A question about quad channel and memory speed in relation to game performance

[Mirkoskji]

Hello everyone,

I'll try to reword the question.

Taking for example Fallout 4, which show incredible performance gains when played on a system with high memory speed, is the same to run memory with half the speed in a quad channel configuration than running it on a different system that has only dual channel memory config but double the speed?

Is 4 channel ddr-x 1600 = 2 channel ddr-x 3200, considering this specific application (games) ?

 

[eidairaman1]

Hello everyone,

I'll try to reword the question.

Taking for example Fallout 4, which show incredible performance gains when played on a system with high memory speed, is the same to run memory with half the speed in a quad channel configuration than running it on a different system that has only dual channel memory config but double the speed?

Is 4 channel ddr-x 1600 = 2 channel ddr-x 3200, considering this specific application (games) ?

It's motherboard dependent, if your motherboard doesn't have quad channeling you wont see the effects of it. Back during DDR1 Days, the fastest Jedec was PC 3200 Ram. Theoretically 3200MB/s in single channel, when in dual channel config it would go up to 6400MB/, those are just maximum transfer rates, and software only use what's needed, hence why Dual channeling is still standard where multi channeling are for servers etc.

https://en.m.wikipedia.org/wiki/Multi-channel_memory_architecture

https://en.m.wikipedia.org/wiki/DDR_SDRAM

 

[TheoneandonlyMrK]

Well it depends more on the platform these days since cpus have the memory controller built in and some are better than others, especially at the minute.
In general going from 2133 16Gb with two sticks to four @1333 32Gb on my fx dual channel platform i get less bandwidth by about 20% but much lower latency like 50% lower and more memory.
The os and applications are faster but 3d game's are definitely slower but not drastically , there's not much fps difference in reality(2-6fps) but it can be exaggerated by particular game engine.

 

[NdMk2o1o]

Well it depends more on the platform these days since cpus have the memory controller built in and some are better than others, especially at the minute.
In general going from 2133 16Gb with two sticks to four @1333 32Gb on my fx dual channel platform i get less bandwidth by about 20% but much lower latency like 50% lower and more memory.
The os and applications are faster but 3d game's are definitely slower but not drastically , there's not much fps difference in reality(2-6fps) but it can be exaggerated by particular game engine.

Going from 2 sticks to 4 doesnt equate to going from dual channel to quad channel though which is what the OP is pertaining to

 

[TheoneandonlyMrK]

Going from 2 sticks to 4 doesnt equate to going from dual channel to quad channel though which is what the OP is pertaining to

Fair point , i kinda drifted off halfway through i think ,utd won , anyway.
It's sort of was covered i said all platforms differ, intel had the better imc , maybe still do but that could change soon.
Obviously with four you still end up running them slower than two because inevitabley they are attached to a buss that has its limited performance potential too ,and the imc being larger in resource terms generates a lot of heat and uses a lot of power.
So can't be run as high without serious cooling, I think with games its All about bandwidth and latency is less critical ,by and large its assets being pre fetched to system memory prior to being used by your graphics that causes hitches as the same memory bandwidth is used to move the frame buffer about so when saturated obviously issues can happen ,this is managed by all concerned normally unless you ignored recommended settings.
The quad channel platform should have more bandwidth and be better, but it isn't definite.

 

[NdMk2o1o]

Loathed to link a PC World article but this shows there is hardly any real-world improvement in going from dual to quad channel RAM at the same speeds, in fact going to quad at half the speeds as you asked about would likely return a different picture in where the dual channel system with 2x the speed would be a lot better performing.

http://www.pcworld.com/article/2982...e-shocking-truth-about-their-performance.html

 

[TheoneandonlyMrK]

Loathed to link a PC World article but this shows there is hardly any real-world improvement in going from dual to quad channel RAM at the same speeds, in fact going to quad at half the speeds as you asked about would likely return a different picture in where the dual channel system with 2x the speed would be a lot better performing.

http://www.pcworld.com/article/2982...e-shocking-truth-about-their-performance.html

Thats from 2015 when ddr4 was new i think it shows perfectly why its not clear cut and different platforms differ , the latest quad channel from intel wont play out like that ,imc designs are soo much better including Amds.
I did mention the bus but interconnect/actual traces/fabric and much more all have a say too.

