Quad channel RAM in CPU-bound scenarios

[yoadknux]

Hello everyone,

I'm running a 1080 GTX on a 4930k @ 4.4GHz and 2x8 2133MHz (CL13 - stock is CL9 1600) memory. I use the system for gaming. I have a 1080p 144Hz monitor, so I play on low resolutions. I suspect that in those resolutions, I am in a situation that is "CPU Bound". I do not feel that the 4930k seriously bottlenecks my 1080; though I suspect that a higher end processor and RAM (example: 7700k + 2x8 3000MHz Ram) could have given me slightly more performance.

Thing is I am not going to buy a new mobo, new processor and new memories for 10FPS. I have done a little bit of research about single vs dual channel ram, and it seems that dual channel can increase gaming performance in situations that are not GPU bound. From my own research, playing with 1 stick does introduce a slight performance hit... This makes me wonder if going quad channel could give me additional performance.

Any predictions? I am probably going to try it anyway because I can get a decent deal on additional 2x8 sticks from the same brand and same timings, though I'd love to hear about the science behind it.

Thanks!

 

[Woomack]

I guess that you can have better results when you tighten the timings. Point is that memory access time improves performance on the way CPU cache <-> RAM. You may even see higher performance at 2133 CL9 in dual channel than 2133 CL13 in quad channel. That's only a theory but sometimes it works like that, depends on used program.

Probably all DDR3 memory kits can make 2133 CL11 so I would play with memory settings and try to set it to 2133 11-13-13 ( or tighter ) or 2400 11-13-13.
In most games changing CPU won't give anything. Also most games are not using additional CPU instructions so older gen will perform about as good as new. Difference is mainly in CPU cache speed and access time to memory and cache. I still wouldn't change CPU only for games. Better wait and upgrade it in 1-2 generations ( still don't know if it will be worth it ).

 

[yoadknux]

I guess that you can have better results when you tighten the timings. Point is that memory access time improves performance on the way CPU cache <-> RAM. You may even see higher performance at 2133 CL9 in dual channel than 2133 CL13 in quad channel. That's only a theory but sometimes it works like that, depends on used program.

Probably all DDR3 memory kits can make 2133 CL11 so I would play with memory settings and try to set it to 2133 11-13-13 ( or tighter ) or 2400 11-13-13.
In most games changing CPU won't give anything. Also most games are not using additional CPU instructions so older gen will perform about as good as new. Difference is mainly in CPU cache speed and access time to memory and cache. I still wouldn't change CPU only for games. Better wait and upgrade it in 1-2 generations ( still don't know if it will be worth it ).

Hi,

I've done two things: One, tried to follow your advice of going 11-13-13, and two, bought another 2x8 CL9 (which turned out to be slightly different than my first kit, as I'll now explain).

First, let's talk about the old 2x8 pack: I tried going 11-13-13 but I get BSOD when I run Prime95 blend. So that's obviously not working out. I tried 12-13-13 and it was stable for a while, but then Skyrim crashed out of nowhere, most likely because of the ram. I cannnot do better than CL13 for this pack. I could try increasing ram voltage, but I'm not sure it's gonna do anything (It sits at 1.52V atm).

Now, let's talk about the new 2x8 pack: I got a great deal on it, but it's not an identical model, my older 2x8 is G-Skill ARES while the new one is G-Skill RipJawsX. When I added the new 2x8 (4x8 now) and tried running CL11, it actually worked fine (so i thought)... Memtest no errors overnight, but turns out only the RipJaws can do CL11 and the ARES cannot, I guess when you have multiple ram sticks and only one of them manages to overclock, the other one is ignored entirely. I confirmed that the OS did not identify 32GB but rather 16GB. Going back to CL13 gives me 32GB.

So now comes the question which is better, CL13 4x8 Quad channel or CL11 2x8 Dual Channel. To be honest I don't think there's a difference between 32GB and 16GB for gaming, from my tests even 8GB is fine (I have another 2x4 kit, and for most games, the performance is similar). I can mix RAM and do 2x4 + 2x8 at CL11 2133MHz Quad channel but I think it could mess things up.

As for the results, here are the differences that I've noticed when doing 4x8:
- Prime95 blend makes the CPU go hotter now. It's not dangerously hot (hottest core after 30m is 75c) but it's about 5-10c hotter than what I had with dual channel. Does it make sense? What's the science behind it?
- Userbenchmark gives me a much higher RAM score (111% vs 72%), but it's weighted mostly on multi-core throughput. I am not sure if the increase in RAM score is due to larger ram or the quad channel kicking in.
- TimeSpy showed almost the exact same score, but it's weighted on CPU+GPU.

I'll do some gaming benchmarks later. We'll see how it goes.

 

[Woomack]

Try 1.65V on the RAM. Anything up to 1.7V ( usually more ) is safe for all IC if temps are below max ( but max is 80°C+ and RAM works at about 40-50°C max ). Can also bump VCCSA voltage a bit to stabilize memory controller.

In 3DMark when you compare memory performance then look at physics tests as almost only there are differences.

 

[yoadknux]

Try 1.65V on the RAM. Anything up to 1.7V ( usually more ) is safe for all IC if temps are below max ( but max is 80°C+ and RAM works at about 40-50°C max ). Can also bump VCCSA voltage a bit to stabilize memory controller.

In 3DMark when you compare memory performance then look at physics tests as almost only there are differences.

