You can help me specify a new computer

[sje]

I estimated 200 W for the video...

I you don't use 3D, the graphic card use only around 20-30 watts.
http://www.hardware.fr/articles/817-3/resultats.html

Actually, all I'm using at the moment from the HD 7850 card is a single DVI port, connected via a converter to a VGA monitor.

The card's spec sheet says 130 W maximum load, so even a mildly out-of-spec card should still run under 150 W. Since I have Linux kill the video signal after 15 minutes idle, the card might be using only 5 W average.

The big question here is not the video card wattage, but rather the memory wattage. There's eight sticks of 2133 MHz DDR4 8 GiB RAM; the maximum allowed by both CPU and the mainboard.

In my 2006 Mac Pro I've got 28 GiB 667 MHz DDR2 FB RAM on eight sticks and these DIMMS run VERY HOT in spite of big heat sinks and generous air cooling. One pair of these 2 GiB sticks failed due to heat although they did work okay for nearly eight years -- if I put them back in, they would work, but only for a few hours.

Since I don't have a wattmeter, I had to rely on test results posted on the web, and that's where I got the 200 W number for CPU plus RAM load plus the WD Black 2 TB drive.

[sje]

Here's a number from a node on our new cluster (2x10cores, looks like 2.9ghz sustained:

time=1:01(94%) nodes=5226911714(5.2B) fh1=87% pred=8 nps=85.1M
checks=144.8M qchecks=280.1M fp=1.8B mcp=264.0M reversible=0
LMReductions: 1/118.5M 2/66.2M 3/27.8M 4/1.6M 5/13.9K 6/9
null-move (R): 3/151.1M 4/8.8M 5/172.2K 6/4.8K 7/48
splits=449.4K aborts=81.1K joins=1.8M data=39% tbhits=0

This is Crafty running 20 threads. It is scaling (NPS) almost perfectly (that 94% number). I am going to run some SMP speedup numbers but that will take at least a week to produce the data.

This is far from the fastest thing Intel makes:

model name : Intel(R) Xeon(R) CPU E5-2660 v3 @ 2.60GHz

E5-2660: http://www.cpubenchmark.net/cpu.php?id=2359
i7-5960X: http://www.cpubenchmark.net/cpu.php?id=2332

At US$1,500, that chip costs US$450 more than the i7-5960X. It's CPUMark figure is 1.5% higher than the i7-5960X.

http://www.newegg.com/Product/Product.a ... 6819117476

----

I'm going to run perft(12) on the new machine as part of further testing. I expect it to complete in under 60 hours. This will add about US$2.25 to my electric bill.

[wgarvin]

For general purpose computing you could put together a very reasonable system for under $2000. You only need an expensive graphics card if you are into PC gaming. I would recommend something like the i7-4790k, which is 4 cores at 4GHz, and overclockable: http://ark.intel.com/products/80807/Int ... o-4_40-GHz

Remember to buy the K model if you want to overclock it (the non-K models are locked). For motherboard, I'd suggest an ASUS z97i-plus mini-ITX.

[Edit: oops, I am a week too late. Anyway, looks like you settled on a good system.]

[Zenmastur]

You are right, usually I also want to have the power supply running at 50% of it's capabilities. It gives you the highest efficiency.

Really? I thought they were supposed to be more efficient the closer you got to full load.

The efficiency drops on both ends of the curve, most power supplies have a sweet spot somewhere.
When you look at the power supply I use at the moment it's highest efficiency is exactly at 50% of it's maximum load.

http://www.anandtech.com/show/8835/cors ... y-review/4

You seem to be overlooking one important fact. If you add up the maximum power draw of all components you will end up with a power draw that can't be reached no matter what programs you are running.

Example: Max CPU draw 140 W, Max Video draw 125 W, Max HD+SDD+Optical Draw 25 W, Max all other peripherals 25W. The CPU can draw more power than it's rating due to power conversion inefficienties so 140 / .75 = ~187 watts. Maximum theoretical power draw is around 362 W. It's almost impossible for this hardware to draw this much power, assuming no over-clocking and no over-volting. The reason is that the main draw in the CPU is the floating point unit so it must be kept fed. The video card must be fed as well. It's difficult to keep both units fed with data unless you have an execeptional memory subsystem. It's even harder to maintain such a high power draw for any length of time.

