$8000 to spend. 2 x E5-2687W, 16 real cores, 64Gb RAM

[JuLieN]

[moderation]
@Terry and Seb
Calm down, now. Just ignore each-others for a while.
[/moderation]

[diep]

None of the 2011 socket-E5 XEONS are overclockable as yet.

This is the heart of the problem.

Sujay we know how glad you help people out, but referring to oldie Xeons is no good idea except when they are dirt cheap;
the SR2 really is socket 1366, and he wants 8 core Xeons, called sandy bridge. Those need the EVGA SR-X Buying socket 1366 processors right now is a waste of money, except when that massively drops in price.

These exactly were my conclusion also, thank you for confirmations.

If I build a cluster per your specification, will I be able to also use it for gp (general purpose)? Please see original post in thread, thank you.

Well first of all clusters are not so easy to setup. Secondly you will need linux.
Sure on paper some will brag about how good it all works with windows,
but that's just ethernet. For performance you need infiniband.

Infiniband is what the supercomputers use nowadays as well to cheaply cluster.

I had confused your gp with gpu. Software doesn't work 'right out of the box' at a cluster.

Also this overclocking the guys speak about is not so easy to achieve, you go way outside of the specs. If you pretend to be scientist, you simply cannot overclock machines and expect scientific software to run well. It won't.

Most chessprograms have tons of money put into improvement of it, meanwhile having 0 sales. See how big the teams is and how hidden behind closeddoors their tuning is. Even the open source programs the tuner has not been released from. Not even from StockFish.

So on one hand some dudes really paid to get these guys to work, as they don't move an inch without getting paid - which explains why they suddenly were beating the old world top so bigtime.

You get what you pay for.

Clusters is a lot cheaper way to get big performance. Unrivalled performance if i may say so, yet programming them is not easy.

Most software won't work at a cluster.

On other hand if you are a scientist a cluster will give unrivalled performance in terms of testing.

On top of that it is possible to make a cluster version of engines.

Now we heard about Rybka, Toga and Jonny running on clusters already in 2008. It is 4 years later now. Where are those versions, as back then they DID get the extra search depth at the world c hamps 2008.

The search depths Toga got back then are like 5 plies deeper than what it got over here. Where are all those cluster versions?

My guess is: you won't see them any soon.

You can expect a Diep cluster version though. Will i sell it? I am not sure about that.

What i do know is this: for your scientific codes it is way way faster.

If you buy a machine only for a chessprogram produced by one of the rybka clone guys, who without fantasy didn't manage to profit much from newer hardware, yet did cash in money for adding 2 elopoints to it, and basically just has a well tested beancounter - i'd argue it's not worth for such thing to buy new hardware. These guys are not worth your attention.

I will not repeat the statement of Guy Haworth there, as some of these programmers felt really insulted when he said that - but it's the truth.

How about paying me or Marc Uniacke or Shay Bushinsky or Stefan Meyer-Kahlen or Fabien Letouzey or Frans Morsch?

Some of the psychopaths who had more information thanks to their job than the above mentionned programmers, they thought: "heh if i grab this code, paid by government no one cares, let's add 2 bytes then i can get bragging rights and attention and maybe cash". That's what happened.

Don't waste $8000 to these guys. If your scientific projects need big crunching power, go buy a simple cluster. It's cheaper and way faster.

Oh and it IS possible to write software that runs real fast on such cluster. Don't let yourself get fooled there by historic claims though of engines already claiming to have run on a cluster in reality hiding an expensive not yet released shared memory machine.

The proof of that last is easy of course - it would've been easy to get a cluster with hundreds of cores back then - yet they 'by accident' ran on an amount of cores that shared memory machines also had.

Rybka at 40 cores and toga at 24 cores and jonny at 16 cores.

By accident just the dunnington had released. Junior team ran on a 24 core shared memory dunning system. So probably rybka ran on similar hardware.

We saw some fake photo somewhere later on on the internet about rybka. Yet its search depth back then was so much deeper that it had to be a shared memory box back then.

What happened later was just cover ups and those cover ups it never got the search depths of back then.

You really want to buy a $8000 machine for that type of persons?

Go buy a real cluster and sit and wait until everyoen runs on one.
Now that i'm building diep for a cluster and have that cluster at home.
A cheap cluster, cheaper than the machine you're about to build,
realize they soon will throw in big cash and get ready for clustering as well.

That's how it usually works.
Usually someone behind scenes gets paid to do it, and then they all take that over.

The mellanox cluster i build here is cheap and will kick butt.

Vincent

[muxecoid]

I am the original poster. Didn't you find what muxecoid Michael Golub wrote is somewhat Bizarre?


NOTE FOR MODERATORS: Please not listen to Mr Terry McCracken who sends me PM implying he can fool moderators that "muxecoid" be an insult and so get me deleted or banned. "muxecoid" is the USERNAME of Mr Michael Golub. I do not know if it has good meaning, bad meaning or (according to dictionary) no meaning. Apology of Mr Julien Marcel is graciously accepted/

Sorry, just random trolling, no offense intended. Please forgive me.

[Dr.Wael Deeb]

I am the original poster. Didn't you find what muxecoid Michael Golub wrote is somewhat Bizarre?


NOTE FOR MODERATORS: Please not listen to Mr Terry McCracken who sends me PM implying he can fool moderators that "muxecoid" be an insult and so get me deleted or banned. "muxecoid" is the USERNAME of Mr Michael Golub. I do not know if it has good meaning, bad meaning or (according to dictionary) no meaning. Apology of Mr Julien Marcel is graciously accepted/

Sorry, just random trolling, no offense intended. Please forgive me.

http://www.youtube.com/watch?v=7x8wPt8xarE&ob=av2n

Sorry,couldn't resist