Socket-2011 vs socket-1155 and quad-channel vs dual channel

[BonzaiDuck]

Should've taken me much less time to surmount the sense of confusion about why Intel released socket 2011 for a few Sandy Bridge E processors, then resumed producing socket-1155 releases of Ivy Bridge.

Now . . . . "I get it . . . . "

Has anyone . . . anyone here . . . . been able to have a sense of a difference in performance for having quad-channel versus dual-channel memory potential?

Or how this might be amplified for using DDR3-2600 (etc.) in quad-channel mode versus any RAM speed choice in dual channel mode?

 

[alyarb]

I'm sure more than a few reputable tech sites have done the z77 vs x79 shootout and in most cases there is not much real world difference in a single environment but probably dramatic differences in the synthetic tests. You can almost always consider the i7 3820 to perform identically to a 2600k at stock.

The system was intended for 2S and 4S machines running 10 or more virtual servers with enough action going on to necessitate such memory utilization, but that isn't going to stop intel from selling it to some kid who runs one instance of a game and one instance of a voip client.

 

[ShintaiDK]

LGA2011 is like LGA1366 made for dualsocket. The purpose of the tripple or quadchannel is to secure not only the local CPU. But also the remote CPU in terms of NUMA memory access via QPI.

Another benefit is it increases the possible memory amount as well.

 

[BrightCandle]

S2011 is about extremes. You have more memory channels, more cores, more cache and it all adds up to higher TDP. In a lot of average desktop software cases none of these things help the performance at all, because they are single threaded applications that for the most part fit in small caches and don't saturate the existing memory bandwidth nor utilise a great deal of RAM.

S2011 isn't about the average application or game, its way to expensive to be worth the zero advantage it brings in that environment. S2011 unlike 1366 before it is a niche product targeting people that have software that can utilise the 'mores'. I have software I run that takes 12GB of RAM, happily pushes through 20-30 GB's a second of memory and would use 100 cores if I had them. The compilers I use can utilise the 6 cores fairly well, my unit and functional tests certainly run on all those CPUs and even get a nice boost from HT. Professional uses can benefit to the tune of a good 50-75% extra performance if the software they have utilises the cores and the RAM.

If you don't have a piece of software that you know will benefit then S2011 it is a waste of money. If you don't know if S2011 is for you then its not. Ivy Bridge serves better performance for the enthusiast and gamer than S2011 does.

 

[BonzaiDuck]

S2011 is about extremes. You have more memory channels, more cores, more cache and it all adds up to higher TDP. In a lot of average desktop software cases none of these things help the performance at all, because they are single threaded applications that for the most part fit in small caches and don't saturate the existing memory bandwidth nor utilise a great deal of RAM.

S2011 isn't about the average application or game, its way to expensive to be worth the zero advantage it brings in that environment. S2011 unlike 1366 before it is a niche product targeting people that have software that can utilise the 'mores'. I have software I run that takes 12GB of RAM, happily pushes through 20-30 GB's a second of memory and would use 100 cores if I had them. The compilers I use can utilise the 6 cores fairly well, my unit and functional tests certainly run on all those CPUs and even get a nice boost from HT. Professional uses can benefit to the tune of a good 50-75% extra performance if the software they have utilises the cores and the RAM.

If you don't have a piece of software that you know will benefit then S2011 it is a waste of money. If you don't know if S2011 is for you then its not. Ivy Bridge serves better performance for the enthusiast and gamer than S2011 does.

See -- I thought this might be the case. Even if I were to resurrect my "programming" hobby or other IT interests, I doubt that the performance gains of skt 2011 (fully utilized and configured) would be worth it. ONe would have to be working under time constraints to make it so.

 

[utahraptor]

I think one of the only benefits a gamer would get would be reduce load times if fancy cache was I use.

 

[BonzaiDuck]

I got into this technology in the early '80s -- so I could do Discriminant Analysis with expensive software or write my own in BASIC or C.

During my "other life" among the working-wounded, I was prone to sneering at gaming and "multi-media" applications. My low tolerance has now transferred to mobile-device junkies -- the "Beam me up, Scotty" crowd . . . [those . . . those . . . . youngsters . .. ], while I do simultaneous gaming and multi-media.

I have a friend -- like many others -- a latecomer to what was once known as "microcomputers." He started building his own in 2003; I started over-clocking in 2004 as my friend only dabbled. Too many BSOD's -- he got cold feet.

So during every build cycle between us, he's plotting to get the latest, greatest, fastest. He tends to think that the best measuring stick is the old dollar-sign. And if he sees this, I hope he doesn't misinterpret my remark. There are many forum members here who may not look at our efforts as a balance between cost, speed, performance, power-usage and other factors.

He's looking to build again -- maybe in the fall. So the question came up: "Ivy Bridge, Sandy-Bridge or Sandy-Bridge E?" "Socket 1155 or 2011?"

 

[ShintaiDK]

SB-E only makes sense in 3 areas for the desktop.

1. You need 6 cores for whatever reason.
2. You need more than 32GB memory.
3. You need the extra PCIe lanes for that quad GPU setup or those 5 RAID cards for your "collection".

And why buy SB when there is IB.

 

[BonzaiDuck]

SB-E only makes sense in 3 areas for the desktop.