So what usually happens is either the video card is maxed and the CPU is at about half draw or the CPU is maxed and the video card is at a quarter draw or less. During either of these the drives are usually idle and the fans are at about 75% power draw. This yields max sustained power draw of around 240 W. Even this high a power draw is unusual for most systems.

So specifying a 750 to 850 watt power supply not only wastes money on an overly large power supply it also wastes enegy due to poor efficiency. It's not possible to reach even a 50% PSU load with such a setup. So you are always operating at less than 50% (i.e. 28%-32% max load) and some of the time you will be operating at 10% or less of design power output. Not a wise choice. Unless you know specifically what a particular set of software will draw when running and the sole purpose of the machine is to run this software its a costly mistake to spec a power supply like this.

Regards,

Zen

[sje]

Excessive heat over time is what kills components in a power supply.

With respect to heat, poor efficiency at lower output power is much worse than the same at higher output power.

An incandescent light bulb which lasts a year at standard input voltage might last a decade at 50% voltage, but only a few hours at 150% voltage.

[Vinvin]

...
The big question here is not the video card wattage, but rather the memory wattage. There's eight sticks of 2133 MHz DDR4 8 GiB RAM; the maximum allowed by both CPU and the mainboard.

In my 2006 Mac Pro I've got 28 GiB 667 MHz DDR2 FB RAM on eight sticks and these DIMMS run VERY HOT in spite of big heat sinks and generous air cooling. One pair of these 2 GiB sticks failed due to heat although they did work okay for nearly eight years -- if I put them back in, they would work, but only for a few hours.

From http://en.wikipedia.org/wiki/DDR4_SDRAM

DDR4 operates at a voltage of 1.2 V with a frequency between 1600 and 3200 MHz, compared to frequencies between 800 and 2400 MHz and voltage requirements of 1.5 or 1.65 V of DDR3. Although a low-voltage standard has yet to be finalized (as of August 2014), it is anticipated that low-voltage DDR4 will run at a voltage of 1.05 V, compared to DDR3's low-voltage standard (DDR3L) which requires 1.35 V to operate.

And DDR2 operated at 1,8 Volts.
So, RAM is less energy consuming now.

[Joost Buijs]

You are right, usually I also want to have the power supply running at 50% of it's capabilities. It gives you the highest efficiency.

Really? I thought they were supposed to be more efficient the closer you got to full load.

The efficiency drops on both ends of the curve, most power supplies have a sweet spot somewhere.
When you look at the power supply I use at the moment it's highest efficiency is exactly at 50% of it's maximum load.

http://www.anandtech.com/show/8835/cors ... y-review/4

You seem to be overlooking one important fact. If you add up the maximum power draw of all components you will end up with a power draw that can't be reached no matter what programs you are running.

So what usually happens is either the video card is maxed and the CPU is at about half draw or the CPU is maxed and the video card is at a quarter draw or less. During either of these the drives are usually idle and the fans are at about 75% power draw. This yields max sustained power draw of around 240 W. Even this high a power draw is unusual for most systems.

So specifying a 750 to 850 watt power supply not only wastes money on an overly large power supply it also wastes enegy due to poor efficiency. It's not possible to reach even a 50% PSU load with such a setup. So you are always operating at less than 50% (i.e. 28%-32% max load) and some of the time you will be operating at 10% or less of design power output. Not a wise choice. Unless you know specifically what a particular set of software will draw when running and the sole purpose of the machine is to run this software its a costly mistake to spec a power supply like this.
Zen

I agree with you that in practice you will never reach the calculated load.
On the other hand these 'Platinum' PSU's are so good that even at 10% of their maximum load they still have a higher efficiency than those cheap '80+' labeled PSU's.

When you build a machine which cost you >$2k you want to have reliable components in it.
The $100 difference in cost between a crappy PSU and a very good one really doesn't matter, at least not to me.

[Zenmastur]

You are right, usually I also want to have the power supply running at 50% of it's capabilities. It gives you the highest efficiency.

