So glad you have a dual-boot machine and could make the comparison across OS's. I don't have Windows on any machines (nor even any Windows media) so couldn't do that myself.

So a relevant question is, how close is the ratio of HT/no HT benchmarks to the ratio of calculations performed on real WUs with and without HT under Linux?

Since C4CW WUs are so consistent, we could use my timings on them to look at this. Let's convert to fraction of a workunit computed by a core in 1 hr.

                2 cores                    4 cores            ratio
float           583                            526           1.108
integer        3537                           2577           1.373
sum            4120                           3103           1.328

frac of        .08752                       .06845           1.279
WU per hr

So this limited data suggests that the sum of the integer and floating pt benchmarks is in the right ballpark for figuring claimed points, but that a weighted average of some kind probably would work better (but the weights probably would be science-specific).

By the way, I stumbled on this blog that actually discusses why HT might work well on an Atom!
http://www.agner.org/optimize/blog/read.php?i=6

Right -- this whole machine can do 1.5642 times as much work per hour with HT on as with HT off (which is why I keep HT on for running BOINC!). (The Atom probably benefits more from HT than a faster processor would.)

But what the benchmarks are trying to assess is calculation speed *per core*, not for the whole machine, because benchmarks are used in determining claimed points for each WU -- not for the composite of work done by the machine. So the 1.56 isn't the relevant ratio for comparing benchmarks. On this computer, *any particular WU* runs about 1.279 times faster with HT off than with it on, and presumably that's what the benchmarks should be trying to get at.

Until now we/me were under the strong impression that Linux-64 is superior in handling integer calculations and HCC running 2x faster still after the last application revision suggests so. I've got only non-HT quad comparisons on the BOINC 6.10.58 benchmarks and fpops don't even differ by 100, but iops on W7-64 is stable at 6750 and the Linux-64 is 12800, almost 2x higher.

I'll reserve revision of opinion until the techs/programmers come up with a new compile that matches the ~1:15 hours on Linux-64 for HCC against the W7-64 of 2:40. This science is probably integer exclusive.

--//--

edit: PS, on 64 bit, I think HAL said something just before it was happening ;P

Did some changes on my harpertown rigs so will update mine:

Dual quad Xeon E5640 Westmere @2.66, running 15 threads on Win7 64
Whetstone. 2694
Dhrystone 6710

Dual quad Xeon E5472 Harpertown @3.0, (800 FSB) running 8 threads on Ubuntu 64
Whetstone 2939
Dhrystone. 15772

Dual quad Xeon E5420 Harpertown @2.5, (677 FSB) running 8 threads on Ubuntu 64
Whetstone. 2448
Dhrystone. 13274

2010 Mac mini @2.66 running SL
Whetstone. 3119
Dhrystone. 6825

Hyperthreading really drops the whetstone if maxing out all threads.

OS: Mac OS X 10.5 (Leopard), 64 bit
CPU: Intel Core 2 T7200 @ 2.0 GHz (mobile CPU for Mac mini 2007 model)
Boinc version: 6.2.18 for x86_64-apple-darwin

Number of CPUs: 2
2149 floating point MIPS (Whetstone) per CPU
7821 integer MIPS (Dhrystone) per CPU

Same machine,
Boinc version: 6.10.58 for x86_64-apple-darwin

Number of CPUs: 2
2055 floating point MIPS (Whetstone) per CPU
4975 integer MIPS (Dhrystone) per CPU