 

[ZenZimZaliben]

4 Channel is faster under large data loads and mutli channel optimized software. Channels are direct connections to the Memory Controller. As an example 4x 64bit channels vs 2x64bit channels. This allows the computer to store 256bits per cycle vs 128bits regardless of memory speeds. It's like video cards with the Bit Ring Bus rating. A very hit bit count and a slower GPU are usually faster than a lower bit rating and faster GPU because it can store more date per cycle. The cpu is able to work with larger chunks of data to the ram.

The problem is software is not programmed well to offer increased parallelism offered by the multi-channel memory configurations.

 

[Mirkoskji]

Ok i'll try to add some detail.
suppose you have a 4 channel enabled pc, like my personal rig, and you have a ddr3 memory that can go up to 2400mt/s.
Also suppose your memory controller has no scaling penality from one channel up to four.
let's keep this on gaming scenario, because I know that productivity applications benefit from higher bandwidth mostly.
If i play fallout 4 in dual channel mode with my memory set to ddr3-2400, and then I try to play it with 4 channel memory at ddr3-1200, would there be a difference?

(also suppose you can linearly scale the latency so if you have 2400 cl10, you can set 1200 cl5)

 

[TheoneandonlyMrK]

I don't know but you can try, bench it and find out it will be quite individual to your platform.

 

[cadaveca]

Not really. There are TWO aspects to memory performance; Bandwidth and latency.

Quad channel adds latency because there are more channels to address, thereby making worse performance in a way, while bandwidth is increased due to the added channels. It would simply either balance itself out, make for a slight deficit, or a slight increase, but not enough for it to really matter.

So, for gaming, increasing speed increases bandwidth and lowers latency, but the real thing that actual affects game FPS is actually the latency, since bandwidth is already more than needed.

You'll see that quad-channel systems have more cores, and thereby more thread, so higher bandwidth is needed if all cores are busy, but that does not reflect higher performance for just a few threads, just like single-threaded apps don't benefit from more cores.

 

[Mirkoskji]

Not really. There are TWO aspects to memory performance; Bandwidth and latency.

Quad channel adds latency because there are more channels to address, thereby making worse performance in a way, while bandwidth is increased due to the added channels. It would simply either balance itself out, make for a slight deficit, or a slight increase, but not enough for it to really matter.

So, for gaming, increasing speed increases bandwidth and lowers latency, but the real thing that actual affects game FPS is actually the latency, since bandwidth is already more than needed.

You'll see that quad-channel systems have more cores, and thereby more thread, so higher bandwidth is needed if all cores are busy, but that does not reflect higher performance for just a few threads, just like single-threaded apps don't benefit from more cores.

thank you, so an old quad channel memory system working at 2400 mt/s is not comparable to, say, a brand new z270 motherboard with dual channel ddr4 4000 or 4200 memory, even if they put out roughly the same bandwidth and can be tuned to have similar total latency.

 

[cadaveca]

Exactly, since the Z270 system would have far lower latency, less than 40ns. But the quad-channel system would be 65ns - 75ns.

That is part of why for gaming, X99 is not the first choice.

 

[Mirkoskji]

Exactly, since the Z270 system would have far lower latency, less than 40ns. But the quad-channel system would be 65ns - 75ns.

That is part of why for gaming, X99 is not the first choice.

53ns the last time i measured. Wow under 40 sounds really impressive!


Anyway now im curious to test. I will find the time to download again fallout 4 and give it a try, just for curiosity. Actually I was only a little lazy, but I found very few information about channel scaling in comparison to frequency scaling in games. Most documentation i found was about frequency only, especially on modern games

 

[cadaveca]

Meh. You can get less than 40ns with lower speeds, even. I'm running some benchmarks right now for memory review; I'll share one with you:

As you can see, I'm starting to approach X99 bandwidth, but with far lower latency already, at just 3600 MHz.

