OT: If you could build a HUGE parallel processing supercruncher what would you use?

[InlineFive]

1. Even though I'm an AMD Fanboy I think I would go with lots of HT P4s because they simply do better.
2. I don't know what type of connection would be powerful enough to provide bandwidth to make parallel processing worth it.

-Por

 

[mikecel79]

I think I might use some big iron and go with Itanium 2s (9MB cache) and connect all the machines using Infiniband or 10Gigabit Ethernet.

 

[zodder]

I would definitely use somebody elses money!

 

[kamper]

There were those co-processor thingy's I read about a while ago. Don't remember what they were called but they ran @ 200MHz with 64 concurrent threads. They only used something like 4watts. I'd look into integrating a bunch of those onto a board with a single regular processor. No P4's though.

Oh, and: what Zodder said

 

[MournSanity]

Opterons.

 

[lobadobadingdong]

Originally posted by: hypersonic5
Opterons.

indeed, they have 4 way blades out now, you can load up a mess of blade servers for some simply massive crunching power.

 

[titanmiller]

1. I would probably lean towards HT P4s even though I tend to be somewhat of an AMD fanboy, those P4s just put out so much/dollar.
2. From my calculations a (decent) 24/7 Dial-up connection could handle 1500 WUs/day, so unless you really have a super cruncher almost any connection could work.

 

[InlineFive]

Originally posted by: lobadobadingdong

Originally posted by: hypersonic5
Opterons.

indeed, they have 4 way blades out now, you can load up a mess of blade servers for some simply massive crunching power.

That sounds like a solid solution. Keep in mind I said COULD so I don't actually have the cash for this stuff.

-Por

 

[RemyCanad]

Power5.

 

[n0cmonkey]

1. POWER 5, mips64, or UltraSPARC4. IBM, SGI, Sun, and Fujitsu (sparc64) have already gotten through big problems. SGI has done some amazing things with hundreds of processors, and the SunFire 10k is just something to drool over.
2. Locally or to the net? Locally it depends on what you are crunching. myrinet(sp?) would be nice.

 

[InlineFive]

Originally posted by: n0cmonkey
1. POWER 5, mips64, or UltraSPARC4. IBM, SGI, Sun, and Fujitsu (sparc64) have already gotten through big problems. SGI has done some amazing things with hundreds of processors, and the SunFire 10k is just something to drool over.
2. Locally or to the net? Locally it depends on what you are crunching. myrinet(sp?) would be nice.

It would (imagination kicking in) be a huge complex comprimised of at least 1Thz of computers configured in multiple "nodes". So each node could be assigned a different task. Then you could allocate these to different projects or put a bunch of them on one project to give it a boost.

-Por

 

[bgeh]

i would say Opterons since they scale much better then Xeons

 

[n0cmonkey]

Originally posted by: PorBleemo

Originally posted by: n0cmonkey
1. POWER 5, mips64, or UltraSPARC4. IBM, SGI, Sun, and Fujitsu (sparc64) have already gotten through big problems. SGI has done some amazing things with hundreds of processors, and the SunFire 10k is just something to drool over.
2. Locally or to the net? Locally it depends on what you are crunching. myrinet(sp?) would be nice.

It would (imagination kicking in) be a huge complex comprimised of at least 1Thz of computers configured in multiple "nodes". So each node could be assigned a different task. Then you could allocate these to different projects or put a bunch of them on one project to give it a boost.

-Por

I thought we were talking about parallel computing.

We're going for versatility here. Easy answer for me, I'm enough of a geek I've been thinking about how I would do that.

Dual Opteron blades, connected between each other using 100mbit connections with a controller node (quad opteron) on gigabit. Run linux for the server features that I just can't find anywhere else, with WINE installed to run clients that are only out there for Windows. You could even setup some nice controls on the main system to control the rest.

 

[ProviaFan]

Originally posted by: n0cmonkey
Dual Opteron blades, connected between each other using 100mbit connections with a controller node (quad opteron) on gigabit. Run linux for the server features that I just can't find anywhere else, with WINE installed to run clients that are only out there for Windows. You could even setup some nice controls on the main system to control the rest.

Sounds like a plan!

Unfortunately, this is one of those "if you have to ask the price, you can't afford it" things.

 

[n0cmonkey]

Originally posted by: jliechty

Originally posted by: n0cmonkey
Dual Opteron blades, connected between each other using 100mbit connections with a controller node (quad opteron) on gigabit. Run linux for the server features that I just can't find anywhere else, with WINE installed to run clients that are only out there for Windows. You could even setup some nice controls on the main system to control the rest.

Sounds like a plan!

Unfortunately, this is one of those "if you have to ask the price, you can't afford it" things.

Yeah, I checked one of the banners here on the forums. You have to ask for a price quote. In IT/technical terms that means: Start playing the lotto.

 

[Dantoo]

If you could build a HUGE parallel processing supercruncher what would you use?

Somebody else's money!

But it might look like this:
Par-kut at BBR?

or if I was a Mac G5 fan maybe like this:
Terrascale Cluster

maybe if I was into Itaniums
MSCF-MPP2

or Pentium 4's
TACC

Opterons do a great job if you get enough - say 2,816:
Los Alamos ASCI - Opteron

Alpha Fans might be tempted by this:
ASCI-Q

If everybody sent in their old Celeron 300a's something like this might arise:
Digital Darwin

In the meantime if you want to get a cluster going, this is probably the cheapest way:
The Basics

 

[n0cmonkey]

Darn, forgot about Alphas! I wish I had one

 

[grendel8]

mmmm, Alpha's

Great post Dantoo! :beer:

 

[xjamesx]

Fantastic links, thanks Dantoo

I can't even begin to imagine the processor power of those clusters:Q