Although most of Rosetta's processing power may come from idle CPU cycles, I'm interested in building a folding farm of dedicated machines.

To that end, I've been evaluating processors, motherboards, and graphic cards. If someone's done the work, please point me in their direction. Otherwise, I'll post my evals as time permits.

The basic ROI variables are hardware cost vs. performance, and power consumption costs over the useful lifespan of a box. In other words, what is the lowest overall average cost per work unit over a given period of time.

I've been looking at the monster rig the University of Antwerp built as a reference for a great GPU intensive system. While not entirely useful for this project yet, perhaps a GPU client will be developed. Ideally, I'd like to find a motherboard that can support quad PCI-e video cards, and dual processors for an 8x8 machine. While dual CPU boards are relatively common, the closest I've come to my ideal so far is a 3x pci-e, dual cpu system.

Any input is appreciated.

Do you have a link to info. about the machine in Antwerp?

Rosetta graphics are for your benefit, not the science, so unless you are planning to run projects that will use those graphics cards for useful work, it would be better to just plan on not displaying the screensaver or the graphic so far as your crunching efficiency, and power use are concerned.

Be sure you adequately plan for the heat produced. Almost every watt of power used ultimately ends up as heat in the room. So, it will be significant.

---

Rosetta Moderator: Mod.Sense

Although most of Rosetta's processing power may come from idle CPU cycles, I'm interested in building a folding farm of dedicated machines.

To that end, I've been evaluating processors, motherboards, and graphic cards. If someone's done the work, please point me in their direction. Otherwise, I'll post my evals as time permits.

The basic ROI variables are hardware cost vs. performance, and power consumption costs over the useful lifespan of a box. In other words, what is the lowest overall average cost per work unit over a given period of time.

I've been looking at the monster rig the University of Antwerp built as a reference for a great GPU intensive system. While not entirely useful for this project yet, perhaps a GPU client will be developed. Ideally, I'd like to find a motherboard that can support quad PCI-e video cards, and dual processors for an 8x8 machine. While dual CPU boards are relatively common, the closest I've come to my ideal so far is a 3x pci-e, dual cpu system.

Any input is appreciated.

I've seen messages saying the the 6.2.* version of BOINC supports using GPUs, but I've seen nothing indicating that any BOINC project makes use of this support yet.

The following project uses GPUs, but making it work on the same machine as any BOINC projects does not seem ready for beginners, and requires devoting a CPU core to communicate with the GPU so you might think of a quad-CPU-core machine if you try it anyway:

http://folding.stanford.edu/

Hyperthreaded CPUs often slow down BOINC projects, so you might want to avoid them or at least turn off the hyperthreading. Multiple CPU cores in the same CPU allow you to work on more than one workunit at the same time, but can run into problems with finding enough memory to support all of them, especially under 32-bit operating systems such as the common varieties of Windows. The 32-bit versions of Windows Vista don't allow you to use memory beyond a limit of about 3.5 GB. 64-bit operating systems, such as some varieties of Windows you can find with extra effort, allow the use of memory beyond the 4 GB limit for 32-bit OSs.

Although most of Rosetta's processing power may come from idle CPU cycles, I'm interested in building a folding farm of dedicated machines.

To that end, I've been evaluating processors, motherboards, and graphic cards. If someone's done the work, please point me in their direction. Otherwise, I'll post my evals as time permits.

The basic ROI variables are hardware cost vs. performance, and power consumption costs over the useful lifespan of a box. In other words, what is the lowest overall average cost per work unit over a given period of time.

I've been looking at the monster rig the University of Antwerp built as a reference for a great GPU intensive system. While not entirely useful for this project yet, perhaps a GPU client will be developed. Ideally, I'd like to find a motherboard that can support quad PCI-e video cards, and dual processors for an 8x8 machine. While dual CPU boards are relatively common, the closest I've come to my ideal so far is a 3x pci-e, dual cpu system.

Any input is appreciated.

I've seen messages saying the the 6.2.* version of BOINC supports using GPUs, but I've seen nothing indicating that any BOINC project makes use of this support yet.

The following project uses GPUs, but making it work on the same machine as any BOINC projects does not seem ready for beginners, and requires devoting a CPU core to communicate with the GPU so you might think of a quad-CPU-core machine if you try it anyway:

http://folding.stanford.edu/

Hyperthreaded CPUs often slow down BOINC projects, so you might want to avoid them or at least turn off the hyperthreading. Multiple CPU cores in the same CPU allow you to work on more than one workunit at the same time, but can run into problems with finding enough memory to support all of them, especially under 32-bit operating systems such as the common varieties of Windows. The 32-bit versions of Windows Vista don't allow you to use memory beyond a limit of about 3.5 GB. 64-bit operating systems, such as some varieties of Windows you can find with extra effort, allow the use of memory beyond the 4 GB limit for 32-bit OSs.

A quad-CPU running Rosetta won't use more than 1GB of RAM (having 4 WU running at the same time). So 3GB is more than enough I believe.

---

@ Mod.Sense

The Antwerp machine: http://www.dvhardware.net/article27538.html

Note that this machine would be terrible for Rosetta, great for folding@home.

I should elaborate that I'd like to build rigs that can provide good CPU as well as GPU performance. I will be folding for both projects simultaneously. In the future, I would like to put the GPUs into play for Rosetta, if and when a GPU client is written.

@ robertmiles

I played with the Boinc GPU support, and it would not work properly for me. I didn't pursue it, because the project that currently uses it wasn't as interesting to me as others. Hopefully Rosetta will bring the GPU to bear soon. As a comparison, folding@home has over 3 Petaflops of power, over half of which is derived from GPUs. Wow.

Good point about the memory limitations of various OS. I'm planning on using 64bit Linux wherever I can. If not that, then 64bit Vista or Server 2k3.

@ Chilean

Good to know.

I've been using 64 bit Linux Ubuntu on a Q6700 with 2 gig of Ram. I don't see any consistently higher crunching rate than when I run the same rig in Windows xp 32 bit. The Linux box shows higher memory bandwidth available, but I suspect it's not being used. Even while I am running other programs concurrently with R@H, memory use does not exceed 1 gig.

A quad-CPU running Rosetta won't use more than 1GB of RAM (having 4 WU running at the same time). So 3GB is more than enough I believe.

It is at the moment, but the memory requirements do change over time - there were some jobs a while back that were using a few hundred MB each. I think the memory requirements of minirosetta have been lower than for the previous versions but that might change.

---