RAM performance test by frame time analysis [PCgameshardware.de]

[Carfax83]

Here's a new article from PCgameshardware.de which verifies what I've always known..

RAM capacity, and RAM frequency matter despite what some naysayers might believe. In the test, PCgameshardware.de show some very significant differences in frame time between 4GB/8GB/16GB configurations in BF4..

If they had used the standard benchmark test method for RAM employed by almost every other hardware site which only uses the average frame rate, or if they're generous they include the minimum frame rate as well, it would likely have resulted in NO DIFFERENCES which would have reinforced the lies that have been circulating around the internet regarding RAM capacity/frequency and gaming.

This review by Techspot is one such example, and because the testing methodology is flawed, there is no appreciable difference between the various RAM capacities that were tested.

But by using frame time analysis like what PCgameshardware.de used, we get the full picture.

As for frequency, it's even more salient, with large differences in frame rate (up to 23%) between the fastest and slowest kit tested.

The impact of RAM frequency will always vary depending on the configuration and the setup. Dual core and pure quad core CPUs have the most to gain, as they have less cache and or lack HT, which helps to hide memory latency.

But I personally can attest that even on a high end hexcore rig and very high graphical settings, the impact of RAM frequency is noticeable if you know what to look for.

The Witcher 3 is such an example. The game has limited multithreading, and so shader compiling can cause mild stutters unless you have shader cache turned on in the GPU control panel.

But I've noticed that at DDR4 3200, the game stutters far less than at say DDR4 2666 when entering (Geralt must be running and shader cache must be turned off) Novigrad from Portside gate fast travel sign after a fresh load. What I'm guessing is that the higher frequency RAM allows a quicker copy of those shaders into RAM from storage, and also allows the CPU to do the compiling much faster as well.

 

[npoe1]

I was on the same boat as everyone else until I saw a video from Digital Foundry about performance with different RAM speeds.

I haven't seen buyer's guides and/or RAM testing in Anandtech for a while.

 

[GlacierFreeze]

Interesting. Will have to keep this in mind for my next build in a couple years.

 

[therealnickdanger]

I always install the maximum amount of RAM allowed by my mobo - you can never have too much, but you can always run out.

Mein Deutsch ist a bit rusty, but it appears that they tested the 1x4GB setup against a 2x4GB setup and 2x16GB setup. I would have liked to see them control for that.

 

[plopke]

Don't most people agree with this ?

Every generation has a certain capacity/speed drop off?
Used to be 4GB , 8gb , now 16Gb ,you might as well get 32Gb specially since it is quiet cheap to get DDR4.

For speed it get more complicated but yes if you max out everything then the slowest part namely memory speed is the bottleneck. The higher end your system the more benefit you might get but memory speed is still at bottom of my recommendation list to go crazy on when building a gaming PC , if you max everything then yes it is silly to not go for high speed memory.

Also these higher memory speeds are only available on the higher-end chipset , X99/Z170 so if you on a budget pointless to go for extra speed.

 

[Carfax83]

Mein Deutsch ist a bit rusty, but it appears that they tested the 1x4GB setup against a 2x4GB setup and 2x16GB setup. I would have liked to see them control for that.

I'm pretty sure they didn't. He was talking about the amount of ranks, not channels..

 

[Carfax83]

Also these higher memory speeds are only available on the higher-end chipset , X99/Z170 so if you on a budget pointless to go for extra speed.

Well, the X99 and Z170 are the only two consumer chipsets that support DDR4 I believe.

Z170 is the mainstream chipset, and X99 the high end. Ironically enough, it's the Z170 platform which benefits more from faster RAM compared to the X99, because:

1) Skylake CPUs have a maximum of 8MB of L3 cache vs up to 20MB in Haswell-E CPUs.

2) Skylake CPUs have dual channel memory controllers, vs the quad channel memory controllers in Haswell-E.

Skylake benefits more from fast RAM because it has to go to main memory more often than Haswell-E. But Haswell-E still benefits, just not as much.

 

[Flapdrol1337]

Nice. Didn't know more than 8GB was already useful.

I wonder what memory speed does with frametimes.

 

[plopke]

Well, the X99 and Z170 are the only two consumer chipsets that support DDR4 I believe.

Z170 is the mainstream chipset, and X99 the high end. Ironically enough, it's the Z170 platform which benefits more from faster RAM compared to the X99, because:

1) Skylake CPUs have a maximum of 8MB of L3 cache vs up to 20MB in Haswell-E CPUs.

2) Skylake CPUs have dual channel memory controllers, vs the quad channel memory controllers in Haswell-E.

Skylake benefits more from fast RAM because it has to go to main memory more often than Haswell-E. But Haswell-E still benefits, just not as much.

H170, H110 , B150, Q170,Q150 all come with ddr4 support but I think more strictly locked to 2133. Z170 on paper only supports 2133 but many higher speed in O.C modes which most vendors enable automatically.

But yeah , interesting point on the Haswell vs Skylake.

 

[Carfax83]

Nice. Didn't know more than 8GB was already useful.

I wonder what memory speed does with frametimes.

Likely reduces it, just like memory capacity

People often seem surprised that RAM can have any effect at all on frametime, as we've been so conditioned over the years to believe that faster, or more RAM has no effect due to bad test methodology.

But RAM stores all the important assets like textures, models, maps, animations, audio, shaders, meshes etcetera.. It stores all the data required for the CPU to set up a scene for the GPU to draw, and that's how it would impact frametime latency.

