RAM is 4x2 faster than 8x1 ?

[Gonk42]

I realise this subject has been done to death on various threads but I've not seen a comparison of like for like. BareFeats state that 8 are better than 4 but don't compare the same amount of RAM in each case, i.e. they don't compare 8GB with 8GB.

I would expect that 4x2GB would have less latancy than 8x1GB so on average would be a tiny bit faster. Has anyone seen any test results?

(I should say that I'm posting this out of scientific curiosity - from the practical point of view its better to go with 4x2 anyway because it leaves slots free.)

[kirkbross]

I realise this subject has been done to death on various threads but I've not seen a comparison of like for like. BareFeats state that 8 are better than 4 but don't compare the same amount of RAM in each case, i.e. they don't compare 8GB with 8GB.

I would expect that 4x2GB would have less latancy than 8x1GB so on average would be a tiny bit faster. Has anyone seen any test results?

(I should say that I'm posting this out of scientific curiosity - from the practical point of view its better to go with 4x2 anyway because it leaves slots free.)I think it's technically true, but I already bought a 6GB (6x1) kit from OWC so I am prone to want to believe that the difference in latency and performance is negligible.:rolleyes:

[QCassidy352]

I thought having 8 slots filled was actually faster than having an equivalent amount of memory with only 4 slots filled?

[barefeats]

I should have an answer by Monday. I received additional 1GB modules so I can do the 4 x 2 vs 8 x 1 scenario and test it with our 64 bit parallel multithreaded copy of STREAM.

My hypothesis is that the 8 x 1 should be faster, but it doesn't "future proof" you if you plan to expand to 8 x 2 in the near future.

[ajx22]

Ram works better with parity - so having all slots filled is faster/better then not. The OS still 'looks' for RAM to be present in ALL present slots.

This is the reason that the MBP's (etc.) ship with all slots filled (2x 1GB); vs only a single 2GB stick.

[tyr2]

I'd expect that 4x2 is faster than 8x1. Mostly because the memory is serially accessed. Any requests for the second pair in each riser must go via the memory controllers on the first pair first, thus increasing latency.

You need 4 slots filled to achieve memory interlacing and the wider bus, increasing this to 8 won't increase this further.

Will be interested to see the results of your testing barefeats.

[Gonk42]

I should have an answer by Monday. I received additional 1GB modules so I can do the 4 x 2 vs 8 x 1 scenario and test it with our 64 bit parallel multithreaded copy of STREAM.

My hypothesis is that the 8 x 1 should be faster, but it doesn't "future proof" you if you plan to expand to 8 x 2 in the near future.

I look forward to seeing your results. People seem to have convincing arguments both ways. I tend towards the second set being slower as they are only accessed when the memory is not found in the first set which must lead to some delay but I am no expert.

By the way, barefeats, by raising the matter of the tests not being like for like in terms of both being the same RAM size I was not trying to detract from the very interesting results that you've published to date - I'm just a slightly pedantic scientist:)

[orpheus1120]

Subscribed!

[fridgeymonster3]

take a look at barefeats.com. They look at the performance difference between 4x2 GB (8GB Total), 4x2GB + 2xGB (10GB total) and 8x2 (16GB Total). They concluded that the best performance comes from running either 4 matched or 8 matched DIMMS. The 10GB setup only saw a very slight performance increase over the 4x2 GB and the 16 GB was best overall. They suggest either 8x2 or 8x1 for best performance. However, if you are ever thinking about adding more RAM, the 8x1 isn't upgrade friendly.

[Umbongo]

I should have an answer by Monday. I received additional 1GB modules so I can do the 4 x 2 vs 8 x 1 scenario and test it with our 64 bit parallel multithreaded copy of STREAM.

My hypothesis is that the 8 x 1 should be faster, but it doesn't "future proof" you if you plan to expand to 8 x 2 in the near future.

Will you be benchmarking mixed sets also? The 6GB (2x2Gb & 2x1GB) and 10GB (4x2Gb & 2x1GB) configurations that many people see to have bought would be of interest I'm sure.

[Father Jack]

4 x 2 should be faster .. :)

[student_trap]

4 x 2 should be faster .. :)

I certainly hope so:D (4x2 on the way!)

