Is it better to have more amount memory or a higher speed of memory?

[EagleEyes6801]

I am wanting to make my own gaming desktop with DDR4 memory but can't decide on having more amount of memory or a higher speed of memory. HELP!!!

[Guest]

Memory speed don't do much, so just buy the amount you need

[Centurius]

Most of the time a higher amount beats a higher speed but as with all things, context is key. If your apps never use more than 16 Gigs of RAM getting 64 is insane, etc.

[Sakkura]

Memory speed is more important than capacity, once you have enough for the things you do.

[typographie]

It depends on what you're doing with the PC, as different types of applications benefit more or less from memory than others.

 

If it's a gaming PC, I'd say go for more memory up to about 16 GB and then spend the rest on improving something else in your system. I don't think there's much need to go beyond DDR4-2666 or so, which is still quite cheap.

[i_build_nanosuits]

what are you gonna do with the machine? what are your most intensive tasks you plan to do?!

[Sakkura]

3 minutes ago, Star Paladin Vinyl said:

Memory speed don't do much

Uhh...

 

Spoiler

 

[IAcKI]

If you're planning on gaming just go for 8-16gb of 2133MHz or higher and you're golden. 

[Vishal Gupta]

Both. But depends on what you want to do. Like for gaming, 16gb ddr3/ddr4 is more than enough.

 

[IAcKI]

1 minute ago, Sakkura said:

Uhh...

 

  Hide contents

 

Jesus christ. They're huge differences! 
Guessing it's due to it being quite a CPU bound game?

[Sakkura]

1 minute ago, IAcKI said:

Jesus christ. They're huge differences! 
Guessing it's due to it being quite a CPU bound game?

Yeah. It is an unusual scaling, but there are examples of memory speed becoming more important than the conventional wisdom implies (and which used to be pretty accurate). It's also worth mentioning that you wouldn't see the same improvement between 2400 and 3466 - there are diminishing returns.

[ybriK]

3 minutes ago, IAcKI said:

Jesus christ. They're huge differences! 
Guessing it's due to it being quite a CPU bound game?

Games like GTA V, Witcher 3 and Far Cry 4 love fast memory. Anything above 2133MHz gives significant performance increase however it will eventually hit a stopping point (or a bottleneck which is the game, not sure if that's the right term) which is around 3000-3200MHz and above, then very minimal performance will be gained after those speeds. Lots of benchmarks online, too lazy to link.

[IAcKI]

Just now, Sakkura said:

Yeah. It is an unusual scaling, but there are examples of memory speed becoming more important than the conventional wisdom implies (and which used to be pretty accurate). It's also worth mentioning that you wouldn't see the same improvement between 2400 and 3466 - there are diminishing returns.

Yeah i guessed so. Be less apparent now I would guess with DDR4 becoming standard and I don't believe I've ever seen lower than 2133MHz with the increase in price being on small up to 2600MHz. I know the 2866MHz Dominator was actually more expensive than 3000MHz Dominator.

2 minutes ago, ybriK said:

Games like GTA V, Witcher 3 and Far Cry 4 love fast memory. Anything above 2133MHz gives significant performance increase however it will eventually hit a bottleneck which is around 3000-3200MHz and above, then very minimal performance will be gained after those speeds. Lots of benchmarks online, too lazy to link.

Is this all still the case with DDR4? Very interested now as I've recommended a friend get some 2600MHz Ram as it was quite abit cheaper than 3000MHz at the time.  

[ybriK]

Should still be, it's noticeable in games that are CPU bound. Meaning significant utilization across all the logical cores of any given CPU. Even though I only have DDR3 1600MHz I did a test between 1066 vs 1600 on Fallout 4 and I saw massive performance drops on 1066MHz.

[IAcKI]

1 minute ago, ybriK said:

Should still be, it's noticeable in games that are CPU bound. Meaning significant utilization across all the logical cores of any given CPU.

