O.K, but you are not loading up the memory entirely. Loading up 16 instances is a brutal test of memory stability, but it's up to you. Once more it is workload dependant! You can feel free to use the machine for your daily tasks, it's personal preference.
You want to be using 85% of memory leaving enough for the OS and overheads. Should look something like below. Leave for a minimum of 200%. 400-600% optimally and 1000% is classed as the 'golden standard'. 400% should be more than adequate. This test can also flag any cache instability.
Example
System Agent voltage is a difficult one to cover, it is not a linear voltage. For example my sample likes around 0.9v for 2400 but needs around 1.016 to 1.048 for 3000mhz. You may find, especially when attempting to tighten down if you are planning on, that this voltage requires a bit of patience. Some samples may be fine in auto, but be warned board dependant you may find yourself above 1.15v. Using this much system agent normally isn't required, and some samples tend to be unstable above this voltage. Worst case do not exceed 1.2v
With the above in mind, I would work to the memory's XMP programming firstly. Then you can work on memory timings. This is more in-depth than I'd like to go into within the first post so this will come later. All I will say is do not be overly obsessed with DDR4s higher latency, due to far better ICs and 3rd tier timings, lowering latency doesn't have an overall massive impact to system performance. Memory write times are also very much uncore frequency dependant.
Uncore
Firstly the below is using an ASUS X99 Deluxe. Non 'OC Socket' motherboards will struggle to exceed 3.8 cache frequency for daily use.
For clocking Cache (uncore) much like when pushing core frequency it's CPU dependant. Some samples can achieve 4.6.-4.8. Cache is some what unpredictable. You may find it to be stable only to experience a random reboot within a week, maybe even a fortnight. So for this part I recommend you see what works best for you. When going above 4Ghz, stability may become tricky. If you are really struggingly with overclocking your cache I would recommend sticking with a manual voltage. When above 4.2 the cache really needs the voltage. If you are using offset ok, stick with it, only switch to manual voltage if you are finding cache stability difficult. You may find the system just shutting itself off when first experiementing with cache. This is Haswell-E's internal voltage regulator basically saying "no." There isn't anything wrong with your machine, it's just your cache isn't stable.
There seems to be a very steep climb with uncore, some samples may be able to do 4ghz with as little as 1.1v. For example my sample seems to like around 1.14v for 4Ghz. Above 4.2 I require 1.3v to eradicate cold boot instability, as I say you may find the system appears stable when it is not. Example one can stress test at 1.275v and appear stable but experience cold boot freezes without more voltage. 1.3v seemed to remove this for my sample.
Others may require in excess of 1.25-1.3v. All I will say is don't be afraid to feed the cache voltage. Up to 1.4v with adequate cooling is fine. I wouldn't exceed 1.45v on this voltage for daily use. AIDA64 is fairly cache intensive so I would recommend the latest stable build to test cache initially, minus FPU tests.
Passing POST
This can sometimes be more difficult than getting your system OS stable. Post stability issues are normally down to either memory or cache related instability. If you are an ASUS user you may find yourself running into the Q-Code 'BF' or bf continually. This code isn't assigned as of yet to the table, but you want to be looking at your cache voltage/frequency or SA voltage. These can be very sensitive to cold booting. At higher memory frequency you may find there is enough drift within memory training that your system is failing POST on occasion. You can disable memory training once you are sure the system is stable, however I would recommend re-running memory tests to insure that the system is stable as it will re-train one last time after disabling.
This is really an intermediate start out than a guide, but I'm welcome to answer any questions. There is an awful lot to cover with X99 which is still in it's infancy!
Links
Memtest Pro http://hcidesign.com/memtest/purchase.html
http://www.tweaktown.com/guides/689....it&utm_medium=twitter&utm_campaign=tweaktown
Last edited:
.