CPU vs GPU DC LL effiency

[nucleon]

Over the last week or so, I did the same work on a Titan GPU vs CPU and wanted to see how it went. The results below is for LL DC work.

Summary figures:

Intel Ivy-Bridge CPU @ 3.1GHz quad core mini ITX system- 28GHz-day/day 100W
Titan GPU 43.5Ghz-day/day 200W

Power figures approx. GPU Shark reports power consumption as percentage points of TDP. Titan TDP is 250W, percentage reported was 80%. CPU power consumption was measured at the wall. Which takes in consideration power supply inefficiencies. Not exactly apples-to-apples comparison. To do it properly, requires more time. :)

I know it's not exactly scientific. I wrote these figures down, but didn't write my calculation method.

I'm interested to see how the NUC Haswell systems fair. A quad core Haswell in NUC form factor could be very cool.

My back of the envelope recommendation at the moment, is a low-clocked quad core, mini ITX motherboard with 1600MHz ram. (low voltage RAM if you can do it).

Multiple systems if you want more performance.

Unless you want to do GPU TF.

I'm power saturated in my apartment at the moment, so the only way I can grow is go more efficient.

-- Craig

[henryzz]

For consumer chips I would imagine sandy bridge-e with 6 cores would be a strong contender.
A ivy bridge-ex server machine would do even better when they come out.

[nucleon]

Originally Posted by henryzz

For consumer chips I would imagine sandy bridge-e with 6 cores would be a strong contender.
A ivy bridge-ex server machine would do even better when they come out.

That's if you want a single machine output. I'm talking performance per watt.

On the top end - I'm keen for Haswell-e with DDR4 see if it resolves the memory saturation we're seeing; for SB-E it was 4 from 6 cores, for SB/IB it was 3 from 4 cores.

A lower clocked quad core, not only is more power efficient. (Lower clocked have generally have lower voltages) But if the ram is held at the higher speed - memory saturation considerations are more favorable.

-- Craig

[science_man_88]

Originally Posted by nucleon

Over the last week or so, I did the same work on a Titan GPU vs CPU and wanted to see how it went. The results below is for LL DC work.

Summary figures:

Intel Ivy-Bridge CPU @ 3.1GHz quad core mini ITX system- 28GHz-day/day 100W
Titan GPU 43.5Ghz-day/day 200W

Power figures approx. GPU Shark reports power consumption as percentage points of TDP. Titan TDP is 250W, percentage reported was 80%. CPU power consumption was measured at the wall. Which takes in consideration power supply inefficiencies. Not exactly apples-to-apples comparison. To do it properly, requires more time. :)

I know it's not exactly scientific. I wrote these figures down, but didn't write my calculation method.

I'm interested to see how the NUC Haswell systems fair. A quad core Haswell in NUC form factor could be very cool.

My back of the envelope recommendation at the moment, is a low-clocked quad core, mini ITX motherboard with 1600MHz ram. (low voltage RAM if you can do it).

Multiple systems if you want more performance.

Unless you want to do GPU TF.

I'm power saturated in my apartment at the moment, so the only way I can grow is go more efficient.

-- Craig

using the quoted Ghz days/day figure I get the titan is roughly 156% of the GHz days/day versus the ivy, but when you do it with per watt it's about 78% as efficient per watt.