Ryzen RAM speeds a myth?

[Mark Kaine]

I suspected this for a while - also because i saw *zero* (thats a 0) difference between 8GB 2933 and 16GB 3200¹ (only diff was my wallet lol) ¹Tested with R3600, GTX 1060

 

So now i tested it again this time with Ryzen 2200g + GTX 1060 + 16GB Corsair Venegance "3200"

 

Superposition scores 16GB :

Stock 2133mhz 

6334 points

XMP 2933mhz

6343 points

 

So conclusion is its absolutely useless to buy "faster" RAM (for gaming)

 

If you disagree please explain, and do not mention GN as he always heavily tweaks RAM and therefore is irrelevant to this discussion. This isnt about overclocking, this is about what you get if buy expensive, "fast" RAM.

 

 

And yes, i might be wrong, but my tests point to this being rather right, unless my motherboard is defictive and shows "fake values" which i cannot verify.

 

PS: and I would actually appreciate if someone can explain to me why there is zero (well 9 points which is within margin of error obviously) difference?  

 

 

 

 

[Levent]

wat. You mean to tell me everyone is wrong and you are correct? Your testing is flawed.

[Moonzy]

7 minutes ago, Mark Kaine said:

Ryzen 2200g + GTX 1060 + 16GB Corsair Venegance "3200"

i think... you might need a better system to truly see the diff...?

though i may be wrong

 

it's mostly prevalent in CPU bound scenarios i would think

[Mark Kaine]

Just now, Levent said:

wat. You mean to tell me everyone is wrong and you are correct? Your testing is flawed.

So did you test it yourself?

Flawed in which way?

[DoctorNick]

Where are the results?  Edit: Superposition isn't great for testing memory.. Is uses almost no CPU

[Levent]

1 minute ago, Mark Kaine said:

So did you test it yourself?

Flawed in which way?

Yes, that is why my 2133mhz CL14 rams are running at 3200 CL20. Even that gives me a performance bump, particularly noticeable in RDR2. Flaw is quite obvious:

1 minute ago, Moonzy said:

it's mostly prevalent in CPU bound scenarios

Not to mention you tested a SINGLE workload.

[Mark Kaine]

Just now, Moonzy said:

i think... you might need a better system to truly see the diff...?

though i may be wrong

 

it's mostly prevalent in CPU bound scenarios i would think

Yes, you might be wrong because that was a big advertising point even back then "Ryzen needs fast RAM"

 

And I would even say the iGPU which I didnt use for this test is irrelevant because the same pitch was (and still is) made for Ryzen CPUs without an iGPU.

 

Now Im not saying its all wrong, Im rather asking for a good explanation how its possible i see literally no difference.

[Moonzy]

Just now, Mark Kaine said:

Im rather asking for a good explanation how its possible i see literally no difference.

isnt superposition more of a GPU benchmark though?

 

that said, some CPU benchmark such as cinebench dont get affected by RAM much either from what i heard

 

best test in CPU heavy games, off the top of my head... MHW is a good candidate though it's not repeatable

maybe some game with built in benchmark, set the graphical settings to low so you'll be CPU bound and test it that way? SOTR for example

[Mark Kaine]

9 minutes ago, Levent said:

Not to mention you tested a SINGLE workload

True, I did test a lot more last time i did this though and the results were the same, not 1FPS more or even a better score in cinebench.

 

Also since this CPU is pretty weak, its already being used like 90% in Superposition, so in theory there should be improvements?

 

 

6 minutes ago, Moonzy said:

isnt superposition more of a GPU benchmark though?

 

that said, some CPU benchmark such as cinebench dont get affected by RAM much either from what i heard

 

best test in CPU heavy games, off the top of my head... MHW is a good candidate though it's not repeatable

maybe some game with built in benchmark, set the graphical settings to low so you'll be CPU bound and test it that way? SOTR for example

Ya, but see above its already using ~90%, and those are good points, it maybe depends on many factors, but that also means fast RAM isnt *that* big of a deal as its made out to be (only in certain scenarios, and if, if, if... lol)

[Levent]

1 minute ago, Mark Kaine said:

True, I did test a lot more last time i did this and the results were the same, not 1FPS more or even a better scorecin cinebench.

 

Also since this CPU is pretty weak, its already being used like 90% in Superposition, so in theory there should be improvements?

If CPU isnt being maxed out then your GPU is your bottleneck and you are trying to test your CPU. It doesnt work that way.

