RAM voltage 1.43?

[Shahmatt]

Hi,

I have a AMD 2200G with MSI B350i board.

I'm now testing my HyperX Fury 2x4GB at 3400 CL14 timings. I've been forced to raise the voltage to 1.43 volts to keep it stable.

I've read in places that it's safe to go up to 1.45 volts, but I just want to be sure with the community here. What do you think?

These memory sticks are Micron B die, Rank 1.

Any insight is appreciated.

 

[Barty1884]

A lot of it is outright guesswork, with too many variables to list.

Some round-about logic.
1. XMP2.0 (Intel) allows an absolute max of 1.5V for certification, with preference for 1.35V or less.
2. The faster kits (3200MHz) that appear on the QVL for the board, all do 3200MHz @ 1.35V.
https://www.msi.com/Motherboard/support/B350I-PRO-AC#support-mem-13
http://www.savewaytech.com/product/avexir-blitz-series-ddr4-163200-dual-channel-8x2-avd4uz132001608g-2bz1gy
https://www.amazon.ca/Corsair-Vengeance-3200MHz-Desktop-Memory/dp/B0143UM4TC


Personally, I think 1.43V is pretty high. IIRC, the max recommended voltage for Ryzen APUs* is 1.25-1.3V for SOC depending on the source.
You're pushing >1.4V through the IMC, part of the chip itself.

If the CPU/GPU aspect isn't recommended for >1.3V, even the 1.35V of rated kits is likely pushing it.


What is your HyperX kit rated at, stock? If it can do 2966-3200MHz @ 1.35V, I'd leave it alone. The performance gains would be (relatively) minimal, and extremely unlikely to be worth the risk... however small.

*Remember, the Ryzen 1X00 series, 2X00 series and the APUs 2X00G/E have different variables. I believe the 1x00 and 2x00 chips are relatively "safe" at up to 1.4V, but the APUs are definitely lower - likely due to the presence of the GPU aspect on the chip too.

 

[Eximo]

That is an interesting question. Since the APUs only have a single CCX are they doing double duty for memory support as compared to the dual CCX Ryzen chips?

 

[Shahmatt]

A lot of it is outright guesswork, with too many variables to list.

Some round-about logic.
1. XMP2.0 (Intel) allows an absolute max of 1.5V for certification, with preference for 1.35V or less.
2. The faster kits (3200MHz) that appear on the QVL for the board, all do 3200MHz @ 1.35V.
https://www.msi.com/Motherboard/support/B350I-PRO-AC#support-mem-13
http://www.savewaytech.com/product/avexir-blitz-series-ddr4-163200-dual-channel-8x2-avd4uz132001608g-2bz1gy
https://www.amazon.ca/Corsair-Vengeance-3200MHz-Desktop-Memory/dp/B0143UM4TC


Personally, I think 1.43V is pretty high. IIRC, the max recommended voltage for Ryzen APUs* is 1.25-1.3V for SOC depending on the source.
You're pushing >1.4V through the IMC, part of the chip itself.

If the CPU/GPU aspect isn't recommended for >1.3V, even the 1.35V of rated kits is likely pushing it.


What is your HyperX kit rated at, stock? If it can do 2966-3200MHz @ 1.35V, I'd leave it alone. The performance gains would be (relatively) minimal, and extremely unlikely to be worth the risk... however small.

*Remember, the Ryzen 1X00 series, 2X00 series and the APUs 2X00G/E have different variables. I believe the 1x00 and 2x00 chips are relatively "safe" at up to 1.4V, but the APUs are definitely lower - likely due to the presence of the GPU aspect on the chip too.

Thanks for your reply. The website and spec sheet say nothing other than 1.2 volts.

https://www.hyperxgaming.com/us/memory/fury-ddr4?partnum=HX426C15FBK2/8
(2666 CL15, 4GB, Kit of 2, Black)

My SOC voltage is set to 1.2 volts and remains there according to HW monitor (I read maximums for this range between 1.2 and 1.25V)

I could lower timings to 3400 CL16 and correspondingly lower voltage to 1.39 volts. However if 1.35 is a hard safe limit I guess I'll have to compromise further.

