Upgrading T2400 to T7200 makes big speed improvement

[kurutta]

Just replaced the T2400 in my T60 with a T7200. After reassembling the machine I performed the following test in Matlab:

r=rand(5000); tic, inv(r); toc, clear all

This simple command line program does the following: (1) Creates a 5000 x 5000 array of random floats; (2) Solves for the matrix inverse of this array. The tic and toc commands cause the time that the matrix inversion took to be displayed and the clear all command just removes all variables from memory.

With the T2400 the matrix inversion required 101 seconds. With the T7200 only 40 seconds was required. This problem is very demanding of CPU and memory but does not require any I/O operations. Both tests were performed with the same 32-bit operating system (Ubuntu 10.04). While solving for the inverse, both cores appeared to be utilized and were working 100% with both the T2400 and the T7200. I think this demonstrates that for certain problems a far greater speed improvement can be had with the core 2 duo then is normally expected.

I was quite surprised that this problem solved 2.5 times faster and am not sure how to account for this. The array required 200 MB of memory which is far too large to fit in the 4 MB cache of the T7200. But this is twice the size of the T2400 cache and being able to work on a larger chunk of the problem may account for the speed difference.

[ArtShapiro]

With the T2400 the matrix inversion required 101 seconds. With the T7200 only 40 seconds was required.

I was quite surprised that this problem solved 2.5 times faster.

How does going from 101 seconds to 40 seconds represent 2.5 times faster???

I did a similar upgrade (2400, I think, to 7600) and it certainly made my T60p a little peppier. I'm still not sure it was worth the $150 or whatever I paid a year or two ago, but as one can't purchase machines anymore with adequate vertical resolution I want to keep my T6x machines running as long as possible.

Art

[kurutta]

How does going from 101 seconds to 40 seconds represent 2.5 times faster???

We are talking about a ratio. So, 100/40 = 5/2 = 2.5 Right?

I am not saying that everything is 2.5 times faster. Just this particular problem that involves crunching a monster array with 25 million numbers. But I must say that I like my T60. It has the 1400 x 1050 IPS panel with the 4:3 ratio that god intended a display to have. It is interesting that market forces have caused laptops to become less useful. It appears that some have screens much wider then 16:9. A friend ordered a Lenovo with a i7 processor for some numerically heavy work and was disappointed because the inadequate vertical resolution made it difficult to get his work done.

[ArtShapiro]

How does going from 101 seconds to 40 seconds represent 2.5 times faster???

We are talking about a ratio. So, 100/40 = 5/2 = 2.5 Right?[/quote]

Whoops, I parsed that as 1:01! A thousand pardons. Glad it's working out so well.

Art

[Specter]

I did a similar upgrade (2400, I think, to 7600) and it certainly made my T60p a little peppier. I'm still not sure it was worth the $150 or whatever I paid a year or two ago, but as one can't purchase machines anymore with adequate vertical resolution I want to keep my T6x machines running as long as possible.

Hi!

This is the reason I keep my T60p (apart from loving it of course), the vertical resolution of reasonably priced ThinkPads is too low. It is strange that so few talk about this problem with wide screens of today. 700 or 800 vertical pixels just doesn't cut it in my opinion.

And by the way, the Matlab code finishes in about 25 seconds on my T60p (T7600). This is when using R2011b. Using R2007b on the other hand it takes about 51 seconds on the same computer. The difference was larger than I would have anticipated.

[dr_st]

Interesting test!

Takes between 35 and 40 seconds on my T60/T7200, with Matlab R2009b, which is consistent with results seen so far.

I don't have any other Core Duo / Core 2 Duo machines to test it on right now, so it's hard to say whether the big difference from T2400 is due to C2D efficiency or larger cache. I'd bet on the latter being more dominant.

On my desktop QX9650 overclocked to 3.6GHz, Matlab R2011a, it takes 9-10 seconds.

And by the way, the Matlab code finishes in about 25 seconds on my T60p (T7600). This is when using R2011b. Using R2007b on the other hand it takes about 51 seconds on the same computer. The difference was larger than I would have anticipated.

Indeed it's surprising. Have you monitored the CPU with R2007b to see that it's actually at 100% and not 50%? Perhaps it was running single-threaded? The following article suggests that multithreading for built-in functions may not be turned on by default until Matlab R2008a:
http://alenblog.wordpress.com/2011/03/3 ... in-matlab/

[Specter]

Indeed it's surprising. Have you monitored the CPU with R2007b to see that it's actually at 100% and not 50%? Perhaps it was running single-threaded?

You are right, I forgot about that! Indeed, CPU-utilization was about 50% when using R2007b with default settings. If I enable multithreading it went to 100% and - more importantly - the time taken decreased to about 29 seconds! Thanks for pointing this out!

[emeraldgirl08]

Thanks for posting this! I have an R60 coming and it has the T2400 so I'll be in the market for a T7200 soon. One thing I noticed while perusing EBay is that the prices are markedly higher for the T7200. Is it a highly sought upgrade for many laptops Sorry not hijacking thread but it is something I noticed as I browsed prices.

[Johan]

In regard to the speed increase in T60's when upgrading CPU's, there are quite mixed user reports as evident from the thread T60 CPU upgrade - currently a T2300 CD, can a T5200 C2D work and especially this post.

In short, the actual "speed increase" greatly depends of the actual usage - whether it is litle or much CPU intensitive. While e.g. Matlab as discussed above may run significintly faster, this improvement is not as evident if doing e.g. plain browsing etc.

Johan

[Banshee365]

It has the 1400 x 1050 IPS panel with the 4:3 ratio that god intended a display to have.

What size screen is that? Sounds like an SXGA+.

[kurutta]

I was thinking that when the array size became large enough that it no longer could fit in the L2 cache, the computation time would suddenly grow much larger. That is, for the T7200 there would be some obvious change in the trend around 4 MB. So wrote this small matalb script that does what my original script does -- measure the time a matrix inverse takes -- but do it many times with increasing array size. The script below goes all the way to 50 MB in steps of 100 kB. When it finishes it plots the results.

There was a noticeable step up in time around 10 MB (it moves around from run to run), but there was far less to see then I had expected. Of course my thinking on this was very unrealistic, as though the cache was not being used simultaneously for lots of other running processes as well. It probably would be a good idea to not have much else running when doing this test.

clear all
start_bytes = 100e3; step_bytes = 100e3; end_bytes = 50e6;
bytes = start_bytes:step_bytes:end_bytes;
n = round(sqrt(bytes./8));
for i = 1:length(n)
r = rand(n(i));
tic, inv(r); t(i)=toc;
fprintf(1, 'bytes = %d n = %d time = %d\n', bytes(i), n(i), t(i))
end
plot(bytes, t)

[netzspannung]

I went the same path with my T60-from t2500 to t7200, there was a noticeable difference... I need to use lightroom, so I liked that. Also I like the added psychological comfort of having a Core (2) architecture and SSE3.
However, my t61 with 7500 is still much faster, even with less RAM installed. There is only so much the T60 old chipset can do.