We don't care about power usage or heat dissipation so long as the thing is reliable. We are using "in-row" cooling where the cluster is in an air-tight box with the freon evaporator cores and fans built into the racks themselves, so that we don't have to provide a frigid room temp to keep cluster heat down. In fact, I have never seen a HPC center give any consideration to power consumption. Most already have a couple of hundred tons of A/C cooling so the power consumed by the machine itself is a small part. The HPC guys care about performance. The old ECL hardware burned power like there was no tomorrow, but everyone was happy with the speed. On a chip, the only constraint I care about is can they dissipate the heat safely and reliably without exotic steps such as those used in older Crays for cooling.diep wrote:Intel has a major problem here of course. they have a great 2 socket cpu now, but it eats way too much power for server farms and their lower clocked E-models as you already indicated have bios problems and achieve very little compared to the power wasting X models that auto overclock 400Mhz to 600Mhz.bob wrote:All I can say is we have been in negotiation with Dell and Intel for a large cluster (2048 cores or more). We were told in a meeting with Intel that the MP I7 would be out late this year. The nodes we are looking at are dual-socket 8-core i7 based...diep wrote:Well intel is already bragging around about the Xeon MP for 3 years now.bob wrote:The xeon-MP version of the I7 is due out within a couple of months according to Intel. I'll take that. I've run on a 4-way prototype a short while back, and with the coming 8 cores and 4 sockets, that will be one fast box.tano-urayoan wrote:Hello Dr. Hyatt.bob wrote:I'd take the quad-core i7, currently at about the same speed clock-wise. When the xeon-MP version comes out later this year, it won't be long until we see 4 socket 32 core machines that will be beyond fast.Sean Evans wrote:Bob, would you prefer Crafty to run on an Intel i7 OR Intel Xeon W5580 @ 3.2GHzbob wrote: Fastest I7 (Intel) processor he can find, with the most cores he can afford.
Which will be your selection, quad i7 or a dual W5580 system?
Now it should release in end 2010 then in end 2009.
I"m guessing end 2010.
by then of course you also could buy for a part of that money a 96 core AMD box.
What's faster 32 core intel or 96 core AMD?
Vincent
Whether that will actually happen or not (MP later this year) is just speculation until we see the chips.
Crafty they have advantage of having optimized it for their compiler for 10+ years now, so for crafty it might seem like a great chip maybe.
for deepsjeng and diep the ipc of the AMD is not far away from the nehalem. it's a few percent difference, which is all caused by compiler, not the actual hardware.
for 4 socket line, intel has a major problem as AMD already for many years is selling 4 socket machines. Took AMD a long time to release a quad core and initially it was really power wasting and low clocked.
Intel until today has nothing to compete iwth the low power 4 core AMD server cpu's. The only thing intel has, is something on big stereoids of 700mm^2 @ 300 watt a cpu or so (TDP is 170 watt by the way).
Now for you and me that's mighty interesting to have 1 such box, but for the HPC world in itself it's just a waste of power.
On other hand, AMD released already 24 core box which is a low power cpu compared to all those intels and amd is busy releasing a 12 core chip also.
Then there is 4 socket mainboards that glueless connect to another 4 socket mainboard. Maybe crafty won't run fine on that, but diep sure does, as it can handle those latencies very well.
That's 96 cores. Without glue, it's a 48 core box and it will be released in time.
With that Xeon MP that's not so sure, they would release it earlier this year, then it was delayed another 1.5 years. I bet they release it half a year before cpu's actually get sold of it.
So be sure if you sign a contract that there is a penalty for each day they deliver the cpu's later.
My government got severely dicked by intel there with the itanium2 cpu there. HPC hardware you HAVE to deliver in time, or you're out of the game. ntel has disqualified itself there.
Also the price is unclear to me.
I'm not sure whether i'm allowed to quote expected yields of those cpu's. I bet they prefer to nuke me instead of even a laymen guess, but those cpu's will be really expensive to print and verify.
many bugs WILL get found at such a huge power usage in those cpu's i predict. That's intel price to seemingly be 'fastest' at the 4 socket domain for a few months.
In reality it's one big AMD show all the way. For the power usage of a single 4 socket intel box you can run probably 3 of those AMD machines.
At introduction of their opteron cpu for 2 socket and 4 socket domain in 2003 and 2004 it was, AMD took over 50% of server market for a short while. That's really a big deal. Intel is only on paper faster now, thanks to overclocking and throwing more power budget against it, whereas the power usage by now is approaching DEC alpha usage.
That's really bad IMHO.
I doubt intel can deliver that Xeon MP against a competative price.
The 24 core machine has a price of $2300 a cpu.
If you'd assemble it yourself you can get it handsdown for $2300 from amd.
probably a lot less if you order a lot of those cpu's.
you're speaking of 512 there. If you ship an email to AMd: we want to order a cluster and are interested in a bid of 6 core opteron cpu's @ 2.6Ghz that can be put in quad socket mainboards, and we won't publish what we paid for it, if you do a real good bid.
I bet istanbul should be possible to get for half the price @ 2.6Ghz a cpu.
they can use the sales. they know that an university like yours will normally spoken only order intel hardware. You really can get a discount of 50% there i bet. Matter of good negotiation. The rest is real cheap. Those mainboards are real cheap and so on. those cpu's are majority of the cost of such a system.
Go for it Bob.
Vincent
p.s. NASA @ intel usually negotiates even tougher. That 10240 processor box they got the cpu's nearly at production price for is rumours of some years ago.
these are fantastic times for negotiations if you have a wallet filled with cash, most manufacturers you can get everything half price now.
The BIOS bug we found was a Dell issue, as Dell writes the bios code for their systems. And it had a simple bug that was fixed pretty quickly. It was overclocking when it should not, and was not overclocking when it should. But the interesting part was, overclocking with all 8 cores running full speed did not cause any heat problems, and I ran the thing with Crafty non-stop for 24 hours monitoring the CPU temps with lm_sensors on linux.