![[VR] optimal_909](https://forums-cdn.zaonce.net/data/avatars/m/109/109067.jpg?1553781733){kind=avatar}
I am currently running my G.Skill Ripjaws V 2x8 Gb sticks at 3200 Mhz, I wonder what headroom I'd have in it, and if it makes sense to explore it.
Hardware & Technical Just in time for the Alpha
- Thread starter Morbad
- Start date
I am currently running my G.Skill Ripjaws V 2x8 Gb sticks at 3200 Mhz, I wonder what headroom I'd have in it, and if it makes sense to explore it.
Would need to know the kit's exact model number or the ICs it uses to get an idea of what they're capable of without just taking a shot in the dark. Most newer 3200 stuff has a fair bit of headroom, and pretty much anything can have it's subtimings optimized, even if it can't be clocked higher.
How much increasing memory performance can benefit would depend on the rest of the system and the workload in question.
Here it is:
3000MHz F4-3000C15D-16GVGB
MR Flight Sim is the only application where I have any bottleneck, especially memory.
Those use Samsung 4Gb D-die ICs, they generally aren't very good. Probably not a whole lot of headroom in them.
This new 5800X sample is quite solid. Finally have a good baseline set of PBO curves for it at +150MHz. Five of the eight cores are reliably boosting to 5GHz, two to 4.9GHz, and the weakest core still hits 4.85GHz...that's under lightly threaded loads of course. Heavily threaded loads range from 4.3GHz in LINPACK to 4.7GHz in demanding real-world applications. Games regularly hit 5GHz and typically stay well above 4.7. Not bad for a 120mm air cooler in a SFF system. I'm running 'motherboard' power limits, which are essentially uncapped and the part can spike to about 170w before becoming thermally constrained. Only one core needed a tiny positive curve offset for the +150MHz and another core could take the full -30 offset, with most somewhere in the middle of the negative (negative is good for the Zen 3 FV curve as it sets target voltage for a given point on the clock slope) range.
My previous sample couldn't do more than stock boost with a decidedly ugly set of curves and probably had an outright defective SoC die. This new part is a good 5-10% faster in practice. They're from the same batch and manufacturing date, so this level of variance was a bit surprising. Earlier part probably barely qualified as a 5800X (they probably should have disabled it's two worst cores and sold the thing as a 5600X), while this new one was likely only a smidge away from making the cut to be the better of the two dies in a 5950X.
I think I'll keep it.
My previous sample couldn't do more than stock boost with a decidedly ugly set of curves and probably had an outright defective SoC die. This new part is a good 5-10% faster in practice. They're from the same batch and manufacturing date, so this level of variance was a bit surprising. Earlier part probably barely qualified as a 5800X (they probably should have disabled it's two worst cores and sold the thing as a 5600X), while this new one was likely only a smidge away from making the cut to be the better of the two dies in a 5950X.
I think I'll keep it.
New AMD graphics drivers seem pretty good. Performance at a given clock speed and clock speed potential at a given voltage are both up, reversing the degradation the 21.4.x and 21.5.x drivers had vs. 21.3.x in this regard.
While cleaning my system I also noticed that I had accidentally cross-wired a pair of my fans the last time I cleaned, resulting in one intake and one exhaust running at the wrong fan speed. Fixed that and a temperature anomaly along with it.
So far, so good:
CPU wise I found an new way to test my PBO curve and discovered that cold could make the CPU boost to the point of instability. Had to run some of the lighter, but still AVX2, y-cruncher tests with the lowest ambients I could manage to get my curve offsets straightened out. Lost a small amount of CPU performance, but I'm now extremely confident in it's stability.
Edit: After a night of stress testing, the clock gains are proving to not be as high as I had hoped. Had to bump GPU core voltages back to prior levels. I did get the ability to max out the memory clock again, however.
While cleaning my system I also noticed that I had accidentally cross-wired a pair of my fans the last time I cleaned, resulting in one intake and one exhaust running at the wrong fan speed. Fixed that and a temperature anomaly along with it.
So far, so good:
CPU wise I found an new way to test my PBO curve and discovered that cold could make the CPU boost to the point of instability. Had to run some of the lighter, but still AVX2, y-cruncher tests with the lowest ambients I could manage to get my curve offsets straightened out. Lost a small amount of CPU performance, but I'm now extremely confident in it's stability.
Edit: After a night of stress testing, the clock gains are proving to not be as high as I had hoped. Had to bump GPU core voltages back to prior levels. I did get the ability to max out the memory clock again, however.
Last edited:
Grabbed a 6900 XT a few months back when they decent AIB version became available for not-complete-insane prices Still cost a fortune and was hoping to have had a modest upgrade to this system so I could move this 6800 XT into my new HTPC (which needs a strong GPU because it's what my VR headset is connected to, as well as pushing my 4k TV). Unfortunately, my first 6900 XT sample was defective and the replacement is just a mediocre overclocker. Works fine at stock, and actually undervolts extremely well, but fully tuned it's still only about 8% faster than my 6800 XT (which is a very good sample) in high-res shader limited stuff, and actually a bit slower in a few tests. So, the 6900 XT is going in the HTPC and this SFF gaming-oriented setup is keeping the 6800 XT for the forseeable future.