Really? I thought they were supposed to be more efficient the closer you got to full load.

The efficiency drops on both ends of the curve, most power supplies have a sweet spot somewhere.
When you look at the power supply I use at the moment it's highest efficiency is exactly at 50% of it's maximum load.

http://www.anandtech.com/show/8835/cors ... y-review/4

You seem to be overlooking one important fact. If you add up the maximum power draw of all components you will end up with a power draw that can't be reached no matter what programs you are running.

So what usually happens is either the video card is maxed and the CPU is at about half draw or the CPU is maxed and the video card is at a quarter draw or less. During either of these the drives are usually idle and the fans are at about 75% power draw. This yields max sustained power draw of around 240 W. Even this high a power draw is unusual for most systems.

So specifying a 750 to 850 watt power supply not only wastes money on an overly large power supply it also wastes enegy due to poor efficiency. It's not possible to reach even a 50% PSU load with such a setup. So you are always operating at less than 50% (i.e. 28%-32% max load) and some of the time you will be operating at 10% or less of design power output. Not a wise choice. Unless you know specifically what a particular set of software will draw when running and the sole purpose of the machine is to run this software its a costly mistake to spec a power supply like this.
Zen

I agree with you that in practice you will never reach the calculated load.
On the other hand these 'Platinum' PSU's are so good that even at 10% of their maximum load they still have a higher efficiency than those cheap '80+' labeled PSU's.

When you build a machine which cost you >$2k you want to have reliable components in it.
The $100 difference in cost between a crappy PSU and a very good one really doesn't matter, at least not to me.

I think you are confusing size with quality. There not the same thing. Quality is determined by design, component selection, output waveform, response to load changes and other factors too numerous to mention here.

Buying "big" doesn't address many of these issues and in fact may cause some of them (response to load changes and output ripple) to be worse. There is a proper way to calculated the required size and it doesn't involve throwing money at the problem in the hopes that it will be resolved.

Regards,

Zen

[Joost Buijs]

Most of the lower powered PSU's are also less well designed because they are meant for an average consumer PC.
Often they use 2nd grade Chinese electrolytic capacitors which are certain to fail within a couple of years.

I'm not confusing size with quality, I have an education in electronics so I know exactly what you are talking about.

Anyway let everybody decide for themselves where they want to spend their money on.

[bob]

Here's a number from a node on our new cluster (2x10cores, looks like 2.9ghz sustained:

time=1:01(94%) nodes=5226911714(5.2B) fh1=87% pred=8 nps=85.1M
checks=144.8M qchecks=280.1M fp=1.8B mcp=264.0M reversible=0
LMReductions: 1/118.5M 2/66.2M 3/27.8M 4/1.6M 5/13.9K 6/9
null-move (R): 3/151.1M 4/8.8M 5/172.2K 6/4.8K 7/48
splits=449.4K aborts=81.1K joins=1.8M data=39% tbhits=0

This is Crafty running 20 threads. It is scaling (NPS) almost perfectly (that 94% number). I am going to run some SMP speedup numbers but that will take at least a week to produce the data.

This is far from the fastest thing Intel makes:

model name : Intel(R) Xeon(R) CPU E5-2660 v3 @ 2.60GHz

E5-2660: http://www.cpubenchmark.net/cpu.php?id=2359
i7-5960X: http://www.cpubenchmark.net/cpu.php?id=2332

At US$1,500, that chip costs US$450 more than the i7-5960X. It's CPUMark figure is 1.5% higher than the i7-5960X.

http://www.newegg.com/Product/Product.a ... 6819117476

----

I'm going to run perft(12) on the new machine as part of further testing. I expect it to complete in under 60 hours. This will add about US$2.25 to my electric bill.

They also run these up to 18 cores and 45mb L3.

We also have a "phi" board in this box. 61 cores of what is basically an intel pentium processor (in order execution). I have not yet done any testing on that as our tech guys apparently broke the software with an upgrade. But Linux runs on it and it looks like a "computer within the computer." I log into this 20 core box, then ssh to the 61 core box, which seems strange, but seems to work.

More later. There might be some opportunities here for additional shenanigans, such as using the phi cores to do evaluations and such sort of "belle-like"