Overclocking Corsair Vengeance 3200 to 3600

mossmotorsport

Bronze Level Poster
So, much like @NoddyPirate I believe, I thought that by buying some of the highest spec kit out there, there wouldn't be a need for any overclocking.

And then....

Well the Crosshair VIII is made for overclocking, right?
The Corsair Vengeance RGB Pro marketing does reference specially selected IC's for further overclocking!
AMD talk about PBO2 and basically encouraging overclocking!
Asus GPU Tweak features a built in OC profile...

... with all of this it's been something of a red flag to a bull, and I've had far too much fun tinkering, with varying levels of success.

My question at the moment is, am I falling into a trap whereby the synthetic benchmark scores really very little real world difference? My RAM is a 64GB Corsair Vengeance RGB Pro Kit with CL16-18-18-36 timings at 3200Mhz. I've reduced the latency a little bit, and seen some minor improvement in benchmark scores.

My problem now however, is I may be reading too much into things, and am unsure on next steps.

I've successfully had the RAM running at 3600Mhz with CL18-19-19-35 timings, but Kahu RAM stress / error tests all show the RAM heating up to 68 degrees c, way too hot for my liking. The voltages persist at 1.35v, and I wouldn't really want to go higher than this. Voltage is the main cause of temperature gains, but this isn’t changing, so I’m not clear why it’s getting so hot? The RAM itself is Micron d-die according to Thaiphoon Burner.

One reason for really wanting to push for 3600Mhz, is so that I can have the FCLK frequency set at 1800Mhz rather than 1600Mhz. Unless I'm misunderstanding, unless my FCLK is running at 1800Mhz the RAM could feasibly be a minor bottleneck for performance on my Ryzen 5950x?

I guess a few questions, and an apology for such an unstructured post:

- 64GB 3600Mhz kits were unavailable via the configurator, so I opted for 64GB at 3200Mhz, was this a mistake?
- Would selling my current RAM, and purchasing a higher speed rated kit help to ensure no CPU bottlenecks, in addition to achieving marginally better FPS (this is matched with an Asus ROG Strix OC 3090)
- Temperatures seem to only marginally increase at 3400Mhz, I can manually set the FCLK in the BIOS to a frequency to match, as this is an odd / unusual frequency, could this be counter productive?
- Should my focus be on achieving the tightest timings possible with 3200Mhz, and accept that the infinity fabric won't be running at its fullest potential?
- Should I stop the benchmarks, and just enjoy what I have?! Although the fettling has been a bit of a giggle!

Cheers in advance!

Edit: The 64GB is used for extensive Photoshop, Lightroom, Virtualisation (Kali Linux, Nessus) and various other legitimate uses. I'd rather stick with this level if possible.
 
Last edited:

Bigfoot

Grand Master
Power is proportional to the square of the voltage, but also proportional to frequency. As you increase the frequency, you will be generating more heat.
 

NoddyPirate

Grand Master
I imagine you will know much of all this already - so forgive the Egg Sucking Lessons for Granny!
- Would selling my current RAM, and purchasing a higher speed rated kit help to ensure no CPU bottlenecks, in addition to achieving marginally better FPS (this is matched with an Asus ROG Strix OC 3090)
In all likelihood any FPS gains in games will be miniscule and not worth all the effort. The primary advantage of higher speed RAM is in memory intensive workloads ultimately. I overclocked from 3000 to 3200 MHz and did see the full 6-7% improvement - but only in memory intensive Lightroom workloads such as large exports or photomerges - which for me is great! Elsewhere it has made no noticeable or visible difference to me really.
- Temperatures seem to only marginally increase at 3400Mhz, I can manually set the FCLK in the BIOS to a frequency to match, as this is an odd / unusual frequency, could this be counter productive?
There is no such thing as an 'odd' frequency - standard RAM speeds actually go up in 66's and 33's like this - 3000 - 3066 - 3133 - 3200 - 3266 - 3333 - 3400 - etc, etc.

As long as the FCLK is running at the same speed as the RAM (i.e. half the working speed for DDR memory - so 1700 MHz against your 3400 MHZ RAM) then all will be well. Ryzen really needs the FCLK and Memory Clock to be synchronised or performance will suffer, no matter what RAM speed you select.

68 degrees is not crazy high temperature for RAM, but I wouldn't want to be over 70 degrees personally, so if it was me I would probably want to dial it down a bit. I imagine RGB RAM will run that bit warmer too.

Question though - what CL could you manage at 3400 MHz?

- Should my focus be on achieving the tightest timings possible with 3200Mhz, and accept that the infinity fabric won't be running at its fullest potential?
It all depends on what you can achieve. I was able to move from CL16 @ 3000 MHz to CL16 at 3200 MHz - so I saw a real tactile improvement in latency in terms of nanoseconds. As a general rule you want the lowest latency in terms of time. The basic calculation is:

(CAS Latency / Speed in MHz) * 2000.

