Does DRAM Bandwidth Really Matter?

[Khader87]

hi,

 

i have i7 4790 and 16gigs 1600mhz ram, with vega 64 on 2560x1080 screen

 

im have issues in running new intensive games, many said after troubleshooting that is due to limited dram Bandwidth and that i need better ram which mean going to a whole new rig.

 

is that true? my memory dram bandwith is around 22gigs on dual channel.

 

games like, Odyessey, warhammer 2, BO4, BF5

[xg32]

all of those games are cpu bottlenecked hardcore, not just the ram.

 

6core/12threads or 8 cores minimum (no 8600k/9600k)

[Waqas409]

6 hours ago, xg32 said:

all of those games are cpu bottlenecked hardcore, not just the ram.

 

6core/12threads or 8 cores minimum (no 8600k/9600k)

its funny how people call the straight away "ou ou bottleneck" without actually seeing the GPU and CPU usage in that specific game.

[xg32]

3 hours ago, Waqas409 said:

its funny how people call the straight away "ou ou bottleneck" without actually seeing the GPU and CPU usage in that specific game.

it's also funny how people assume i haven't seen the bottleneck on a 8600k/2600k in those specific games.

 

Sure there could be other issues, but the 4790 is 100% a bottleneck in those games.

[rightmeow]

Bullshit, 4790 is not a bottle neck. Those games run on counsel, and that has a potato for a cpu. Dram frequency makes no difference. There's nothing in the world that reads and writes 35 Gigabytes a Second to the ram. What issues are you running into? Do you have the newest drivers? Doubt it's anything hardware related. 

[MageTank]

6 minutes ago, rightmeow said:

Bullshit, 4790 is not a bottle neck. Those games run on counsel, and that has a potato for a cpu. Dram frequency makes no difference. There's nothing in the world that reads and writes 35 Gigabytes a Second to the ram. What issues are you running into? Do you have the newest drivers? Doubt it's anything hardware related. 

There is an awful lot that is wrong about this statement. DRAM frequency not making a difference is part of it, but the "nothing in the world that reads/writes 35 gigabytes a second to the ram" is outright wrong. Plenty of applications make full use of that bandwidth, and specific instruction sets such as AVX (and their derived variants) still require excessive amounts of memory bandwidth to avoid bottlenecks. You can see this when running Linpack, especially the MKL flavor. 

 

As for the original statement of DRAM frequency making no difference, this is simply not true. There are plenty of published benchmarks both from forum users (such as myself) as well as those in the tech review industry that have demonstrated the impact of memory bandwidth/latency on the overall gaming experience. 

 

Here is a post from 2014 on OCN that first kicked things off: https://www.overclock.net/forum/18051-memory/1487162-independent-study-does-speed-ram-directly-affect-fps-during-high-cpu-overhead-scenarios.html

 

Next, we have the Digital Foundry benchmarks that are still shown to this day as being extremely impactful:

https://www.eurogamer.net/articles/digitalfoundry-2018-intel-core-i9-9900k-review

https://www.eurogamer.net/articles/digitalfoundry-2017-intel-coffee-lake-core-i7-8700k-review_1

https://www.eurogamer.net/articles/digitalfoundry-2017-intel-kaby-lake-core-i7-7700k-review

The trend continues as you go back to older generations as well.

 

If you prefer GN's article, they saw a 5-7% performance boost across all titles going from 2666 to 3200: https://www.gamersnexus.net/hwreviews/3086-intel-i5-8400-cpu-review-2666mhz-vs-3200mhz-gaming/page-4

 

The point is, memory speed does have an impact on gaming performance, especially when it comes to minimum framerates. It's not as dramatic as say, a CPU or GPU upgrade, but if your platform supports it, memory OCing is free, and should be considered by those that intend to get the most out of their hardware. I won't pretend to know why it matters, as I don't think it's related to frequency (my personal tests showed scaling improve as RTL or Round Trip Latency decreased, which frequency only plays a small part in RTL) but some have questioned whether or not it's related to an application being bottlenecked by CPU I/O requests. Either way, there is a pretty consistent difference in performance, and in my personal tests, it's done wonders for my minimum framerates, with some titles showing 20% higher FPS on absolute minimums. 

 

I am not biased at all, but there is a memory overclocking guide in my signature, cough cough.

 

If you need additional sources, let me know. I have an entire folder full when I get off work.

[rightmeow]

@MageTank 5% to 7% is nothing! Practically irrelevant. Why did intel stop supporting quad channel memory? Even though it's twice the bandwidth? Because the performance gain is in the single digitals even in the best niche cases. Shit I wouldn't be surprised if 900MHZ Single Channel is 90% of the best quad channel 3200MHZ chips. All that work for so little gain. 

[MageTank]

23 hours ago, rightmeow said:

@MageTank 5% to 7% is nothing! Practically irrelevant. Why did intel stop supporting quad channel memory? Even though it's twice the bandwidth? Because the performance gain is in the single digitals even in the best niche cases. Shit I wouldn't be surprised if 900MHZ Single Channel is 90% of the best quad channel 3200MHZ chips. All that work for so little gain. 

Intel never stopped supporting quad channel memory. X299 supports it as we speak, and LGA 3647 supports Hexa-channel memory. I have no idea where you are getting your facts from, but they are far from true. With your logic, JEDEC shouldn't bother raising frequency with each memory standard that gets released. Also, people do far more to gain 5-7% performance, let's not paint with that broad of a brush. My personal tests saw a 20% increase in minimum framerates, far more than what most get out of other forms of overclocking.

 

I've offered additional sources if you wanted to see them. At this point, it's your turn to prove otherwise with some sources of your own. Hopefully ones with a testing methodology that we can recreate and test for ourselves. 

[Khader87]

i got i5 8500 and 16GB ddr3 2666mhz 8x2 and games run like butter now. seems like the 4790 or the ram were the bottleneck

[Code.016]

9 hours ago, MageTank said:

Intel never stopped supporting quad channel memory. X299 supports it as we speak, and LGA 3647 supports Hexa-channel memory. I have no idea where you are getting your facts from, but they are far from true. With your logic, JEDEC shouldn't bother raising frequency with each memory standard that gets released. Also, people do far more to gain 5-7% performance, let's not paint with that broad of a brush. My personal tests saw a 20% increase in minimum framerates, far more than what most get out of other forms of overclocking.

 

I've offered additional sources if you wanted to see them. At this point, it's your turn to prove otherwise with some sources of your own. Hopefully ones with a testing methodology that we can recreate and test for ourselves. 

that guy might be using an apple desktop. but seriously, this is the first time i saw a guy that doesn't believe that DRAM freqs is relevant, just like those people believe that the earth is flat.