Anyway now im curious to test. I will find the time to download again fallout 4 and give it a try, just for curiosity. Actually I was only a little lazy, but I found very few information about channel scaling in comparison to frequency scaling in games. Most documentation i found was about frequency only, especially on modern games

Adding channels isn't exactly a direct comparison, since there are core changes and cache changes that make the CPU side of things affect the results as well. That lack of correlation is why you cannot find such comparisons. Just comparing the difference in system performance, however, you can find, which is in essence the same thing.

 

[biffzinker]

Wow under 40 sounds really impressive!

Not under but close enough.

 

[cadaveca]

Not under but close enough.
View attachment 88346

That's DDR3, not DDR4, so not a good compare. However, this is evidence of why some people said that DDR4 was slow, and they said the first DDR4 paltforms weren't as good for gaming. It is the latency that mattered.

For DDR3, going under 30ns is where the magic is.

 

[Mirkoskji]

Meh. You can get less than 40ns with lower speeds, even. I'm running some benchmarks right now for memory review; I'll share one with you:

I'm using ddr3 2400 cl10. I could go for ddr3 2133 cl9, or 1866 cl8, but probably i wont squeeze more out of it

 

[Mirkoskji]

That's DDR3, not DDR4, so not a good compare. However, this is evidence of why some people said that DDR4 was slow, and they said the first DDR4 paltforms weren't as good for gaming.

For DDR3, going under 30ns is where the magic is.

well i guess latency got better with the maturation of the platform

 

[cadaveca]

well i guess latency got better with the maturation of the platform

As DDR4 CPU could scale the speeds up higher, yeah, but at the same time, 3200 MHz DDR4 has been available since the very beginning of DDR4, and at that point, DDR4 is most definitely better than DDR3.

However, you'll notice that Intel just started supporting 2400 MHz DDR4 with KabyLake and Broadwell-E... if more bandwidth was truly needed, they could have easily increased that support, but it's not needed because the performance benefits aren't exactly linear when increasing memory alone.

 

[Mirkoskji]

of course, we were talking about a single case scenario (fallout 4) that is mostly an anomaly when dealing with performance and memory speeds

 

[ZenZimZaliben]

Honestly I think you are all missing the mark here. The CPU treats quad channel as a 4x64bit device, capable of writing a 256bit in 1 cycle...where dual channel can only write 128bit blocks. So anything under 256bit takes quad channel 1 cycle..while anything above 128bit takes dual channel 2 cycles..192bit takes dual channel 2 cycle while quad would do it in 1. The only instance where it would be similar is when addressing 128bit blocks otherwise quad will always be superior when utilized correctly.

Regardless of how fast the ram is it can only talk to the cpu when asked. So yeah latency is a big deal.

 

[Mirkoskji]

Honestly I think you are all missing the mark here. The CPU treats quad channel as a 4x64bit device, capable of writing a 256bit in 1 cycle...where dual channel can only write 128bit blocks. So anything under 256bit takes quad channel 1 cycle..while anything above 128bit takes dual channel 2 cycles..192bit takes dual channel 2 cycle while quad would do it in 1. The only instance where it would be similar is when addressing 128bit blocks otherwise quad will always be superior when utilized correctly.

Regardless of how fast the ram is it can only talk to the cpu when asked. So yeah latency is a big deal.

This also seems a good point, anything above 128 bit would suffer from the time ram needs to prepare for a new cycle. it takes 32 cycles to write 1024 byte on a 4 channel configuration, and 64 cycles on a dual channel, right?

 

[RejZoR]

Going from 2 sticks to 4 doesnt equate to going from dual channel to quad channel though which is what the OP is pertaining to

It does on a quad channel motherboard...

 

[NdMk2o1o]

It does on a quad channel motherboard...

But the quote I was replying to didn't have an quad channel board so his experience is different to what the op was asking...