[fridgeymonster3]

I guess nobody is interested in the actual truth - you all just want 4x2 GB to be better than 8x1 GB when that is not the case:

Barefeats.com :
"if you want maximum memory throughput under all situations, we recommend either 8 matching 1GB FB-DIMMs, 8 matching 2GB FB-DIMMs, or 8 matching 4GB FB-DIMMs."

I rest my case!:D


See article: http://barefeats.com/harper3.html

Maybe the Thread Starter should look at previous threads before posting a completely redundant topic that has been covered extensively here.

[gazfocus]

I don't think the difference will be that noticeable.

For me, I'll be getting 4 x 2GB because then it will give me the option to add more ram later where as if I filled all 8 slots, I would have no free slots.

[zdobson]

The 10GB setup only saw a very slight performance increase over the 4x2 GB

I would assume that you would want to leave the stock 2GB in if you order 4x2. Even though the difference isn't much, the stock 2GB will be faster in your machine than on the shelf.

Also, I can't image the benefit of having all slots full with 8x1 would be greater than having the expandability of 4x2.

[fridgeymonster3]

I can't image the benefit of having all slots full with 8x1 would be greater than having the expandability of 4x2.

I agree with you completely, as you can see in my signature that is exactly what I did; however, the OP stated that he was interested in scientific curiosity not practicality in terms of expandability.

[zesta]

Based on what I have read from multiple sources:

2 GB modules should be slightly faster than 1 GB modules.

8 dimms should have better throughput, but slightly more latency than 4 dimms.

I don't think the difference will be noticeable, and I would recommend getting the 4x2, so you can upgrade easier later. That is what I am doing. You can also keep your original 2x 1GB for a total of 10 GB, which may be slightly slower than 4x 2GB, but probably not noticeable.

-Randy

[barefeats]

Will you be benchmarking mixed sets also? The 6GB (2x2Gb & 2x1GB) and 10GB (4x2Gb & 2x1GB) configurations that many people see to have bought would be of interest I'm sure.

Yes. I know that those two configs are very popular. They will be tested.

[barefeats]

Also, I can't image the benefit of having all slots full with 8x1 would be greater than having the expandability of 4x2.

That's a good point. If you are future proofing your investment and the difference turns out to be slight, then 4 x 2GB makes more sense that 8 x1.

[barefeats]

We tested the different combinations with a special 64-bit parallel multi-threaded version of STREAM. I averaged the results of Copy, Scale, Add, and Triad to produce an overall speed rating in gigabytes per second:

4 x 1GB = 4GB = 6.5GB/s

6 x 1GB = 6GB = 6.7GB/s

2 x 1GB + 2 x 2GB = 6GB = 6.5GB/s

4 x 2GB = 8GB = 6.5GB/s

8 x 1GB = 8GB = 7.5GB/s

2 x 1GB + 4 x 2GB = 10GB = 6.8GB/s

6 x 1GB + 2 x 2GB = 10GB = 7.5GB/s

6 x 2GB = 12GB = 6.8GB/s

4 x 1GB + 4 x 2GB = 12GB = 7.5GB/s

2 x 1GB + 6 x 2GB = 14GB = 7.5GB/s

8 x 2GB = 16 GB = 7.5GB/s

Conclusion: Any combo of matching pairs that fills all 8 slots = fastest.

[FastMan]

6 x 1GB + 2 x 2GB = 10GB = 7.5GB/s

8 x 2GB = 16 GB = 7.5GB/s



barefeats, any explanation for this? 10 as fast as 16?

[Spikeanator6982]

8GB 8x1GB the same speed as 10GB and 16GB. As Barefeats said, the speeds are the same if all slots are full. amount of ram and speed dont go together. As far an explaination, bc 10GB and 16GB both take up the 8 slots.:rolleyes:

[ksz]

4 x 2GB = 8GB = 6.5GB/s

2 x 1GB + 4 x 2GB = 10GB = 6.8GB/s

Excellent! Now I will keep the stock 2 x 1GB when adding the new OWC 4 x 2GB for a total of 10GB. I was really expecting to discard the stock 2 x 1GB.