Might give it a few benchmarks my self with my Ram at 2133 and also at 3000 to see what the performance differences are.

[Centurius]

51 minutes ago, Sakkura said:

Uhh...

 

  Hide contents

 

You're comparing an absolute crazy difference in speeds on a different memory technology with a very unique game. 

 

 

[Snadzies]

Having enough is far more important that having faster memory.

Usually if you don't have enough then any extra data that normally would be held in RAM goes to the HDD/SSD and accessing the data from the HDD/SSD is much, much, slower.

Alternatively if the data doesn't go to the HDD/SSD then the data has to be compressed and decompress to fit in the allocated amount of RAM which also takes up a fair bit more time and resources compared to having enough RAM in the first place.

 

Once you have enough then having more does you next to no good and faster RAM becomes far more important.

[Sakkura]

55 minutes ago, Centurius said:

You're comparing an absolute crazy difference in speeds on a different memory technology with a very unique game. 

 

 

Hidden Content

 

I'm comparing on a couple of older memory controllers, meaning the impact of memory speed will be underestimated compared to DDR4.

 

You're relying on benchmarks of GPU-bound games. And from some random unreliable site.

[typographie]

2 hours ago, Centurius said:

You're comparing an absolute crazy difference in speeds on a different memory technology with a very unique game. 

Well, if you want to detect differences in memory performance you need to look at platform-bound games like Fallout. It may be a relatively rare example, but conversely most games are not as far to the other end of the scale as Crysis 3 or The Witcher 3, either.

[MageTank]

3 hours ago, Sakkura said:

Yeah. It is an unusual scaling, but there are examples of memory speed becoming more important than the conventional wisdom implies (and which used to be pretty accurate). It's also worth mentioning that you wouldn't see the same improvement between 2400 and 3466 - there are diminishing returns.

Very close, but some slight inaccuracies. The diminishing returns aspect depends entirely on how much CPU overhead is involved. There will be more CPU overhead if you had say, a GTX Titan X, when compared to say, a GTX 770. Meaning at around 2800mhz, you won't really notice a difference in performance on the GTX 770 build. However, the Titan X might see scaling all the way up to 3200mhz. SLI/Crossfire configurations see even more improvement with memory bandwidth. However (and I will reveal more about this in a guide on this forum later) memory latency is far more important in SLI/Crossfire configurations than it is in single GPU configurations. So one will have to find a great balance between raw bandwidth for the sake of bandwidth (quad channel) and bandwidth + latency (high clocked dual channel). 

 

This begs the question of whether memory bandwidth will matter in DX12 games (I do not own any DX12 games, so I cannot test) as DX12 plans to remedy the CPU overhead situation. If such is the case, raw capacity is a better option. However, seeing as the vast majority of games are not DX12, speed (with at least enough capacity to run your games) is the better option. I would personally invest in 16GB minimum these days, with at least 8GB sticks for dual rank (for Rank Interleaving) to get the absolute best performance out of your system. 16GB should be enough for any game, while also offering room for upgrades, and again, rank interleaving. Free bandwidth is free bandwidth.

 

I will be working on an all encompassing memory guide for this forum soon, whenever i have enough time off work. Going to be doing it on Haswell and Skylake i5's, with varying amounts of graphics cards (from a 750 Ti up to say, a GTX 970 for now, until Pascal Titans come out). I will test memory speeds from 1333mhz DDR3 to 2400mhz DDR3, and DDR4 2133mhz to DDR4 3400mhz DDR4. If anyone has any games they want me to test, feel free to PM me so I can catch them on sale before i start testing. Hopefully once I am finished, we can finally say without a doubt that memory speed is important (unless DX12 changes that). 