[Mark Kaine]

2 minutes ago, Levent said:

If CPU isnt being maxed out then your GPU is your bottleneck and you are trying to test your CPU. It doesnt work that way.

Nope Im testing game performance, your cpu should never be maxed out while gaming.  You just proofed my theory, wow!

[Moonzy]

14 minutes ago, Mark Kaine said:

fast RAM isnt *that* big of a deal

it isnt

that's why im using 2666 CL16 on my 3900x system because it's half the price of a 3600 CL18

at best im losing... 10-15%, and only in CPU bound games, by then my fps is probably already in the 100s, so -shrug-

 

but i would say it's wrong to say it doesnt affect things at all

[mariushm]

15 minutes ago, Mark Kaine said:

I suspected this for a while - also because i saw *zero* (thats a 0) difference between 8GB 2933 and 16GB 3200¹ (only diff was my wallet lol) ¹Tested with R3600, GTX 1060

 

So now i tested it again this time with Ryzen 2200g + GTX 1060 + 16GB Corsair Venegance "3200"

 

Superposition scores 16GB :

Stock 2133mhz 

6334 points

XMP 2933mhz

6343 points

 

So conclusion is its absolutely useless to buy "faster" RAM (for gaming)

 

There's performance difference between SINGLE CHANNEL ( one stick of ram)  and DUAL CHANNEL (two sticks , or four sticks )

 

There's performance difference (but very small) between SINGLE RANK and DUAL RANK memory sticks. 

If you have 2 x 4 GB sticks, those were most likely SINGLE RANK.  The 2 x 8 GB (if that was the case) sticks were most likely DUAL RANK sticks, which should give a very tiny performance increase. 

 

Memory frequency matters A LOT for APUs, as it increases the performance of the integrated graphics.  As you were using the dedicated graphics card, it matters less. 

The combination is also bad because your 2200g has only 8 pci-e lanes going to the pci-e x16 slot, so you're artificially limiting the maximum bandwidth of your video card to 8 pci-e 3.0 lanes (though in real world it matters very little)

 

The benchmark was badly chosen, as it doesn't use the CPU a lot, and doesn't work with memory a lot.

  

Choose a benchmark which deals with lots of small computations, like.... some database benchmark  or some file compression benchmark, even an application that renders a video may show a performance difference.

Get 7zip and use the built in benchmark and see how fast it can compress and decompress some things.  You can argue you don't care about compression, because you care only about games, but guess what... a lot of games compress content, and decompress content during game level loads, so you may get slightly faster load times.

 

 

 

 

 

 

[mahyar]

Just now, Mark Kaine said:

Nope Im testing game performance, your cpu should never be maxed out while gaming.  You just proofed my theory, wow!

"game performance" with ram speed is dependent on the engine

so to prove that you are right you have to test 10-15 games with different engines and different version of an engine

[Chris Pratt]

20 minutes ago, Mark Kaine said:

I suspected this for a while - also because i saw *zero* (thats a 0) difference between 8GB 2933 and 16GB 3200¹ (only diff was my wallet lol) ¹Tested with R3600, GTX 1060

 

So now i tested it again this time with Ryzen 2200g + GTX 1060 + 16GB Corsair Venegance "3200"

 

Superposition scores 16GB :

Stock 2133mhz 

6334 points

XMP 2933mhz

6343 points

 

So conclusion is its absolutely useless to buy "faster" RAM (for gaming)

 

If you disagree please explain, and do not mention GN as he always heavily tweaks RAM and therefore is irrelevant to this discussion. This isnt about overclocking, this is about what you get if buy expensive, "fast" RAM.

 

 

And yes, i might be wrong, but my tests point to this being rather right, unless my motherboard is defictive and shows "fake values" which i cannot verify.

 

PS: and I would actually appreciate if someone can explain to me why there is zero (well 9 points which is within margin of error obviously) difference?  

 

 

 

 

You're 100% wrong. I'm not sure what you did in your testing that caused erroneous results, but this has already been tested *to death* by people, no offense, far more qualified than you, and there is definitely a difference between RAM speeds on Ryzen.

 

Not only that, but it's just logically obvious as well. Unlike Intel, Ryzen has its infinity fabric, which is the communication layer between the CCXs. This must run in a lock step 1:1 ratio with the RAM, so the slower the RAM, the slower the infinity clock, and the slower everything the CPU does is.

