What hardware do you recommend?

I am not sure how Hyperthreading works with Einstein, but in genereal Boinc really seems to profit from it, so i would go with an I7 that has it, as it will look to boinc like an 8 core Processor. If you are like me you would also probably roam some overclocking forums and see what cpus are good at the moment for that and buy a inexpensive one and clock it to the speed of around the top model or higher and invest part of the saved money to better cooling. But i am not that up to date with the I7 and overclocking, i only know that those CPUs already do a lot of that on their own.

For GPU i have to agree. I run an 8800GT in one of my PCs and it is only around half as fast as most GTX 580s that got to validate my units. That might change with a better Cuda app, but for the moment it might be best cost effetive wise to go with a powersaving GTX 2something or maybe even 9800 card or better 2 or 3 of them. If Einstein would support AMD/ATI cards i would suggest a 5850 wich has a lot of computing power and uses compared to Nvidia a lot less power. Over at Milkyway@Home those cards run circles around everything Nvidia has to offer.

http://einsteinathome.org/host/3810653/tasks&offset=0&show_names=0&state=3 That is my 8800GT doing its work, its interesting to look at the Work Unit Details, compare the runtimes and see what the other Computer has as hardware (but you won't know if and how it is overclocked). My PC is a Pentium Dual E2180 clocked from 2GHZ to 3GHZ and the 8800GT running at stock speeds.

Best you get is about +30% improvement on throughput running HT to give two virtual cores on one physical core vs running the physical core without HT.

It's still one physical core regardless of what the HT 'fools' the OS with. The "+30%" throughput depends highly on how sympathetic your application is for running two threads to take advantage of the HT. Boinc is good for that in that you can run two tasks simultaneously. Whether they take twice as long to execute or "30%" less than twice as long depends on the task.

Best is to get something out of the latest generation so that you can take advantage of the latest architecture and programming features. Hence, go bottom end GT-400 if you want low cost or power saving (passive cooling), or whatever. You at least still get the advantages of the Fermi architecture support for CUDA and OpenCL and double precision arithmetic.

Also important is to ensure you have enough video RAM on your video card. 512 MBytes is an absolute minimum. Best is to go for 1GByte so that you can run both tasks and have enough RAM to display your desktop!

Happy fast crunchin',
Martin

I recently ran some tests on an E5620 (Westmere). The report message gives hyperthreading throughput improvement ratios for my system in three different configurations, varying in number of memory channels provisioned

1-channel 4.4% (severely memory starved, no one would run this way)
2-channel 33% (I overclocked CPU and not memory, so most folks would be a bit more generously supplied)
3-channel 38% (as socket 1156 lacks three channels, this configuration, though lightly clocked on RAM, was probably a bit more richly supplied than most consumer systems of the Nehalem generation parts)

As the question posed was specific to Einstein, so is this answer, but in the broader topic the range of improvements seen have varied quite widely with:

1. specific MPU architecture (HT is not always the same at all)
2. application (some aps like HT a lot, and a very few hate it)
3. balance of RAM vs. CPU speed in specific system
4. cache sizes and speeds

and doubtless other stuff.