What is your OS, Christos?

My Phenom II X6 1075T on Ubuntu 10.10 (64bits - Synaptic) got just:
2323 Whetstone - (I'm disappointed)
14101 Dhrystone

This machine:

Phenom II X6 1075T on Ubuntu 10.10 (64bits - Boinc from Synaptic):
2323 Whetstone
14101 Dhrystone

On W7, got a better whetstone:
2497
Dhrystone was much worse, but I dont remeber the exact number :(

So, a newbie's question:
What matters most to the crunching? Dhry, Whet or a mixture of both?

The benchmark and claim approach is 'bend' to say the least. No matter if Windows 32 or 64 bit or Linux 32 or 64 bit, the Whetstone test for FPOPS is pretty much the same... they are 32 bit, period. When it comes to Dhrystone / Integers, that's not the case at all... there is a very big difference between the bit sizes of the OS/client and the platforms. Simply though, Whet+Dhrystone is summed and divvied by about 480, which produces the credit claim per hour.

On the server side, it's all thrown in 1 pot... OPS. Then regardless of whether a science is heavy on Floaters or Integers, the credit claims are evaluated and a grant is computed from that.

--//--

Inclination to think this is an OS/CPU quirk, not BOINC and does explain some outlandish credit claims I've observed on various wingman, who of course are being offed by being declared outlier. Better not make it a habit of repeating this "undocumented trick'' :D

It it apparently OS-related.
Here is a comparison between Win 7 and Ubuntu 10.10 using the same hardware and the same BOINC version:

Intel i7 2600K @ 4.0GHz, HT on, turbo off

Windows 7 64 bit
BOINC 6.10.58 for windows_x86_64

# cores       8     6     4     1
float      3827  3804  3809   4105
integer   10063 10091  9947  10008

On Windows 7 the scores are quite stable, independent of the number of active cores.

Ubuntu 10.10 64 bit
Boinc 6.10.58 for x86_64-pc-linux-gnu (from Ubuntu Software Center)

# cores       8     6     4     1
float      3468  3617  4003  4010
integer   14596 18232 23007 25554

On Ubuntu the score increases as the number of active cores decreases, particularly for integer.

It seems to me that at least the Dhrystone score is calculated differently on the Linux client. I think it runs the benchmark on all cores available to Boinc, takes the total score and divides by the number of cores.
When the number of "Boinc cores" is no more than the number of physical cores, the score is stable. But when the "Boinc cores" also include virtual cores, the score per CPU drops since virtual cores are quite wimpy compared to real ones.
If we compare the total Linux score for 4-core integer (all real cores: 4 * 23007 = 92,028) and 8-core integer (all real + all virtual cores: 8 * 14596 = 116,768) the performance increase from using HT is ~25%.
Anandtech reports that HT on Nehalem gave a performance gain of 7% floating point, 14% integer and up to 30% "mixed" (presumably from reduced task switching overhead). Sandy Bridge HT may be slightly improved over Nehalem, so I can believe that that the 25% integer gain from HT is real, not just an artifact of the benchmark.
(The Atom HT is actually superior percentage wise, according to kateiacy's benchmarks )

I don't know what the Windows benchmark does, but it's apparently something different than the Linux version.

I suppose the offshoot is that the Linux and Windows benchmarks may both be true (for some value of "true"), but they are not directly comparable. At least not when hyperthreading enters the picture.

No worries I guess it just means that we will have to live with some spread between Linux and Windows point claims for a while yet

Edit: Rewrote benchmarks to more compact table format
Edit: Reverted table listing for Ubuntu 4-core benchmark to original - had numbers from different benchmark runs in the table and in the percent-gain-from-hyperthreading calculations. There are variations up to ~10% from run to run.