What to build is a tough question because who knows what protiens are going to be available or what will be done to the point system that many have become unhappy with.
Personally, I think we'll see more and more bigpacket WUs. The p148x series will see more and more complex simulations with bandwidth requirements exceeding that of current QMDs. This bodes well for the A64 architecture and Dothan architecture (provided the bandwidth requirement doesn't get too high. However I think the current bonus structure is in trouble. I can't see p13xx, and p114x (600 pointers) keeping their 100% bonus. QMDs will remain a small percentage of the available WUs. The computation is so complex that current 450 point simulations are run on the simplest of protiens and the water around them. I don't think much more complex simulations can be done on current PC hardware, limiting QMD cores use. There may however, always be some.
Now what the heck was the question?
Oh. folding farms.
You want it to last, use very little power and provide great ppd.
Silver's work with the Celeron D proved they can fold QMDs at a good clip, 128ppd/GHz, running on a 200 MHz FSB. A 2.26 would have to OC to 3.4 to get there and you'd make 435ppd/layer as long as the QMDs hold out. Getting the FSB up is crucial and production doesn't scale linearly. Celerons on less than an 700 MHz FSB produce around 95 ppd/GHz on QMDs. 3 layers = 1300ppd which should put you over 2000 ppd with the X2. Celerons Ds are not the way to go unless you plan to massively OC them.
I don't think Socket A is the way to go since it has relatively limited bandwidth. On p147x an XP-M (200MHZ FSB) produces 125 ppd/GHz, A64s with 1 MB caches produce over 200ppd/GHz. With the WUs headed toward greater complexity and thus, probably, greater bandwidth requirements I think A Socket A layer will get old too soon. But Almost every XP-M gets to 2.4 GHz which gets production on the average big WU to 266 ppd. It would take 5 layers to get to 2000 ppd with your X2. Run out of big WUs and it'll take a lot more. Production on regular Gromacs is around 55ppd/GHz.
With the Dothan and a steady diet of p1477 it would take just over 1 so we have to call it two. The Dothan is magic on p147x and p148x. You should shoot for maximum bandwidth in any Dothan approach you choose. The whole stink about QMDs started with a beta WU that used more than the bandwidth available on a X2 when folding two instances. Caveat here is if p147x and 148x run out, at 155ppd/GHz and 2.8 GHz production is 434 ppd per layer. If the project suddenly went to nothing but the Tinker core, you'd be screwed as the Dothans suck at Tinkers. QMDs will be coming to the Dothan but I don't expect them to be great QMD folders.
P4 6xx layers would be expensive, but overclocked, you'd only need two folding QMDs. The 6xx series actually makes more ppd on p147x and 148x than they do on QMDs. @ 3.6 GHz each layer would produce around 560ppd on QMDs and 600 on p147x.
If I were to build a rig today, or a farm, I'd gamble on the Dothan set up using a 740 or 750 on a 200MHz FSB (800 effective) using the ASUS adapter.
As I said earlier, a change to the points system or the available WUs could devalue your investment.
Personally, I think we'll see more and more bigpacket WUs. The p148x series will see more and more complex simulations with bandwidth requirements exceeding that of current QMDs. This bodes well for the A64 architecture and Dothan architecture (provided the bandwidth requirement doesn't get too high. However I think the current bonus structure is in trouble. I can't see p13xx, and p114x (600 pointers) keeping their 100% bonus. QMDs will remain a small percentage of the available WUs. The computation is so complex that current 450 point simulations are run on the simplest of protiens and the water around them. I don't think much more complex simulations can be done on current PC hardware, limiting QMD cores use. There may however, always be some.
Now what the heck was the question?
Oh. folding farms.
You want it to last, use very little power and provide great ppd.
Silver's work with the Celeron D proved they can fold QMDs at a good clip, 128ppd/GHz, running on a 200 MHz FSB. A 2.26 would have to OC to 3.4 to get there and you'd make 435ppd/layer as long as the QMDs hold out. Getting the FSB up is crucial and production doesn't scale linearly. Celerons on less than an 700 MHz FSB produce around 95 ppd/GHz on QMDs. 3 layers = 1300ppd which should put you over 2000 ppd with the X2. Celerons Ds are not the way to go unless you plan to massively OC them.
I don't think Socket A is the way to go since it has relatively limited bandwidth. On p147x an XP-M (200MHZ FSB) produces 125 ppd/GHz, A64s with 1 MB caches produce over 200ppd/GHz. With the WUs headed toward greater complexity and thus, probably, greater bandwidth requirements I think A Socket A layer will get old too soon. But Almost every XP-M gets to 2.4 GHz which gets production on the average big WU to 266 ppd. It would take 5 layers to get to 2000 ppd with your X2. Run out of big WUs and it'll take a lot more. Production on regular Gromacs is around 55ppd/GHz.
With the Dothan and a steady diet of p1477 it would take just over 1 so we have to call it two. The Dothan is magic on p147x and p148x. You should shoot for maximum bandwidth in any Dothan approach you choose. The whole stink about QMDs started with a beta WU that used more than the bandwidth available on a X2 when folding two instances. Caveat here is if p147x and 148x run out, at 155ppd/GHz and 2.8 GHz production is 434 ppd per layer. If the project suddenly went to nothing but the Tinker core, you'd be screwed as the Dothans suck at Tinkers. QMDs will be coming to the Dothan but I don't expect them to be great QMD folders.
P4 6xx layers would be expensive, but overclocked, you'd only need two folding QMDs. The 6xx series actually makes more ppd on p147x and 148x than they do on QMDs. @ 3.6 GHz each layer would produce around 560ppd on QMDs and 600 on p147x.
If I were to build a rig today, or a farm, I'd gamble on the Dothan set up using a 740 or 750 on a 200MHz FSB (800 effective) using the ASUS adapter.
As I said earlier, a change to the points system or the available WUs could devalue your investment.
Last edited: