GPU to 72 status... - Page 239

chalsall · 2014-01-08, 23:19

Quote:

Originally Posted by TheMawn

As far as I know, the "speed" at which bits are moved between the components is basically negligible.

Incorrect. Memory bandwidth can be very important.

As I discovered during this exercise, tuning your CPUs to your caches, and your memory, can increase your throughput significantly.

tha · 2014-01-08, 23:23

Well, I would like to know for example if it would make sense to use a X79 based motherboard. It has a quad channel memory instead of a dual channel memory. Does anyone have one?

I see serious differences between systems with identical processors. That may be just different memory speeds. If a 2,0 Ghz Xeon can outpace an ordinary quad core running at 2,4 Ghz by such a wide margin, I would like to know which component contributes what to the final results.

TheMawn · 2014-01-08, 23:28

Quote:

Originally Posted by chalsall

Incorrect. Memory bandwidth can be very important.

Oh no no I didn't mean that. I thought this discussion was about the bits moving between components (i.e. between the RAM and CPU). I know full well the importance of the RAM bandwidth (I need 20% faster RAM to unbottleneck myself, as a matter of fact)

kracker · 2014-01-09, 00:15

Quote:

Originally Posted by TheMawn

Oh no no I didn't mean that. I thought this discussion was about the bits moving between components (i.e. between the RAM and CPU). I know full well the importance of the RAM bandwidth (I need 20% faster RAM to unbottleneck myself, as a matter of fact)

??

This is about that, what else? RAM moving within RAM?

TheMawn · 2014-01-09, 03:48

I donno. The connections between the RAM and the CPU? Like the actual electrical signals moving across the board? To be fair I didn't exactly understand what the question meant.

kracker · 2014-01-09, 03:58

Now I'm confused too

. Memory bandwidth is the memory moved from RAM to CPU, CPU to RAM, (chipset, HDD, blah blah)

Please tell me if I'm wrong, I'm not exactly the most knowing in this area.

LaurV · 2014-01-09, 15:19

Memory bandwidth is when bits move inside of my computers.
Blah blah blah is when bits move inside of your computers.
No matter from which component to which..

blahpy · 2014-01-10, 01:43

Quote:

Originally Posted by kracker

??

This is about that, what else? RAM moving within RAM?

I think he's meaning latency between RAM and CPU (not the bandwidth), id est the fact that the signal can only travel at c (which is not an issue) and not the actual amount of data that can be transmitted in any given time.

henryzz · 2014-01-11, 00:22

Memory bandwidth is the total data throughput of the connection between ram and the cpu.
On a fast cpu it is possible to saturate this with only 3 cores. In this case four 3 Ghz cores would run as fast as four 4 Ghz cores because the memory bandwidth can only handle 12 Ghz of throughput in total.
The cpu caches have much higher bandwidth than ram but are usually way too small.
In the haswells there are a few chips with suffix R that have 128MB of eDRAM built into them as L4 cache. This has very high bandwidth but is not large enough for the tests done by GIMPS currently(for 4 tests at least). LLR tests could benefit massively as they are in general smaller and will fit in it.
Latency doesn't really matter that much for GIMPS because of the way it is optimized.

chalsall · 2014-01-11, 00:47

Quote:

Originally Posted by henryzz

Latency doesn't really matter that much for GIMPS because of the way it is optimized.

Happy to be corrected, but what I have found (empirically) is having each processor leverage on their individual immediate memory caches whenever possible is optimal.

henryzz · 2014-01-11, 12:01

Quote:

Originally Posted by chalsall

Happy to be corrected, but what I have found (empirically) is having each processor leverage on their individual immediate memory caches whenever possible is optimal.

The difference between cache and ram might be worthwhile but the difference between 9-9-9-27 and 8-8-8-24 isn't very big.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Status	Primeinator	Operation Billion Digits	5	2011-12-06 02:35
62 bit status	1997rj7	Lone Mersenne Hunters	27	2008-09-29 13:52
OBD Status	Uncwilly	Operation Billion Digits	22	2005-10-25 14:05
1-2M LLR status	paulunderwood	3*2^n-1 Search	2	2005-03-13 17:03
Status of 26.0M - 26.5M	1997rj7	Lone Mersenne Hunters	25	2004-06-18 16:46

2014-01-08, 23:23	#2620
tha Dec 2002 2×11×37 Posts	Well, I would like to know for example if it would make sense to use a X79 based motherboard. It has a quad channel memory instead of a dual channel memory. Does anyone have one? I see serious differences between systems with identical processors. That may be just different memory speeds. If a 2,0 Ghz Xeon can outpace an ordinary quad core running at 2,4 Ghz by such a wide margin, I would like to know which component contributes what to the final results.

2014-01-09, 03:48	#2623
TheMawn May 2013 East. Always East. 11·157 Posts	I donno. The connections between the RAM and the CPU? Like the actual electrical signals moving across the board? To be fair I didn't exactly understand what the question meant.

2014-01-09, 03:58	#2624
kracker "Mr. Meeseeks" Jan 2012 California, USA 2³·271 Posts	Now I'm confused too . Memory bandwidth is the memory moved from RAM to CPU, CPU to RAM, (chipset, HDD, blah blah) Please tell me if I'm wrong, I'm not exactly the most knowing in this area.

2014-01-09, 15:19	#2625
LaurV Romulan Interpreter Jun 2011 Thailand 7×1,373 Posts	Memory bandwidth is when bits move inside of my computers. Blah blah blah is when bits move inside of your computers. No matter from which component to which..

2014-01-11, 00:22	#2627
henryzz Just call me Henry "David" Sep 2007 Cambridge (GMT/BST) 2³×3×5×7² Posts	Memory bandwidth is the total data throughput of the connection between ram and the cpu. On a fast cpu it is possible to saturate this with only 3 cores. In this case four 3 Ghz cores would run as fast as four 4 Ghz cores because the memory bandwidth can only handle 12 Ghz of throughput in total. The cpu caches have much higher bandwidth than ram but are usually way too small. In the haswells there are a few chips with suffix R that have 128MB of eDRAM built into them as L4 cache. This has very high bandwidth but is not large enough for the tests done by GIMPS currently(for 4 tests at least). LLR tests could benefit massively as they are in general smaller and will fit in it. Latency doesn't really matter that much for GIMPS because of the way it is optimized.