Anyway, I've been learning a lot about tuning these Navi 21 parts and have some settings I'm pretty happy with on the 6800 XT:
Except for the last part, all of that was done via MorePowerTool, which allows one to edit the Soft PowerPlay Tables of AMD GPUs. Peak stable clocks have varied with drivers version a bit, earlier drivers allowing for higher clocks, but having worse actual performance and generally more bugs, but that seems to have stabilized a while back, probably around the 21.8.x drivers. Another factor that caused me to reduce clocks for a while was the discovery that certain loads could be prompted to crash far below the peak loads I was primarily testing with. This I mostly resolved by increasing minimum SoC voltage.
CPU wise, I've been a bit more conservative with my tuning as I'm more cooling limited and exhaustive testing of the per-core curves is extremely time consuming. There have also been some serious issues with AMD's recent AGESA versions, namely anything after 1.2.0.3c introducing some very annoying bugs. 1.2.0.4 introduced a CLDO VDDG voltage cap of 1v, which is not quite sufficient for 1900MHz FCLK on my 5800X sample. 1.2.0.5 kept that bug and introduced a voltage limiter when passing 140A EDC. So, 1.2.0.3c is the best option for Vermeer currently.
I'm tentatively planning two hardware upgrades for this system in the near future:
Anyway, still can't run Odyssey at 4k with the custom settings I want to use, without occasionally dipping below 60 fps...
Anyway, I've been learning a lot about tuning these Navi 21 parts and have some settings I'm pretty happy with on the 6800 XT:
- Reduced GPU core voltage (vs. stock) by 100mV at the low end and 50mV peak
- Increased GPU SoC voltage by 100mV at the low end, but reduced peak SoC voltage by 50mV. This was necessary to make the SoC, FCLK, and memory clock increases I'm using unconditionally stable in certain edge case tests with extreme transient loads and peak frame rates in the four figures.
- Mild SoC clock (which dictates the speed of the memory controller) increase from 1200MHz to 1269MHz and an FCLK increase (L3 cache performance) from 1940 to 2052MHz.
- Jacked up the stock power limit to 375w/300A.
- Set 2550-2650MHz core clock, 2100MHz memory clock with fast timings, and +15% power target (432w peak, which is never actually reached, but it's there if something needs it).
Except for the last part, all of that was done via MorePowerTool, which allows one to edit the Soft PowerPlay Tables of AMD GPUs. Peak stable clocks have varied with drivers version a bit, earlier drivers allowing for higher clocks, but having worse actual performance and generally more bugs, but that seems to have stabilized a while back, probably around the 21.8.x drivers. Another factor that caused me to reduce clocks for a while was the discovery that certain loads could be prompted to crash far below the peak loads I was primarily testing with. This I mostly resolved by increasing minimum SoC voltage.
CPU wise, I've been a bit more conservative with my tuning as I'm more cooling limited and exhaustive testing of the per-core curves is extremely time consuming. There have also been some serious issues with AMD's recent AGESA versions, namely anything after 1.2.0.3c introducing some very annoying bugs. 1.2.0.4 introduced a CLDO VDDG voltage cap of 1v, which is not quite sufficient for 1900MHz FCLK on my 5800X sample. 1.2.0.5 kept that bug and introduced a voltage limiter when passing 140A EDC. So, 1.2.0.3c is the best option for Vermeer currently.
I'm tentatively planning two hardware upgrades for this system in the near future:
- A set of Phanteks T30-120 fans for my lower intake, to give the GPU heatsink a bit more airflow.
- An AMD Ryzen 7 5800X3D drop-in CPU replacement. This will be the final hurrah for socket AM4 and is a slightly lower clocked 5800X, but with three times the total L3 cache (96MiB vs. 32MiB). If I were building a system from scratch, I'd probably never consider this part and just build an Alder Lake or wait to see how AM5 performs, but it should be a fair upgrade for CPU limited games that I don't have to completely rebuild this system for, and would let me move this 5800X to my HTPC (which currently has a 3900X).
Anyway, still can't run Odyssey at 4k with the custom settings I want to use, without occasionally dipping below 60 fps...
The differences between XMP and default JEDEC settings can be quite large on some platforms with some memory kits. This is especially true when the ICs used are heavily binned (some of the fastest DDR4 around is still based on ICs that are officially rated for 2133MT/s by the semiconductor manufacturer) and on platforms that tie other aspects of performance to the memory clock (AM4 processors, for example, run their FCLK at the memory clock, by default).
That said, while XMP can be a good starting point, I generally do all my tuning myself and don't fuss over anything other than memory ICs used, PCB layout, and price...unless I have reason to believe something else in the specs is particularly indicative of binning.
For example, this system has some of the cheapest memory I could find at the time I built it (auto SPD settings highlighted, XMP settings at the bottom...most subtimings not shown, but they are very loose without manual changes, even with XMP):
And these are the settings I settled on (with 50mV less than XMP gives it):
There is definitely faster memory, but since the practical limit for Vermeer CPUs is 1900MHz FCLK and the platform is not hugely sensitive to primary timings, there is really little reason to spend too much on memory. I have other memory kits that could do the same clock and similar subtimings (though about half the tRFC, which is significant), with straight 14s on the primaries, but the performance is maybe 5% different in purely memory limited scenarios, which are rare.