[zdobson]

2 x 1GB + 2 x 2GB = 6GB = 6.5GB/s


I thought this combo wasn't possible according to info on this page:
Memory Test (http://www.barefeats.com/harper3.html)

"If you add two more sticks, they have to match those first two. You can add two more 1GB sticks in slot 2 on both risers for a total of 4GB. You can NOT add two 2GB sticks to that slot or your system will only report 2GB total."

Let me know. I'm hope it is possible because it's what I'd like to do.

[zesta]

I thought this combo wasn't possible according to info on this page:
Memory Test (http://www.barefeats.com/harper3.html)

"If you add two more sticks, they have to match those first two. You can add two more 1GB sticks in slot 2 on both risers for a total of 4GB. You can NOT add two 2GB sticks to that slot or your system will only report 2GB total."

Let me know. I'm hope it is possible because it's what I'd like to do.

You need to move the 1GB from one riser, to make a matched pair. So you end up with:

in the top riser: 2GB, 2GB
in the bottom riser: 1GB, 1GB

They just need to be in matched pairs.

-Randy

[Gonk42]

We tested the different combinations with a special 64-bit parallel multi-threaded version of STREAM. I averaged the results of Copy, Scale, Add, and Triad to produce an overall speed rating in gigabytes per second:

4 x 2GB = 8GB = 6.5GB/s

8 x 1GB = 8GB = 7.5GB/s


Conclusion: Any combo of matching pairs that fills all 8 slots = fastest.

Thank you Barefeats, you've answered my question. It is an interesting result as to me it is counter-intuitive given that there are only 4 serial channels though given that these operate at a much higher frequency than the memory itself I guess what happens on the modules themselves is more important than being first in line.

[Gonk42]

I guess nobody is interested in the actual truth - you all just want 4x2 GB to be better than 8x1 GB when that is not the case:

Barefeats.com :
"if you want maximum memory throughput under all situations, we recommend either 8 matching 1GB FB-DIMMs, 8 matching 2GB FB-DIMMs, or 8 matching 4GB FB-DIMMs."

I rest my case!:D


See article: http://barefeats.com/harper3.html

Maybe the Thread Starter should look at previous threads before posting a completely redundant topic that has been covered extensively here.

In my defense, I did read the previous threads and referred to them in my OP. I couldn't find any test results, at the time, where 4 slots were compared with 8 with the same total RAM (the barefeats results that I saw went up in RAM as the slot occupancy went up so it may have been that more RAM led to an improvement instead of more slots).

Barefeats has now answered the question in a very comprehensive manner and I think I was not the only one interested in the new results.

(I hope the above doesn't sound peevish fridgeymonster - it is not meant to be. :) )

[thePhilster]

Interesting thread, as I've just added my 7th and 8th sticks of RAM to my Mac Pro. Mine are as follows:

Riser A: 4 x 1Gb
Riser B: 2 x 512Mb (original memory) in Slots 1 & 2, 2 x 2Gb in Slots 3 & 4

All RAM slots are filled, they're all in matching pairs, OS X is recognising that there is 9Gb RAM in the system... So, my question is: is there an optimum order to install these RAM chips? For instance, is having 4Gb on Riser A and 5Gb on Riser B gonna cause any problems? Should the 2Gb sticks be in slots 1 & 2 on Riser B rather than slots 3 & 4? Should I just content myself with the fact that OS X is regognising 9Gb RAM and get on with things?? ;)

Phil

[Umbongo]

Interesting thread, as I've just added my 7th and 8th sticks of RAM to my Mac Pro. Mine are as follows:

Riser A: 4 x 1Gb
Riser B: 2 x 512Mb (original memory) in Slots 1 & 2, 2 x 2Gb in Slots 3 & 4

All RAM slots are filled, they're all in matching pairs, OS X is recognising that there is 9Gb RAM in the system... So, my question is: is there an optimum order to install these RAM chips? For instance, is having 4Gb on Riser A and 5Gb on Riser B gonna cause any problems? Should the 2Gb sticks be in slots 1 & 2 on Riser B rather than slots 3 & 4? Should I just content myself with the fact that OS X is regognising 9Gb RAM and get on with things?? ;)

Phil

Riser A should probably be:
slot1: 2GB
slot2: 2GB
slot3: 1GB
slot4: 1GB

Riser B should probably be:
slot1: 1GB
slot2: 1GB
slot3: 512mb
slot4: 512mb

That's based on the first slots starting with riser A being the fastest. You probably wouldn't notice any real world performance differences though. On the 2006 Mac Pro memory chart that often gets posted it's suggested you have the 1GB sticks all on Riser B and the 2GB in slots 1 & 2 on Riser A. However I don't know if that chart is designed to display optimized setups.

