CPU for single threaded string algorithms

[protochar]

Hi ppl,

I am building a system for testing string algorithms.
I'm not an expert for hardware so any suggestions at all
are welcome, here are some facts, conclusions and questions:

1. Algorithms to be tested resemble Lempel-Ziv (~Zip) algorithms.
2. Only integer operations performance is important
3. Algorithms are very memory intensive and frequently read from random
locations spread across many GBs of ram.
4. Algorithms are single threaded.

I think that memory bandwidth+latency and processor integer
operations performance and speed are most important, cache not so much.

Currently I'm considering Core i5 680 as and Core i7 950
(they're within the budget).

One comparison shows that i5 660 (slower than i680) is
better than Core i7 975 (faster than i7 950), for
all the archiving benchmarks.

The only thing that could be an advantage of i7 is greater memory
bandwidth (3 vs. 2 memory controllers), which could speed up thing
when jumping randomly over GBs of RAM and reading chunks into cache.

How much difference will it be is questionable,
but do you think more memory controllers could speedup ram access
for single threaded apps? Would optimal ram configuration for
2/3 mem. controllers mean 2/3 ram modules?
Do you think Xeon's would make some difference
concerning integer operations speed or mem. bandwidth?

Thanks a lot for any comments

 

[13thmonkey]

cuda? yes 2 channel and 3 channel controllers work best with 2 and 3 sticks of ram.

 

[protochar]

cuda? don't know, quite parallel, we are 1threaded for now, but I'll take a look. tnx

 

[geofelt]

Some thoughts:

1) With current products, I would favor one of the 32nm clarkdale i5 duos like the i5-680. But, The cost of a few extra starting multipliers seems excessive. For maximum compute power, you will want to overclock a bit. With a decent cpu cooler, it is very easy on the 32nm parts. I would suggest the i3-550 http://www.newegg.com/Product/Product.aspx?Item=N82E16819115065 All of the clarkdale 32nm chips are essentially the same, and should be able to be overclocked to the same high limit.

2) Extra ram channels, fast speeds, and lower latencies do not seem to be a major factor in real application performance on the nehalem architecture cpu's. The impact is in the 1-3% range. That is because the integrated memory controllers do a very good job of keeping the cpu fed with data. Do not be seduced by synthetic benchmark results of fast ram and triple channels which look wonderful.

3) Unless your need is urgent, I suggest waiting for sandy bridge, due to be available Jan 9. The new chips will be some 15% faster on a clock for clock basis. They also have more aggressive single thread turbo's which fits your needs. Read a preview here:
http://www.anandtech.com/show/3871/the-sandy-bridge-preview-three-wins-in-a-row

 

[protochar]

Tnx geo, good advices. No worries about mem bandwidth than. As for overclocking, doesn't it make the
system less reliable (in terms of crashing) and short-lived? And unfortunately i need to buy the system this
month, so I can't wait for sandy bridge. I think I'll go with i5 680. Does anybody know if it supports ECC?
No info on Intel's pages....

 

[geofelt]

Tnx geo, good advices. No worries about mem bandwidth than. As for overclocking, doesn't it make the
system less reliable (in terms of crashing) and short-lived? And unfortunately i need to buy the system this
month, so I can't wait for sandy bridge. I think I'll go with i5 680. Does anybody know if it supports ECC?
No info on Intel's pages....

1) I don't think ECC support is there. I don't know that it matters, or is worth the extra cost anyway.

2) As long as you leave the voltages at stock or auto, I think a conservative OC is safe. CPU's will be long obsolete before any extra clocking has an effect. I might hazard a wild guess that a 20 yearcpu lifespan might be reduced to 18? A cpu is designed to downclock itself, or even shut off if it passes safe thermal levels.

 

[protochar]

1) I don't think ECC support is there. I don't know that it matters, or is worth the extra cost anyway.

Yes, unfortunately no ECC for i3:

http://www.intel.com/support/processors/corei3/sb/CS-031175.htm

I'm also wondering how likely is it for soft ram error to occur. Asessments range from "now and then" to "never and not even then" (once in a century or so). But many people that do serious computation see ECC as a must. Don't know...

 

[13thmonkey]

How about running 2 identical computations and comparing the results? or is there randomness involved?