 

There's a point of diminishing returns. The FCLK on Zen 2 tops out around 1800MHz stock, for example. Depending on the particular part, you can sometimes overclock it higher. However, as a result, you'll see mixed results with speeds faster than 3600Mhz, and if you go higher than the FCLK can match, it actually *hurts* performance, as it has to switch to a 2:1 ratio, effectively halving the performance of the RAM.

[Levent]

4 minutes ago, Mark Kaine said:

Nope Im testing game performance, your cpu should never be maxed out while gaming.  You just proofed my theory, wow!

That is the second dumbest statement you made this topic so far. Unless there is a bottleneck from CPU/RAM or IF you will not benefit from higher clock RAMs. Somehow you are not able to understand the concept of bottleneck or anyone else here.

[Moonzy]

2 minutes ago, Chris Pratt said:

This must run in a lock step 1:1 ratio with the RAM

not necessarily, I run 2666mhz and still crank FCLK

 

 

this would give me some of the performance from using slower RAM, but of course not all of it

[Mark Kaine]

14 minutes ago, Chris Pratt said:

You're 100% wrong. I'm not sure what you did in your testing that caused erroneous results, but this has already been tested *to death* by people, no offense, far more qualified than you, and there is definitely a difference between RAM speeds on Ryzen.

 

Not only that, but it's just logically obvious as well. Unlike Intel, Ryzen has its infinity fabric, which is the communication layer between the CCXs. This must run in a lock step 1:1 ratio with the RAM, so the slower the RAM, the slower the infinity clock, and the slower everything the CPU does is.

 

There's a point of diminishing returns. The FCLK on Zen 2 tops out around 1800MHz stock, for example. Depending on the particular part, you can sometimes overclock it higher. However, as a result, you'll see mixed results with speeds faster than 3600Mhz, and if you go higher than the FCLK can match, it actually *hurts* performance, as it has to switch to a 2:1 ratio, effectively halving the performance of the RAM.

Oh yeah, I get the theory, I just dont.see the results, I even tried different IFs, zero difference (within margin of error)

 

I mean its possible motherboards can make a difference too, but so far everything points to this being made a far bigger deal than it really is, especially by so called "influencers". 

[Chris Pratt]

10 minutes ago, Moonzy said:

not necessarily, I run 2666mhz and still crank FCLK

 

 

this would give me some of the performance from using slower RAM, but of course not all of it

Then you're running in asynchronous mode and actually hindering your performance trying to crank it higher than the RAM. I guess I should have been less definitive. It doesn't *have* to run lock step, but it makes absolutely no sense to do it any other way.

[Chris Pratt]

Just now, Mark Kaine said:

Oh yeah, I get the theory, I just dont.see the results, I even tried different IFs, zero difference (within margin of error)

 

I mean its possible motherboards can make a difference too, but so far everything points to this being made a far bigger deal than it really is, especially by so called "influencers". 

The board does matter, as there's timings that aren't included in XMP profiles, and still further timings motherboards don't let you tweak at all. As a result, board to board, you'll see variations, depending on how well each is optimized overall.

 

However, you have a sample size of 2 here, which from a statistical standpoint means exactly jack. We have no real info on your RAM, and things like timings or rank can skew the results. In your particular limited testing, there may be no difference, but that says absolutely nothing about anything of value to the rest of the planet.

[Mark Kaine]

1 hour ago, Chris Pratt said:

The board does matter, as there's timings that aren't included in XMP profiles, and still further timings motherboards don't let you tweak at all. As a result, board to board, you'll see variations, depending on how well each is optimized overall.

 

However, you have a sample size of 2 here, which from a statistical standpoint means exactly jack. We have no real info on your RAM, and things like timings or rank can skew the results. In your particular limited testing, there may be no difference, but that says absolutely nothing about anything of value to the rest of the planet.

Well, I disagree about that, I think Im a pretty good representation of the  "average gamer" and as such would be interested in knowing if Im falling for a hoax or not (which this increasingly looks like)

 

And I already said I understand the theory , but Im just seeing zero evidencece of it being true in my testing, which was pretty extensive the first time around , I tested like 10 games and several benchmarks there wasnt any tangible difference between 2933 and 3200.

 

Your argument basically boils down  to im measuring "wrong" thats not very convincing.

 

1 hour ago, Chris Pratt said:

You're 100% wrong.

I dont think so?

 

2133

 

2933 

 

 

Btw 98% CPU usage

 

I mean I would want to try with a different board, maybe mine is "faulty" somehow, but I have to do with what I have. Also its totally possible there are some exceptions but overall it just doesnt look like it matters much, or at all.