I've already noticed that going from 2x8 to 4x8 caused my CPU temperature to increase under load, is anything going to happen to the CPU if I increase the RAM voltage to 1.65V? Does it strain the CPU in some way? What about the VCCSA? If I remember correctly it sits at 1.1V for me at the moment

 

[Woomack]

Check VCCSA at 1.2V. Generally 1.25V was fine for 2400 memory as I remember. Memory voltage affects only memory and it won't cause CPU to heat up more. VCCSA on the other hand can add 1-2°C but not much more.

 

[yoadknux]

Check VCCSA at 1.2V. Generally 1.25V was fine for 2400 memory as I remember. Memory voltage affects only memory and it won't cause CPU to heat up more. VCCSA on the other hand can add 1-2°C but not much more.

Went with Ram 1.65V VCCSA 1.2V, didn't help. CL11 2133 only RipJawsX is detected (Ares detected in bios, but seems to be disabled). CL13 2400 no POST. What do you think is going on? Also, do you think I should try to throw in my 2x4 Vengeance with the Ripjaws? They can probably do CL11 together @ 2133 but not sure if it's still gonna be quad channel and how well they're gonna work together.

Update: Plugged the 2x4 with the 2x8. Running CL11 2133. CPU-Z reports this configuration as Quad channel, and the Userbenchmark shows a similar RAM score (126% vs 123%, tiny bit higher because of latency). As for stability, I'll run memtest overnight but was able to run 1h of Prime95 blend so it has some stability. I'll try to game with it to check the speed.

What do you think? Any downsides to running this configuration (2x4 2x8 CL11) compared the the slightly slower but balanced configuration? (4x8 CL13)

Update 2: Interesting, I can run 2x4 2x8 2400 CL13 @ 1.58V Ram 1.15 VCCSA.

 

[mackerel]

- Prime95 blend makes the CPU go hotter now. It's not dangerously hot (hottest core after 30m is 75c) but it's about 5-10c hotter than what I had with dual channel. Does it make sense? What's the science behind it?

Prime95 and similar software running large FFT works on a large data set that doesn't fit in the CPU caches. As such it is often ram bandwidth limited on modern higher core/clocked Intel CPUs. As a rough rule of thumb, to be "practically ram bandwidth unlimited" I'd suggest dual channel ram running around the same rated speed as a Haswell or newer CPU quad core clock. Assuming Ivy Bridge is same performance as Sandy Bridge, 6 cores at 4.4 GHz is equivalent to Haswell 4 cores at 4.4 GHz (Haswell has 1.5x IPC relative to Sandy Bridge at this task due to AVX2). Dual channel 2133 simply isn't going to cut it, quad channel is enough to go into practically unlimited territory. I should add, Prime95 performance responds primarily to ram bandwidth, and latency doesn't make much of a difference.

Note that only applies to large FFT. Small FFT does fit in CPU cache, and there I'd expect to see no significant difference due to the ram, unless there is some power overhead even when the ram isn't actively being hit hard.

 

[wingman99]

lowered memory timings and more system memory channels above dual channel will yield little. Most of the work for FPS is the Vram on the video card. System memory is slow compared with Vram also data transfer takes time with CPU and the PCI-E bus working together to transfer to the Vram. Most game manufactures concentrate on having the data needed in Vram first, then loading Vram from system memory before you reach that part of the game map.

 

[yoadknux]

Prime95 and similar software running large FFT works on a large data set that doesn't fit in the CPU caches. As such it is often ram bandwidth limited on modern higher core/clocked Intel CPUs. As a rough rule of thumb, to be "practically ram bandwidth unlimited" I'd suggest dual channel ram running around the same rated speed as a Haswell or newer CPU quad core clock. Assuming Ivy Bridge is same performance as Sandy Bridge, 6 cores at 4.4 GHz is equivalent to Haswell 4 cores at 4.4 GHz (Haswell has 1.5x IPC relative to Sandy Bridge at this task due to AVX2). Dual channel 2133 simply isn't going to cut it, quad channel is enough to go into practically unlimited territory. I should add, Prime95 performance responds primarily to ram bandwidth, and latency doesn't make much of a difference.

Note that only applies to large FFT. Small FFT does fit in CPU cache, and there I'd expect to see no significant difference due to the ram, unless there is some power overhead even when the ram isn't actively being hit hard.

Great explanation! I haven't seen higher small FFT temps. I guess the gaming temperatures shouldn't be different for that reason.

lowered memory timings and more system memory channels above dual channel will yield little. Most of the work for FPS is the Vram on the video card. System memory is slow compared with Vram also data transfer takes time with CPU and the PCI-E bus working together to transfer to the Vram. Most game manufactures concentrate on having the data needed in Vram first, then loading Vram from system memory before you reach that part of the game map.

You're probably right. Was just wondering if it'll help in high CPU games for which I already get high FPS to begin with such as Overwatch. Didn't help so much. I will try the 3dmark physics test in a bit.

 

[yoadknux]

Sorry for the double post! Just wanted to add some additional information. I've continued my research. Throughout this topic I mostly discussed two different 2x8 sets that I mixed and got them to work together at CL13 2133. At some point I have also mentioned owning a 2x4 Corsair set.

I've decided to continue mixing, find a guy selling some cheap G-Skill RipJawsX 1.65V CL9 2133MHz, and told myself, why not mix them with my Corsair Vengeance 2x4 CL11 2133. First, I can run them at CL10 2133, which is pretty reasonable. I also bumped VCCSA to 1.2V and can run them at 2400MHz! Pretty interesting. I wonder how much further I can go.

My only concern is that the Corsair RAM is designed for 1.5V while the RipJaws go for 1.65V... and The VCCSA is high according to the internet. I'll keep testing!

 

[wingman99]

I'm glad to hear mixing the memory works fine.