The water coolers don't take any real amount of wattage.

http://www.asetek.com/products/oem-standard-p...120mm-products/510lc.html is the one in my pc. 12 VDC and 2.6 watts. You just pop in the power connector to the cpu fan connection or any fan connection on the motherboard.

The real appeal about air cooling fans and water cooling meant for gaming rigs is that they keep things real quiet. I don't hear my pc running at full blast for example.

Stock cpu fans sound like a jet engine taking off when the cpu demand ramps up.

2.6 watts? That's it? I am surprised a water pump can be run at that wattage.

There is a very thorough post on the FX-8120 for crunching here:

http://www.worldcommunitygrid.org/forums/wcg/...ad,32104_offset,10#351221

@ PPL: When you look at the numbers for the FX remember that when BOIC does it's benchmark I believe that it "suspends" all cores and then runs the benchmark process on one core.

Why is this important? The FX architecture is based on two cores per module. Each core within a module has its own integer unit but shares one 256 bit floating point unit (which I understand can be divided into two logical 128 bit units)

The fact that the BOINC CPU baseline is run on only one core goes a long way towards explaining why FP performance on the test appeared to be better than expected when compared to integer performance.

However, in the "real world" of number crunching both cores within a module are always crunching and there is going to be some contention for that shared floating point unit.

Another thing that I see with my FX unit is that while it overclocks like a dream via the CPU multiplier jacking my 8120 from up 3.1 to 3.8 ghz did no make that much difference in the BOINC benchmark. However, bumping the FSB up by a conservative 10% to 220 resulted in a noticeable increase in total system throughput.

I have read that one of the real bottlenecks with the FX chip is the cache. That would make sense as tweaking the CPU multiplier is not going to do anything for cache speed but bumping the FSB up will.

I've only had my 8120 for a few days and am still learning about it, but so far it is kind of disappointing.

I don't think this is the case under Linux. There is a thread somewhere (I'll try to find it later today) in which benchmarks for machines with hyperthreading were compared under Windows and Linux. Linux benchmarks showed lower speeds when BOINC was set to run more threads than cores, whereas the Windows benchmarks stayed about the same regardless of how many threads BOINC was set to use. So my guess is that benchmarks under Linux time multiple cores and average over them, whereas benchmarks under Windows do what you wrote.

Benchmark results tend to be very different under Windows and Linux even for the same machine without HT. I have only one dual-boot machine (solely Linux on the others). It's a netbook with an AMD Fusion E-350 chip (no HT capability). After reading the comments about POEM being a good choice for making performance comparisons, I decided to run a few POEM WUs on both sides. Here's what POEM reports for benchmarks:

CPU type AuthenticAMD
AMD E-350 Processor [Family 20 Model 1 Stepping 0]
Number of processors 2
Coprocessors CAL ATI unknown (384MB) driver: 1.4.1016
Operating System Microsoft Windows 7
Home Premium x64 Edition, Service Pack 1, (06.01.7601.00)
BOINC version 6.10.58
Memory 3689.9 MB
Cache 512 KB
Measured floating point speed 1564.81 million ops/sec
Measured integer speed 2240.97 million ops/sec


CPU type AuthenticAMD
AMD E-350 Processor [Family 20 Model 1 Stepping 0]
Number of processors 2
Coprocessors ---
Operating System Linux
3.0.0-13-generic
BOINC version 6.12.33
Memory 3562.1 MB
Cache 512 KB
Measured floating point speed 1444.06 million ops/sec
Measured integer speed 5242.06 million ops/sec

The integer speed reported under Linux is more than twice as high as under Windows.

Two POEM WUs have completed and validated under Linux, averaging 9863 secs or 2.74 hrs. The first two are still running under Windows -- it looks as if they're going to take about 5 hrs. (Edit: It turned out that the average was 4.59 hrs.) So for this project, it looks as if the runtime difference is somewhat consistent with the benchmark difference.