So for example, my 3000 MHz RAM at CL16 has a latency of 10.6 nanoseconds.

Getting it to 3200 MHz still at CL16 reduced that latency to 10.0 nanoseconds. So I get higher bandwidth and lower latency - win win.

If the best I could do was overclocking it to 3200 MHz but at CL18, that would give me a latency of 11.25 which is worse. The benefit of the higher bandwidth would be lost in the higher latency. Even with the higher FCLK I would most likely still be worse off in the long run.

If I could manage CL17 at 3200 MHz, then the latency would actually be about the same at 10.6. Although the true latency is unchanged the higher bandwidth would still see better performance so it's still worth it, particularly since I get a higher FCLK too. I do think I saw an improvement in the CPU performance at my higher FCLK, but it was marginal and most likely more to do with me wanting to see a difference than there actually being one. :)

The above is exactly the case for you with 3200 MHz at CL16 vs 3600 MHz at CL18. Your latency is identical between the two, but the 3600 MHz will perform better in theory at least. The gain might just be small and hard to spot depending on what you're doing.

So I would say go for as high a speed as you can - with the lowest latency you can achieve at that speed - as long as you are not getting a true latency that is worse that what you started with. Don't worry about the FCLK - as long as it's synchronised correctly then maximising your RAM speed, without suffering a latency penalty, will get you the best of both worlds.

Your RAM temperature comfort level will probably be the upper limitation in what you settle on I think.

- Should I stop the benchmarks, and just enjoy what I have?! Although the fettling has been a bit of a giggle!
Completely YES! And also emphatically NO! Cos it's fun! :)

As you allude to though - don't rely on Benchmarks. Actual real life performance is all that matters. If you aren't seeing a real, tangible benefit in your own workflows, then go back a step and try again, or go back to stock and keep it all happy as it came.
 

Bigfoot

Grand Master
Is there a point when over locking RAM might harm performance, for example when the fabric clock and memory clock cannot match?
 

NoddyPirate

Grand Master
Is there a point when over locking RAM might harm performance, for example when the fabric clock and memory clock cannot match?
Absolutely it seems. Everything I have read says that Ryzen needs the FCLK and RAM to be perfectly synchronised or performance will suffer as the actual working latency increases as a result of the mismatch.

Ryzen 2nd Gen was considered to be frequently unstable above 1800 MHz FCLK, but I think 3rd Gen is better now and 1900 MHz - or perhaps even more - may work perfectly fine.
 

mossmotorsport

Bronze Level Poster
That’s excellent advice thanks, I’ll need to read and digest all of that. I’m just thankful it’s posting now, the safe timings on the ryzen dram timing tool seemed to be very unsafe and unfortunate for me! First time messing with gear down and anything other than the first few digits though. Thankfully the clear cmos button worked a treat and saved my bacon!
 

NoddyPirate

Grand Master
That’s excellent advice thanks, I’ll need to read and digest all of that. I’m just thankful it’s posting now, the safe timings on the ryzen dram timing tool seemed to be very unsafe and unfortunate for me! First time messing with gear down and anything other than the first few digits though. Thankfully the clear cmos button worked a treat and saved my bacon!
Just remember that clearing CMOS may well lose all your BIOS settings. Save your default settings under a User Profile slot under the TOOLS page in BIOS so you can go back to them with a single click. (y)
 

mossmotorsport

Bronze Level Poster
For anyone interested, my Corsair Vengeance Pro RGB 3200MHz 64GB kit (CL16-18-18-36) now runs at a comfortable 3400MHz with CL16-16-17-36 timings. This has seen notable benchmark improvements:

Baseline, stock D.O.C.P:

mem_baseline.PNG


After OC'ing:

mem_3673.PNG


Voltage was increased slightly from 1.350v to 1.375v. Temperatures are slightly raised, which I'm still trying to learn more about. Corsair themselves state that this is fine, with no issues until much closer to 80+ degrees. Other (older) forum posts suggest RAM can have issues at 50 degrees, including corruption.

Stock was ~ 40 degrees c, increasing to 5 degrees c under load.
OC is now hovering ~ 47 degrees c, increasing up to a maximum of 62 degrees c under load, but more commonly ~ 60.

A 9 hour RAM stress test returns no errors too, which makes me happy!

ram.PNG
 

NoddyPirate

Grand Master
For anyone interested, my Corsair Vengeance Pro RGB 3200MHz 64GB kit (CL16-18-18-36) now runs at a comfortable 3400MHz with CL16-16-17-36 timings. This has seen notable benchmark improvements:

Baseline, stock D.O.C.P:

View attachment 26059

After OC'ing:

View attachment 26060

Voltage was increased slightly from 1.350v to 1.375v. Temperatures are slightly raised, which I'm still trying to learn more about. Corsair themselves state that this is fine, with no issues until much closer to 80+ degrees. Other (older) forum posts suggest RAM can have issues at 50 degrees, including corruption.