Upgrading this system to a 5800X3D will also make main memory performance even less relevant, by significantly reducing the cache miss rates.
Well, I upgraded to a 5800X3D a while back, which proved to be a major boon for Odyssey and several other games of dubious optimization.
Right now I'm getting the system ready to accept an RTX 4090 when they drop. Assuming I can find a version that will fit.
I've swapped out the PSU with the 750w Seasonic SFX platinum version that has been in my test bench for the last six months or so, both for the higher capacity (was previously using the 650w SFX-L version) and the third PCI-E 8-pin power connector on it, which I'll need if I max out on the GPU I can conceivably cool. Main reason for choosing it over something like the Corsair version was that I could reuse all the cabling that was already in place, as it's a significant pain to swap out of this box. Downside is that the fan profile on the this model is ...PSU is commonly the loudest part of my system now.
I've also traded the two Noctua NF-A12x25's in the bottom of my case for a pair of Phanteks T30-120s and I'm quite impressed with them. Smaller hub diameter combined with being 5mm thicker allows them to move a fair bit more air through the filters and fin stack for a given noise level. The trick will be finding an RTX 4090 that has a fin stack short enough to fit in the space I have remaining (after stripping any stock fans and other excess equipment from the card), which is down to about 38mm. Still, anything that's only three slots or less should fit and I'm sure some AIB versions will qualify.
Right now I'm getting the system ready to accept an RTX 4090 when they drop. Assuming I can find a version that will fit.
I've swapped out the PSU with the 750w Seasonic SFX platinum version that has been in my test bench for the last six months or so, both for the higher capacity (was previously using the 650w SFX-L version) and the third PCI-E 8-pin power connector on it, which I'll need if I max out on the GPU I can conceivably cool. Main reason for choosing it over something like the Corsair version was that I could reuse all the cabling that was already in place, as it's a significant pain to swap out of this box. Downside is that the fan profile on the this model is ...PSU is commonly the loudest part of my system now.
I've also traded the two Noctua NF-A12x25's in the bottom of my case for a pair of Phanteks T30-120s and I'm quite impressed with them. Smaller hub diameter combined with being 5mm thicker allows them to move a fair bit more air through the filters and fin stack for a given noise level. The trick will be finding an RTX 4090 that has a fin stack short enough to fit in the space I have remaining (after stripping any stock fans and other excess equipment from the card), which is down to about 38mm. Still, anything that's only three slots or less should fit and I'm sure some AIB versions will qualify.
Looking at these RTX 4090 leaks and inferring from RTX 3090 Ti boards and coolers, many of which may be nearly identical to RTX 4090 parts...fitting one of these monsters in this system is going to take some work.
Even if I dump the shroud and backplate, the heatsink proper on many of these cards looks like it will be too long to fit the case, and thick enough that I'd have to swap back to my noctua fans, or something even lower profile. So, I'm in the awkward position of looking for the worst cooling money can buy on a new high-end card.
I'm also slightly annoyed with my Seasonic SPX-750. It's definitely the loudest component in my system. I might demote it back to testbench PSU and grab one of the new ASUS SFX-L power supplies with the 12VHPWR connectors...if I can ensure the pinout is close enough for me to not have to pull all my cabling.
Even if I dump the shroud and backplate, the heatsink proper on many of these cards looks like it will be too long to fit the case, and thick enough that I'd have to swap back to my noctua fans, or something even lower profile. So, I'm in the awkward position of looking for the worst cooling money can buy on a new high-end card.
I'm also slightly annoyed with my Seasonic SPX-750. It's definitely the loudest component in my system. I might demote it back to testbench PSU and grab one of the new ASUS SFX-L power supplies with the 12VHPWR connectors...if I can ensure the pinout is close enough for me to not have to pull all my cabling.
No real concerns about Great Wall, but I still need to find a pin-out. It's likely ASUS had that customized to match their other units, but I'd like to know for sure.
Well, I snagged an RTX 4090. Not one of the models I was looking for, but a less optimal model I can get is better than the perfect one I can't.
Anyway, this necessitated a new case, and the only case that checked most of my boxes that will also hold this monstrosity was the ASUS Prime AP201.
Going to keep this system mostly intact and swap it with my HTPC, which has an identical motherboard and drives (I've gotten in the habit of building duplicate systems, so if my main one breaks, I can just reposess the HTPC and load my backups over it).
Plus side is that I no longer have to use this loud SFX PSU and can use my better full size ATX unit.
Anyway, this necessitated a new case, and the only case that checked most of my boxes that will also hold this monstrosity was the ASUS Prime AP201.
Going to keep this system mostly intact and swap it with my HTPC, which has an identical motherboard and drives (I've gotten in the habit of building duplicate systems, so if my main one breaks, I can just reposess the HTPC and load my backups over it).
Plus side is that I no longer have to use this loud SFX PSU and can use my better full size ATX unit.