Fallout 4 is the latest game to really benefit from fast RAM, because of how the engine handles data. But it's not the only one. I suspect The Division might benefit from fast memory as well, but I don't have the data to back up that assertion.

The reason why I believe The Division would also benefit, is because it doesn't seem to cache data from what I've seen. The World is so large and detailed, that it seems like it continually loads and reloads assets into RAM.

This is probably done to stop the game from consuming too much RAM, because as I said, it's massive and EXTREMELY detailed.

What made me suspect this, is because I noticed that after I played the game, Windows task manager showed no increase in the amount of cached data. This is the opposite of every other game I have, of which Windows does cache data..

I also noticed a performance increase when my memory is at DDR4 3200 compared to DDR4 2666.

 

[Carfax83]

H170, H110 , B150, Q170,Q150 all come with ddr4 support but I think more strictly locked to 2133. Z170 on paper only supports 2133 but many higher speed in O.C modes..

Thanks for the clarification :thumbsup:

 

[Flapdrol1337]

Doesn't it also depend on the cpu?

On my H81 board my pentium was limited to 1333, but now I put an i5 in and I can run 1600. Can't go higher, so I've lowered the timings instead.

 

[TheELF]

Doesn't it also depend on the cpu?

On my H81 board my pentium was limited to 1333, but now I put an i5 in and I can run 1600. Can't go higher, so I've lowered the timings instead.

Yes it does,but on skylake 2133 is the minimum, even the smallest celeron supports 2133,and with overclocking you can support even faster mem.

 

[coercitiv]

H170, H110 , B150, Q170,Q150 all come with ddr4 support but I think more strictly locked to 2133.

Anything besides Z170 should only support 2133 DDR4 according to Intel, but your mileage may vary depending on MB manufacturer. Some manufacturers enable support for faster memory on the other chipsets as well. For example, take a look at this memory report for an Asus H170 mainboard. The same for H110 and Q170.

The CPU used affects max memory speeds as well, but I do not know if there will be any differences between i3 and the higher tiers.

 

[AlienTech]

On the commodore 64 with 64MB RAM and 1 1MHZ CPU, you had a hard time with bitmaps, the entire windows operating system display is a bitmap.. But For C64 all the characters are translated bitmaps so games use tiles, which replaced characters and hence it worked fine and seemed fast. Kind of like using logical addressing in SSD's, displaying a character would be 8x8 pixels in a bitmap.. So even though it could not manage a real bitmap display the trick of using characters would show a similar result..

Then a game came along called killer bunny and it had a full screen bitmap display that scrolled, th C64 only had hardware to scroll 8 pixels, but moving each 8x8 pixel character up to scroll was impossible. As you had to move 32KB of data to scroll one line (8x8 pixels).. What they did was the had this huge move routine, that moved 8KB each frame and the move routine was like 16K long. It was just a long string of just move, move, move, move, They could not do any branches as it took more time. And the entire memory was filled with this bitmap and the routine to move/scroll it.. No one ever did this trick before or since. But it was a show off to a dumb game and people bought it to see something that was considered impossible. Having slightly faster memory or faster CPU would make this feat a standard.. So instead of wasting 32KB which is half the memory on a move routine which took like less than 100 bytes..

So even something like a single cycle or two makes a huge difference in things. One of the reviews showed a 100% gain in frame rates with memory going from 2133 to 3300mhz.. Not so in other games but you can understand why such a huge difference exist sometimes while at other times it makes no difference. Just missing a hit causing a cascading effect causing a bottle neck can slow everything to a crawl. The same thing with transfer rates from hard drives. For the vast majority of things this is a non event but saying it dont make a difference is valid only in the majority of cases. It does matter.. But only in certain cases..

Also the fastest DDR4 test I saw was on an MSI H110 mother board, liquid cooled and ran at like 4666MHZ.. Although my MSI H110 mother board wont even allow me to run it at 2134MHZ.. Even though the XMP profile says it is 2400MHZ DDR4 RAM. But the biox setting "try it out" ???? it says I can run it at a CL11 profile at 1.37volts.. Which would be equal to like running the RAM at like 3300MHZ or so.. I have not tried it..

 

[Soulkeeper]

This is one reason I like the idea of intel IRIS and AMD going HBM
I've felt for awhile that memory speed is a big part of the future.
As cpu/gpu speed has outpaced memory speed over the years, it's gotten to the point where the cpu/gpu are constantly starved for bandwidth.
Also the big cache size increases have stopped, they no longer double that every other cpu release or so like we used to see.

 

[know of fence]

I always install the maximum amount of RAM allowed by my mobo - you can never have too much, but you can always run out.

Mein Deutsch ist a bit rusty, but it appears that they tested the 1x4GB setup against a 2x4GB setup and 2x16GB setup. I would have liked to see them control for that.

Boom, very good catch. The simple fact is you can't get DDR4 with modules less than 4 GiB in size. The demo is pointless in the first place.
However they tested DDR3. Which means that to test all 4 configs, they'd need a 4 DIMM x 2 GiB DDR3-2166 Kit. It's possible but I find it unlikely, they also only recommend dual channel kits.
In any case PCGH is still trying to sell printed magazines, so this is just an obscure and truncated preview/product-link-advertisement article.

 

[Elixer]

It isn't just the RAM amount & frequency, it also has to do with timings. There are some really big sticks out there with high frequency, but that have crap timing which basically makes the RAM act slower than what it is rated at.