 

Also Ive noticed if I dont set XMP the IF downclocks automatically, thats very nice and I think I'll just leave it like that for now since theres no difference in performance, but surely in energy consumption and heat. 

 

[Chris Pratt]

1 minute ago, Mark Kaine said:

Well, I disagree about that, I think Im a pretty good representationof the  "average gamer" and as such would be interested in knowing if Im falling for a hoax or not (which this increasingly looks like)

 

And I already said I understand the theory , Im seeing zero evidencecof it being true in my testing, which was pretty extensive the first time around , I tested like 10 games and several benchmarks there wasnt any tangible difference between 2933 and 3200.

 

Your argument basically boils down  to im measuring "wrong" thats not very convincing.

 

I dont think so?

 

2133

 

2933 

 

 

Btw 98% CPU usage

 

I mean I would want to try with a different board, maybe mine is "faulty" somehow, but I have to do with what I have. Also its totally possible there are some exceptions but overall it just doesnt look like it matters much, or at all.

 

Also Ive noticed if I dont set XMP the IF downclocks automatically, thats very nice and I think I'll just leave it like that for now since theres no difference in performance, but surely in energy consumption and heat. 

 

My point was that 1) your testing is entirely unscientific (too many variables you're not controlling for) and 2) not statistically significant (extremely limited data set).

 

It's like saying the sun doesn't exist, because you went outside one night and it was dark.

[Mark Kaine]

17 minutes ago, Chris Pratt said:

1) your testing is entirely unscientific (too many variables you're not controlling for)

Oh, so a youtube show with live chat and merchandise is more "scientific" ? I also didnt claim what im doing is scientific, more like "wait a minute... why am I seeing zero difference when everybody and their subscribers says there *is* a difference?"

17 minutes ago, Chris Pratt said:

2) not statistically significant (extremely limited data set).

At what point is it relevant then? Ive already tested 10+ games (I do remember the first time well, it was two back to back runs of the same games/benchmarks over the course of two days, and I didnt want to test any more games and benchmarks because  the results were pretty clear cut at this point) And its relevant to me, either the whole faster RAM = more FPS thing is simply not true, or something is wrong with my motherboard - I dont think its my testing per say as setting  XMP and reading some numbers isnt exactly rocket surgery, its quite difficult to mess that up...

 

 

[AMD A10-9600P]

You're using a 3200G in those screenshots, the limit there isn't the RAM but the CPU itself, as it is comparatively very weak. I would guess as well that you're using a lower end motherboard.

 

There is definitely a performance boost to be had by using higher frequency memory, not just on Ryzen, but also on Intel. This performance benefit will of course vary based on the task, hence why in the one task you have presented you see no benefit, but I can guarantee that is not the case for all tasks, otherwise people wouldn't be so convinced about recommending higher frequency memory.

 

Your testing is not testing, it is just "I saw this once, in one situation, so it's right", which ain't how scientific testing works. Proper testing takes into account all the different variables that may affect the performance, across different scenarios and situations. Come back after testing in 20 different games, with 20 different engines, whilst keeping all other variables controlled, then I may consider believing you. As for now, what you have said is not true, and higher frequency memory does offer benefits.

 

Even on my Intel desktop I see benefits using XMP to get my RAM up to 3200MHz, versus using it at 2666. My Intel laptop uses 4200MHz memory, which gives huge benefits in iGPU loads, as this brings the system memory up to speeds that are similar to what a dedicated mobile GPU (like the MX450) can offer. I can imagine that if you used the iGPU on your 3200G, you would see performance gains using faster memory.

[WhitetailAni]

Try running Minecraft Java. It's HUGELY CPU bound and when I had my Ryzen 5 3600 underclocked to 3.4 GHz (to avoid overheating) vs. when it was overclocked to 4.3 GHz there was a huge difference. At 4.3, I could easily get 600+.

At 3.4, I could barely crack 250.

For fun, I decided to set my RAM to 2133 and test both CPU clockspeeds.

2133 + 3.4 = ~140 FPS

2133 + 4.3 = ~300 FPS

3200 + 3.4 = ~450 FPS

3200 + 4.3 = ~800 FPS

 

This is not definitive in any way, shape or form. This is one CPU and one kit of memory, and my graphics card is very old (GTX 650 Ti BOOST, then it was on stock).

 

But perhaps you could try. Java has a free trial option last I checked.