Stock was ~ 40 degrees c, increasing to 5 degrees c under load.
OC is now hovering ~ 47 degrees c, increasing up to a maximum of 62 degrees c under load, but more commonly ~ 60.

A 9 hour RAM stress test returns no errors too, which makes me happy!

View attachment 26061
Bravo! 👏

Is that Ram Test by Karhu? If yer bored - we’d recommend MemTest86 as the most thorough stability test. It’ll take a similar length of time to the above so you could leave it run overnight....
 

mossmotorsport

Bronze Level Poster
Bravo! 👏

Is that Ram Test by Karhu? If yer bored - we’d recommend MemTest86 as the most thorough stability test. It’ll take a similar length of time to the above so you could leave it run overnight....
Yeah that's Karhu, I read a few different posts regarding memory stress testers. Older posts, admittedly, stating that MemTest86 was not as likely to identify errors compared to Karhu? As you've said though, if it gets left to run over night it won't do any harm to run that too, and it's free!

Hopefully that's my finished tweaking for now. PBO appears to be working well with a manual configuration and curve offset, with negative values varying depending on the strongest cores.
 

NoddyPirate

Grand Master
Yeah that's Karhu, I read a few different posts regarding memory stress testers. Older posts, admittedly, stating that MemTest86 was not as likely to identify errors compared to Karhu? As you've said though, if it gets left to run over night it won't do any harm to run that too, and it's free!

Hopefully that's my finished tweaking for now. PBO appears to be working well with a manual configuration and curve offset, with negative values varying depending on the strongest cores.
Super! Well done!

Don’t forget to consider changing the Windows Power Plan if it still has the Ryzen plan as an option... 🙂
 

ubuysa

The BSOD Doctor
Yeah that's Karhu, I read a few different posts regarding memory stress testers. Older posts, admittedly, stating that MemTest86 was not as likely to identify errors compared to Karhu? As you've said though, if it gets left to run over night it won't do any harm to run that too, and it's free!

Hopefully that's my finished tweaking for now. PBO appears to be working well with a manual configuration and curve offset, with negative values varying depending on the strongest cores.
The thing with RAM testing is that 'hard' RAM faults can be found by almost any testing tool, but intermittent or 'soft' faults often require the RAM to be tested and retested again and again before you can make it fail. We know that Memtest86 runs 13 different tests each of which is designed to stress the RAM in different ways, they are described in detail here. Memtest86 also has a free version that runs up to four iterations of these 13 tests. With many 'soft' RAM errors you need several iterations in order to reveal the error. Any 'hard' errors will be caught in just one iteration but more obscure (but no less fatal) errors require multiple iterations. The generally accepted wisdom is that 8 iterations of Memtest86 will find 99% of RAM issues, though the paid-for Pro version does up to 15 iterations.

I confess that I don't know the Karhu product but a look around their website doesn't tell me what their testing philosophy is and they don't seem to have a free version to try, so I can't find out even how it can be configured nor how it tests RAM. I'll agree that the cost of a license is minimal but I'm naturally wary of paying for software that I can't at least trial. Their readme file doesn't describe their testing methodology at all either, though there is talk of a random number generator. Hopefully that's not the entirety of their testing methodology?

I will keep Karhu in mind and if they ever produce a free or trial version I'll give it a go, but Memtest 86 is a known quantity, we know how it tests and that is really very important. Thanks for the tip though. :)
 

mossmotorsport

Bronze Level Poster
The thing with RAM testing is that 'hard' RAM faults can be found by almost any testing tool, but intermittent or 'soft' faults often require the RAM to be tested and retested again and again before you can make it fail. We know that Memtest86 runs 13 different tests each of which is designed to stress the RAM in different ways, they are described in detail here. Memtest86 also has a free version that runs up to four iterations of these 13 tests. With many 'soft' RAM errors you need several iterations in order to reveal the error. Any 'hard' errors will be caught in just one iteration but more obscure (but no less fatal) errors require multiple iterations. The generally accepted wisdom is that 8 iterations of Memtest86 will find 99% of RAM issues, though the paid-for Pro version does up to 15 iterations.

I confess that I don't know the Karhu product but a look around their website doesn't tell me what their testing philosophy is and they don't seem to have a free version to try, so I can't find out even how it can be configured nor how it tests RAM. I'll agree that the cost of a license is minimal but I'm naturally wary of paying for software that I can't at least trial. Their readme file doesn't describe their testing methodology at all either, though there is talk of a random number generator. Hopefully that's not the entirety of their testing methodology?

I will keep Karhu in mind and if they ever produce a free or trial version I'll give it a go, but Memtest 86 is a known quantity, we know how it tests and that is really very important. Thanks for the tip though. :)

Good news, memtest86 showed a nice clean pass :)

memtest.PNG
 
Top