paul-
14 computers and you want more? wow!
your going to have to get a second electric line run to your house to service all those beasts. I would think your not going to have to do much heating of the house with all those systems blowing out hot air.

With 14 systems, heat and electricity are an issue. My electric bill typically hits $500 per month in the summer and we have a geothermal HVAC system. I expect the bill to fall this winter.

I would like to get more AMD or Core 2 systems into the farm and move some of the P4s to the garage. The P4s consume way too much power and generate too much heat. This was Intel's biggest failing with that family of processor. When I move them to the garage, they will have limited environmental control so they will likely die over time.

Thanks to everyone for all the help. The cure is out there.

---

Thx!

Paul

The AMD vs Intel issue generally seems to relate to the amount of L2 cache on the CPU. Rosetta's application is much larger then many other BOINC projects. This means much of it is stored in L2 cache and so systems with small L2 caches are constantly running out to memory to get subroutines. This slows overall efficiency.

If this is the reason, why AMD cpus are slowed down so massivly, then we will see Intel´s new Nehalem failing on rosetta aswell - maybe even more then AMD. Nehalem comes with only 256k L2 Cache, hence they introduce the same memory controller thechnology which AMD is using since the K8 / AMD 64 series.

The L2 as a reason to why AMD is so bad in Rosetta is hard to accept, tho I dont see any other reason. I have 2 pretty similar CPUs, a 4000+ and a 4200+, they differ by 200 MHz in clock speed, and the 4000+ L2 cache is 1024kb vs 512kb on the 4200+. Now in Rosetta, the 4000+ can get 45~50 creds for 3 hours, while the 4200+ never goes near that (around 30), when im working on this machine, I get horible results like 12 credits for 3 hours. This would also support the L2 problem argumentation.

Well, lets wait for the first Nehalems to crunch Rosetta then :)

---

http://www.MIAteam.eu

The AMD vs Intel issue generally seems to relate to the amount of L2 cache on the CPU. Rosetta's application is much larger then many other BOINC projects. This means much of it is stored in L2 cache and so systems with small L2 caches are constantly running out to memory to get subroutines. This slows overall efficiency.

If this is the reason, why AMD cpus are slowed down so massivly, then we will see Intel´s new Nehalem failing on rosetta aswell - maybe even more then AMD. Nehalem comes with only 256k L2 Cache, hence they introduce the same memory controller thechnology which AMD is using since the K8 / AMD 64 series.

The L2 as a reason to why AMD is so bad in Rosetta is hard to accept, tho I dont see any other reason. I have 2 pretty similar CPUs, a 4000+ and a 4200+, they differ by 200 MHz in clock speed, and the 4000+ L2 cache is 1024kb vs 512kb on the 4200+. Now in Rosetta, the 4000+ can get 45~50 creds for 3 hours, while the 4200+ never goes near that (around 30), when im working on this machine, I get horible results like 12 credits for 3 hours. This would also support the L2 problem argumentation.

Well, lets wait for the first Nehalems to crunch Rosetta then :)

Nehalems will have (up to) 2MB L3 cache per core - shared like Barcelona/Phenom - beyond the 256Kb per core (although it's inclusive so it's really 1.75MB per core). The AMD chips aren't bad a rosetta, but they have a strong FPU so they get a high benchmark, so the granted credits are rarely that high. Core2 is just newer and better suited to Rosetta.

---

I think the L3 cache is pretty useless here. I was just looking around for some scores from Barcelona/Phenoms, who have this L3 cache also and i found some numbers:

AMD 8358 SE (thats a 2,4 GHz Barcelona quad) goes from 34 to 39 credits/3 hours link
AMD Phenom x4 9550 ( 2,2 GHz Phenom Quad) is in the same rage, with some hitting around 25 credits
link

so on the x4 9550 I see pretty much the same numbers like on my x2 4200+, the run the same clock speed, the same cache, the L3 cache seem to make no difference nor the newer K10 technology.

The 2,4 GHz Barcelona isnt pulling away either, so core-per-core I dont see an AMD with 512 kb cache thats going passed my old 1024 kb L2 cache 4000+.

Now when you look on an Intel CPU with loads of L2 Cache ( I was looking for a lower clocked C2D, couldnt find one so fast) , like this E 8600 the CPU gets around 3 times the credits then any AMD out there is able to pull.

Now Im gonna search for an new Intel with low L2 cache, anyone able to find an
Pentium Dual-Core E2220 or E2200 ? they have only 512 kb cache aswell, I wouldnt be surprised to see them get ~30 credits for 3 hours aswell, then the L2 thing would be proven I think :)

---

http://www.MIAteam.eu

bingo :) Intel(R) Pentium(R) Dual CPU E2220 @ 2.40GHz

---

http://www.MIAteam.eu

This is an E4300 which has 1MB per core (it's running at 2.4GHz): https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=492052
and this one is an E2180 with 512KB per core (2GHz): https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=659985

---

Thx for the infos, that should be enough to make conclusions.
I just had a look over at QMC@home and looked for E4xxx and then E6xxx/E8xxx CPUs, and when you find CPUs running at the same clock speed, they earn the same credit, unlike here. So L2 cache really makes the difference here.

My conclusion would be: If you plan to build system only for Rosetta, this doesnt leave many CPUs to chose from, The Intel E8xxx series or the new Quads with big L2 (Q9450 and higher).

Just wondering if even more L2 cache then those CPUs have would help even more.

---

http://www.MIAteam.eu

A couple of random thoughts on the topic.

IME quad cores are far cheaper overall than dual cores. Mostly because you get to amortize a single mobo and case across more cores. I allocate 512 MB ram / core on my crunch boxes, so RAM comes out a wash.

However, power consumption is the final killer for dual cores. I bought a kill-a-watt meter from ebay, it has been one of the best investments I ever made.

In particular, it reports that my dualies consume 80w when running, while the quad cores only take 110w. If you do the math, that's 40w/core vs 28w/core, a clear win for the quads.

That also means you do not want to pony up a bundle of cash for a case with a 400w PSU, you're just not going to use it. I use these cases from ebay. 200w PSU, and $25 delivered to my door. Since the kill-a-watt says I'm only taking 110w, 200w is plenty enough.

If you're going the Linux route, take a look at how to use pxe to boot them over the network - just about all modern mobos will support a network boot.

---

A couple of random thoughts on the topic.

IME quad cores are far cheaper overall than dual cores. Mostly because you get to amortize a single mobo and case across more cores. I allocate 512 MB ram / core on my crunch boxes, so RAM comes out a wash.

However, power consumption is the final killer for dual cores. I bought a kill-a-watt meter from ebay, it has been one of the best investments I ever made.

In particular, it reports that my dualies consume 80w when running, while the quad cores only take 110w. If you do the math, that's 40w/core vs 28w/core, a clear win for the quads.

That also means you do not want to pony up a bundle of cash for a case with a 400w PSU, you're just not going to use it. I use these cases from ebay. 200w PSU, and $25 delivered to my door. Since the kill-a-watt says I'm only taking 110w, 200w is plenty enough.

If you're going the Linux route, take a look at how to use pxe to boot them over the network - just about all modern mobos will support a network boot.

you gotta be careful with matching a PSU's rated spec to measured power draw for two reasons:
first, cheap PSUs rarely measure up to their rated spec, and second, HDs have a power spike at startup (if running from flash or netbooting then not a problem!).

I'd also pay more for a more efficient PSU - I like the Seasonic S12 range but there's a fair bit of choice now. There's a new article here:
http://www.anandtech.com/casecoolingpsus/showdoc.aspx?i=3413&p=10

---