1. You need 6 cores for whatever reason.
2. You need more than 32GB memory.
3. You need the extra PCIe lanes for that quad GPU setup or those 5 RAID cards for your "collection".

And why buy SB when there is IB.

Even the latter point is valid . . . despite any over-clocking game changes . . . price not much of a consideration if any at all . . .

I briefly had an inspiration to build a Gulftown skt 1366. Looking at the lower of the two price options on the CPU, it still didn't make sense. One is hardpressed to get full use of just a quad-core system with a bundle of mainstream application programs and applications.

 

[rickon66]

Two years ago as i was preparing to buld a new rig, I was scouring the tech web sites and bulletin boards to get info about dual channel vs. triple channel. After extensive research, the general opinion of all the experts was socket 1366 - with triple channel memory and it would be more future proof. I went with an I7-930, but would have been just as well served with an I7-860 and saved a few bucks. Well, as we know now socket 1366 = dead end and triple channel memory was no real advantage in most applications. Unless you have very specific needs or very deep pockets, just go with the mainstream CPU/socket and put the saving towords that next build.

 

[pantsaregood]

I doubt even a pair of eight-core Xeons would really need quad-channel RAM, honestly. DDR3-2133 and DDR3-2400 are actually quite common - use of high-bandwidth DIMMs can increase overall bandwidth just as easily as more RAM channels. A quad-channel board using DDR3-1333 (pretty typical) provides only marginally higher theoretical bandwidth than dual-channel DDR3-2400.

Current six-core CPUs aren't really capable of taking advantage of the ridiculous bandwidth available to them. Perhaps a dual-CPU setup would be different, but even a 3960X at 4.5 GHz won't see a discernable difference between RAM running at 533 MHz and 1066 MHz.

EDIT: The Xeon E5-4xxx platform apparently supports quad-CPU setups. I would assume high memory bandwidth would be beneficial to these.

 

[mtnd3vil]

LGA2011 is like LGA1366 made for dualsocket. The purpose of the tripple or quadchannel is to secure not only the local CPU. But also the remote CPU in terms of NUMA memory access via QPI.

1366 was also a 2S capable NUMA platform.

3/4 channel can't really be justified by the logic of "well...multi-cpu" because with that additional cpu comes another (NUMA node) memory controller, 3/4 channels, and DIMM slots.

What do you think?

 

[mtnd3vil]

SB-E only makes sense in 3 areas for the desktop.

1. You need 6 cores for whatever reason.
2. You need more than 32GB memory. or 32GB @ 4GB DIMM prices
3. You need the extra PCIe lanes for that quad GPU setup or those 5 RAID cards for your "collection". Yeah, totally right. Any PCIe device Raid, PCIe based SSD. How about 10Gbe NIC? Long before the switches become affordable, I would love to have a direct 10Gb ethernet link from my LAN server to my workstation.

And why buy SB when there is IB.

in quote

 

[ShintaiDK]

1366 was also a 2S capable NUMA platform.

3/4 channel can't really be justified by the logic of "well...multi-cpu" because with that additional cpu comes another (NUMA node) memory controller, 3/4 channels, and DIMM slots.

What do you think?

If both CPUs accesses the memory located in the memory of 1 CPU. What happens?

 

[mtnd3vil]

If both CPUs accesses the memory located in the memory of 1 CPU. What happens?

It's possible and normal for both CPUs to access RAM in any node, but with less performance than accessing memory within the CPU's home NUMA node. Non-Uniform Memory Acess is named to describe this.

The x79 based LGA2011 systems are only one NUMA node which is kind of funny since that in-turn means the memory access performance IS uniform across all RAM...

 

[ShintaiDK]

And now you know why LGA1366 had 3 channels and LGA2011 4.

 

[sefsefsefsef]

The only reason I would consider socket 2011 for my personal machine is to load it up with 64 GB of DDR3. The bandwidth advantage will not be noticeable. Client CPUs are already pretty much over-provisioned for bandwidth with 2 channels. Even server workloads with many cores running full-bore don't saturate the memory bandwidth they have available to them. That excess bandwidth goes unused unless you can do some useful pre-fetching with your otherwise-idle DRAM channels.

 

[Magic Carpet]

i wonder, how Windows would prefetch 64 gigs of RAM. Can't wait to see

 

[BonzaiDuck]

There also seem to be differences of opinion about choice of DDR3 rated speeds, with articles published by Anandtech and Hexus which suggest that the caching innovations of both Sandy and Ivy "blunt the hopes of high-performance RAM kits."

In my case, I have a spare set of RipJaws 2x4GB DDR3-1600 9-9-9-24 to fill my remaining sockets and get 16GB. I can get a 2x8GB set of DDR3-1600 RipJaws or RipJaws X with the same latencies and 1.5V spec for about $120. Or I can spend $60 more and get a 2x8GB DDR3-2133 running at 1.60V and sporting CAS=9 with the remaining latencies looser by a value of 2 (or -11-11 . . . )

Right now, 8GB seems adequate . . . . and I like the idea of just putting the money in my savings account . . .

 

[Ferzerp]

i wonder, how Windows would prefetch 64 gigs of RAM. Can't wait to see

It stops after a while. Just to see I logged in to a 256GB RAM box, and despite currently having almost all that free, it's only using 17.5GB for that.