[Sakkura]

7 minutes ago, MageTank said:

Very close, but some slight inaccuracies. The diminishing returns aspect depends entirely on how much CPU overhead is involved. There will be more CPU overhead if you had say, a GTX Titan X, when compared to say, a GTX 770. Meaning at around 2800mhz, you won't really notice a difference in performance on the GTX 770 build. However, the Titan X might see scaling all the way up to 3200mhz. SLI/Crossfire configurations see even more improvement with memory bandwidth. However (and I will reveal more about this in a guide on this forum later) memory latency is far more important in SLI/Crossfire configurations than it is in single GPU configurations. So one will have to find a great balance between raw bandwidth for the sake of bandwidth (quad channel) and bandwidth + latency (high clocked dual channel).

With the same driver, the CPU overhead is the same. It sounds like you're talking about bottlenecking. A more powerful GPU will relieve GPU bottlenecking in more situations, shifting the bottleneck towards CPU and/or system memory. So yeah, the bottom line results should line up as you indicate. But the principle of diminishing returns of faster memory (on a given system) holds.

[MageTank]

20 minutes ago, Sakkura said:

With the same driver, the CPU overhead is the same. It sounds like you're talking about bottlenecking. A more powerful GPU will relieve GPU bottlenecking in more situations, shifting the bottleneck towards CPU and/or system memory. So yeah, the bottom line results should line up as you indicate. But the principle of diminishing returns of faster memory (on a given system) holds.

You are completely right. I was confusing overhead with the amount of draw calls a card can handle from a CPU. With a faster card being able to handle more draw calls than the slower card. All I know for certain, is that faster cards seem to care more about faster memory than slower cards. For slower cards (my GTX 770 for example) gained no extra FPS from 2800mhz memory after showing a steady increase at 2400 and 2666mhz. Faster cards, like the Titan X on the other hand, showed an improvement at every single step, from 2400, 2666, 2800, 3000, 3200, etc. 

 

To put it in much more simple terms, my theory is this: Faster cards require faster resources to saturate them. You could be right, in that the bottleneck is simply changing from one component to another, but I need to sufficiently test that for me to proclaim that as fact. As i said, i really don't know the "why", i just know the "what", lol. 

[EagleEyes6801]

I'm want to make a computer that supports 4 people all on at the same time all on the same tower, the ram I was going for was 16gb of 4266mhz with a i7-6700k processor and an Asus Maximus VIII Extreme motherboard, and 4 Asus ROG Matrix GeForce 980's.

[southpaw5271]

On 3/8/2016 at 10:14 AM, Vishal Gupta said:

Both. But depends on what you want to do. Like for gaming, 16gb ddr3/ddr4 is more than enough.

 

Here here

[MageTank]

32 minutes ago, EagleEyes6801 said:

I'm want to make a computer that supports 4 people all on at the same time all on the same tower, the ram I was going for was 16gb of 4266mhz with a i7-6700k processor and an Asus Maximus VIII Extreme motherboard, and 4 Asus ROG Matrix GeForce 980's.

I highly doubt your IMC will handle all 4 dimms running at that speed. In fact, I am willing to bet it doesn't. Your board does not even list memory on its QVL list beyond 3800mhz, and that says a lot about the quality of its traces. 

 

Not only that, the memory speed you are wanting is worthless. You obtain that speed, you are forced to use single rank 4GB sticks, meaning you lose rank interleaving. Allowing 4 people to play at once, means you are going to be giving each client 3GB of ram (Remember, the host machine is going to need ram too, plus you are giving out ram to each hypervisor client) which means you are going to end up swapping to your drives, completely killing all performance benefits of memory. That's before we even get into what it takes to hit 4266mhz on those sticks. They sacrifice RTL's and tertiary timings to reach those speeds. You would have more memory bandwidth using 8GB sticks at 3800mhz with tight RTL's and tertiary timings, i assure you of that. At 3500 CL14-14-14-28, I am actually beating a 4000mhz XMP profile as we speak. 

 

In short, what you are trying to do with that ram is a very bad idea. Go with "slower" memory, in a higher capacity. Combine the automatic boost in bandwidth from Rank Interleaving with the extra performance gained from tighter RTL's and tertiary timings, and you will end up with a faster setup and more memory to spare.