According to Passmark at least, I do see around a 5% performance gain moving from 3200 CL16 to 3400 CL16, but no advantage moving to 3400 CL14. I suspect this is because my CPU bottlenecks the system clocked at only 3400 MHz (to keep temperatures low in my SFF system).

 

[Karadjgne]

Did you also look up the gains of 3200 C14 vrs 3400 C16? You'd probably have better luck with lower speed, tighter timings, since those APU's don't like to go above 3200. Not to say they won't, just seems like they don't like to without extreme tinkering.

 

[Shahmatt]

Did you also look up the gains of 3200 C14 vrs 3400 C16? You'd probably have better luck with lower speed, tighter timings, since those APU's don't like to go above 3200. Not to say they won't, just seems like they don't like to without extreme tinkering.

3200 CL14 @1.35 Passmark 3D Mark yields 2785
3400 CL16 @1.37 yields a score of 2860 (around 2.7% gain).

So I've lowered my voltages to 3400 CL16 @ 1.37. I'll try to loosen my timings a bit and see if I can bring it down to 1.35.

 

[Karadjgne]

Here's what I found several years ago when I was doing similar things with ddr3, wish I could find the documentation.

A hair =2-3%

1866/10 is a hair faster than 1600/9
1600/8 is a hair faster than 1866/10
2133/11 is 2x hair faster than 1600/9
1600/7 is a hair faster than 2133/10
1600/7 = 1866/8

1600/7 Dominators were 2x the price of my Patriot 1866/10, so for ½ the cost I got equitable performance by tightening up the primary timings.
Granted I run 1866/8 at 1.51v but that's no big deal.

On the other hand, I personally saw exactly no difference at all in anything other than boot. When windows loads, it's got the windows 4 color flower that spirals up. At stock speeds, I see the flower come fully, at 1866/8 it skips the last color. If I was doing something where actual ram performance was critical, maybe I'd push it, but I can't see a 2200G build as being good for any such program. That extra 5% performance benefit is only in a benchmark, IRL it's not anything that important. In gaming you'd be looking at an average of maybe 2-3 fps at best between minimums and maximums which are only boundaries not actual play as fps varies greatly in any given time period.

I can understand the want or need to push a pc as hard as it'll go, but there comes a time when diminishing returns outweigh the the costs of such effort.

 

[Shahmatt]

I suppose it's because this is quite a small build with limitations due to thermals (build image here). I've had to downclock the CPU to 3.4GHz to minimize temperatures. These were also the cheapest 8GB kit I could find on the market here.

So it has been personally interesting to me try to push the memory as far as possible in order to benchmark better and compensate for CPU frequency loss. It is surprising that these can be pushed (seemingly stably) to such high frequencies from the stock 2666. I understand that real world performance may be marginal, but this is for science.

I have discovered that playing TF2 is an excellent litmus test to check RAM stability, more so than running AID64 and Prime 95 blend for 30 mins, Unigine Heaven or Passmark. All of the above may give me a clean bill of health but 5 mins of TF2 and all sorts of errors crop up.

Anyway, I'm trying 3333 CL16, 1.36V now - I've been squeezing the voltages down. TF2 approves. Next stop, 1.35V, followed by 3400 CL16 at slightly looser timings.

Thanks for all your help!

 

[Karadjgne]

Price aside, Kingston ram is not exactly 'cheap' ram. It's actually quite dependable, extremely compatible and performance is right up there with many of the more expensive brands, who often times charge more for a fancier heatsink.

Much depends on the OEM. Especially with Ryzen cpus, which originally were temperamental at best.

 

[Shahmatt]

Word is that the Samsung B die pair very well with the APUs.

My experiment has come to a dead end.

3333 is stable at 1.35V but the Passmark yields a slightly lower score. Perhaps there is instability. So I am keeping it at 1.36V. Though 1.36V is allocated the RAM seems to operate at 1.346V. So acceptable!

3400 CL16 at looser timings are unstable at any voltage less than 1.38. I decided not to go higher. 3400 CL18 (CLs 15 and 17 do not stick) does not pass at looser timings either at 1.36V. So perhaps it is time to call it a day.