[MilesS2111s]

We tested the different combinations with a special 64-bit parallel multi-threaded version of STREAM. I averaged the results of Copy, Scale, Add, and Triad to produce an overall speed rating in gigabytes per second:

....

8 x 1GB = 8GB = 7.5GB/s

8 x 2GB = 16 GB = 7.5GB/s

Conclusion: Any combo of matching pairs that fills all 8 slots = fastest.

Some newbie questions:
1) Does this mean that I've just wasted my cash buying 8x2GB as this is just as fast as the 8 x 1GB?
2) If I retain the supplied 2 x 1GB to take myself to 18GB this will actually result in a performace drop?

Many thanks

Miles

[Umbongo]

Some newbie questions:
1) Does this mean that I've just wasted my cash buying 8x2GB as this is just as fast as the 8 x 1GB?

Not if you need more than 8GB of memory.

2) If I retain the supplied 2 x 1GB to take myself to 18GB this will actually result in a performace drop?

Well without 4GB DIMMs you can't get 18GB anyway.

[MilesS2111s]

Not if you need more than 8GB of memory.

I'm buying RAM to make the machine run faster but (to my inexperienced eyes) the excellent data suggests that there is no speed difference which seems odd! Or have I got the wrong end of the stick in terms of how RAM makes the machine faster!?

Well without 4GB DIMMs you can't get 18GB anyway.
DUH!:rolleyes: Ebay here I come:cool:

[barefeats]

You need to move the 1GB from one riser, to make a matched pair. So you end up with:

in the top riser: 2GB, 2GB
in the bottom riser: 1GB, 1GB

They just need to be in matched pairs.

-Randy

That's right. Though they come from the factory with 1 x 1GB on the top riser and 1 x 1GB on the bottom riser (base 2GB config), if you only want to add 2 x 2GB, you have to put the two 1GB FB-DIMMs on the same riser. That's what we did when we tested.

[Boemer]

Interesting that 8 chips is faster than 4.

Xeon has a 256 bits bus, meaning that you need 4 chips, better have them the same size, if you take like 2x 1GB, 2x 2GB, when your test program will make its test within the first 1GB, it will be faster. But when in the real world, you really use that memory, it will become slower when you in the second GB (so like in the 5th GB and 6th GB.

I'm not sure why 8 chips is faster, because it will still have 256bits access... I could imagen that it is faster because of having 2 xeon chips. Too bad, I just ordered 4x 2GB, and I really don't need those 8GB extra after the first 8GB....

[Gonk42]

Interesting that 8 chips is faster than 4.

Xeon has a 256 bits bus, meaning that you need 4 chips, better have them the same size, if you take like 2x 1GB, 2x 2GB, when your test program will make its test within the first 1GB, it will be faster. But when in the real world, you really use that memory, it will become slower when you in the second GB (so like in the 5th GB and 6th GB.

I'm not sure why 8 chips is faster, because it will still have 256bits access... I could imagen that it is faster because of having 2 xeon chips. Too bad, I just ordered 4x 2GB, and I really don't need those 8GB extra after the first 8GB....

I've been looking at the Wikipedia entry on FB-DIMMs and this refers to a MSc Thesis which I also looked at (skim read). The main thing of interest that struck me was a statement saying that having two DIMMs in a channel reduces read after write and write after read delays if they are from different DIMMs because each DIMM module has a buffer as part of the serial to parallel conversion. This will tend to overcome the inherent latency of the second DIMM module only being accessed when the first has had a look in.

In other words, in some circumstances the two DIMM modules in the same channel can act in parallel.

(I apologise to any hardware engineers for my rather simplistic and probably wildly inaccurate reading of the facts.)