Bargain Hardware Thread

[Xyzzy]

It would be neat to use a bunch of water-cooled video cards to heat a home's water supply.

[Mark Rose]

Originally Posted by Xyzzy

It would be neat to use a bunch of water-cooled video cards to heat a home's water supply.

Shouldn't be any harder than leaving the computers in the hypocausts.

[kladner]

Originally Posted by Xyzzy

It would be neat to use a bunch of water-cooled video cards to heat a home's water supply.

The water plan would probably work best with a heat exchanger coil in a preheat tank. This could then be fed to a conventional water heater, which would not have to work as hard. If you wanted to get fancy, the heat exchanger could be part of a circulating system, which would probably extract more heat than a passive setup.

[paulunderwood]

I know next to nothing about GPUs, but would these Tesla M2090 be of interest at $145

Or this one for $160?

[chalsall]

Originally Posted by paulunderwood

I know next to nothing about GPUs, but would these Tesla M2090 be of interest at $145

I often rent two (2) M2050s in a AWS cg1.4xlarge instance for about $0.14 an hour (~ $100 a month) and get about 280 GHz Days / Day out of each. So, definitely, at $145 each these would be attractive to someone who didn't spend as much as I do on local electricity!

[Mark Rose]

I can get used GTX 580's for CA$100 locally that trial factor at about 430 GHz-d/d at roughly the same wattage and include built in cooling.

They might be a better deal for applications requiring double precision floating point though.

[bgbeuning]

I am looking at buying some machines to crunch on prime stuff.
There are lots of options, from ebay used machines, to off the
shelf machines, to build your own machines. I have been looking
for a scoring method to plug in some numbers to pick the best
machines, where best is most crunching for least money.

The scoring should use the prime95 benchmark (p=70M) pages,
the number of cores in the machine, and the price.
The formula

score = benchmark / cores * cost

where a lower score is better sounds reasonable to me.

Lets try a couple of examples.

"HP Desktop Computer Z210 XEON E3-1240 (3.30 GHz) 4 GB DDR3 250 GB HDD"
http://www.newegg.com/Product/Produc...82E16883282047

benchmark = 23.21
cores = 4
price = $295
score = 1711

"Dell C6100 XS23-TY3 Server 8x 2.26GHz 4C E5507 96GB"
http://www.ebay.com/itm/Dell-C6100-X...8AAOSwl9BWJk8R

benchmark = 66.8 (closest I could find was Intel Xeon E5462)
cores = 32
price = $790
score = 1649

Build your own i5-6400

benchmark = 17.91
cores = 4
price = $538 (CPU $190, MB $100, RAM $108, case = $40, PS = $60, cooler = $40)
score = 2847

So do you think this method of scoring is valid?
Show me an example where the score does a bad job of assessing a system.
How do you score systems?

[axn]

Originally Posted by bgbeuning

The formula

score = benchmark / cores * cost

where a lower score is better sounds reasonable to me.

<snip>

So do you think this method of scoring is valid?
Show me an example where the score does a bad job of assessing a system.
How do you score systems?

As per your formula, higher benchmark number is bad. Is that intended?

EDIT:- By benchmark, you mean iteration times in ms? If so, then it is fine. But probably the inverse calculation would have been more intuitive. i.e Cores/Iteration time/$ = Thruput/$

For cost, you might look at purchase cost + cost of electricity for running for, say, 2 years. That will factor in power efficiency as well.
Also, for multicore systems, the scalability might not be great. You might not get the same iteration times when more cores are running.

[retina]

Somewhere in there you need to include the RAM size and speed. The RAM that is installed affects the benchmark greatly. It is unfortunate that the benchmarks page does not include RAM figures. So perhaps you can consider looking to other places on this board to find where people have detailed the RAM specs and adjust your benchmark values accordingly.

[LaurV]

Originally Posted by axn

As per your formula, higher benchmark number is bad. Is that intended?

he said "where a lower score is better", so yes, it was intended.

Originally Posted by axn

EDIT:- By benchmark, you mean iteration times in ms?

I think he means the benchmark on PrimeNet server benchmark list. The line "closest I could find was Intel Xeon E5462" supports this assumption.

Originally Posted by axn

probably the inverse calculation would have been more intuitive. i.e Cores/Iteration time/$ = Thruput/$

For cost, you might look at purchase cost + cost of electricity for running for, say, 2 years. That will factor in power efficiency as well.
Also, for multicore systems, the scalability might not be great. You might not get the same iteration times when more cores are running.

This is all correct. There are lots of things to consider, not the last is the memory bottlenecks and the mobo chipset, cooling, electricity expenses, etc. (i.e. your performance may not scale when you use more cores).

[VBCurtis]

Originally Posted by bgbeuning

"Dell C6100 XS23-TY3 Server 8x 2.26GHz 4C E5507 96GB"
http://www.ebay.com/itm/Dell-C6100-X...8AAOSwl9BWJk8R

benchmark = 66.8 (closest I could find was Intel Xeon E5462)
cores = 32
price = $790
score = 1649

This system looks very very interesting for NFS factoring work. Nice find! If I had any confidence at setting up the 4 individual nodes and remotely managing it all, I'd be inclined to buy one.