That´s why i use: IIRC ... i only try to point him where to find the builds, there he could find the one he needs.

Anyway is interesting to ask why exist a 64 bit MB version for Linux and the x41zc are not avaiable in 64 bits for Windows.

I think on Linux the majority of hosts are x86_64, with i686 in minority, So far Urs has supplied x86_64 apps, be it CPU or ATI,
For the Linux x41g port, the 32bit version didn't produce good results, so a 64bit x41g app is the only possibility for Linux Cuda crunching,
For windows, 64bit addressing slows the x41 Cuda app down, since we don't need 64bit addressing a 32bit Cuda app is what we have.

Edit: Jason beat me to it.

Claggy

[edit:] Yeah a working Linux 32 bit x41zc build would be faster too...

LoL, perhaps I can simplify further though.

For GPUs, 64 bit does not mean faster or better in any way, only can access more memory.... which we don't need (yet). IF we do need it we get less registers, GPU code goes slower.

For CPU Code it's different, because there is a translation layer is OS that must translate 32 bit to 64 bit. Also on CPU 64 bit lets you use more registers, instead of chewing them up.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

That was fast, thanks for the info.

---

on a topic: at current moment we have numbers ( get that on motherboards with i965 chipset, cpu core 2 duo 1.8 - 2.2 gHz range, OS - linux slackware64 14.0 - 14.1, x64).
seti@home, lunatic mb and Ap apps for nvidia / ati.
last nvidia drivers ( 331.49 or so on), and last ati catalyst 13.1 legacy.
only one ati card - msi hd4350, passive cooling, 512 mb VRAM.
because of driver gluck, it not crunch MB tasks, but AP tasks crunching with ~24.5 GFLOPS rate ( based on 14 tasks result).

Nvidia Geforce 9400 GT with 512 mb VRAM from ASUS crunch about the same rate - AP with about 24 GFLOPS, MB rate is unknown. On i3 3220 system i got one AP task with standart seti (not lunatics) - (opencl_nvidia_100) at 33.95 GFLOPS rate.
then i have Geforce 210 with 512 Mb VRAM on i3 3220 system - it shows about 9.66 GFLOPS on seti MB tasks (based on 26 tasks results).

then i have core2duo 2.2 GHz with ASUS GT630 with 1 Gb VRAM, bit underclocked from factory memory speed ( 1600, instead standart for that 1800 MHz), rev 2.0 card (gk208, 384 CUDA cores) - it looks AP speed about 77.25 GFLOPs (based on 4 AP tasks), and MB (based on 84 tasks) about 17.54 GFLOPS

ASUS GTX650 with 1 Gb VRAM on core 2 duo 1.8 GHz show about 63.68 GFLOPS on MB tasks(based on 94 tasks result) . because of shortage of AP tasks now, i do not AP data for GTX650.

i think, i soon get scores for ASUS GT630 rev 2.0 with 2 Gb VRAM and standart memory clock ( 1800 MHz), and maybe also GT640 numbers.

on glxgears my GT630 show very impressive 14 000 FPS. 9400GT shows about 4018 FPS, and HD4350 - about 2624 FPS. GTX650 numbers is be about 9000 - 10 000. strange, but GT630 numbers is about a half higher, than really higher performance GTX650, in "glxgears".

---

That´s why i use: IIRC ... i only try to point him where to find the builds, there he could find the one he needs.

Anyway is interesting to ask why exist a 64 bit MB version for Linux and the x41zc are not avaiable in 64 bits for Windows.

I think on Linux the majority of hosts are x86_64, with i686 in minority, So far Urs has supplied x86_64 apps, be it CPU or ATI,
For the Linux x41g port, the 32bit version didn't produce good results, so a 64bit x41g app is the only possibility for Linux Cuda crunching,
For windows, 64bit addressing slows the x41 Cuda app down, since we don't need 64bit addressing a 32bit Cuda app is what we have.

Edit: Jason beat me to it.

Claggy

[edit:] Yeah a working Linux 32 bit x41zc build would be faster too...

LoL, perhaps I can simplify further though.

For GPUs, 64 bit does not mean faster or better in any way, only can access more memory.... which we don't need (yet). IF we do need it we get less registers, GPU code goes slower.

For CPU Code it's different, because there is a translation layer is OS that must translate 32 bit to 64 bit. Also on CPU 64 bit lets you use more registers, instead of chewing them up.

So my question would be is there a 64-bit AVX CPU application for MB and AP6 on windows and would that result in the best optimization for a Windows box (Windows 7)?

---

So my question would be is there a 64-bit AVX CPU application for MB and AP6 on windows and would that result in the best optimization for a Windows box (Windows 7)?

For the short to medium term, as a matter of directed effort, I'm entirely focussed on Cuda multibeam and Boinc(api) issues. So installer dudes (official Lunatics title :P) will have to chip in there, or devs with alpha/beta work in the area.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

do some additional tests - 9400gt and gtx650

looks like, most of MB tasks ( as again - use lunatics nvidia apps, sadly, but ap tasks not available due to server kaput) on 9400 gt took about 5700 seconds per task, and some "long tasks" take about 11 000 seconds ( 95 minutes, and 183 minutes respectively).
about 12 GFLOPs on MB.

on a gtx650 there is a hug difference - most ap tasks be done only in about 16 minutes! ( ~950 seconds), one long MB task takes a 2765 seconds ( 46 minutes).
about 73 GFLOPS on MB



and GT630 with underclocked memory (1600 mhz) shows 4600 - 15 000 seconds ( 76 - 250 minutes) per task - often the smaller number.

sadly, i do not have a gt640, but based on that data, i think, there no any point to look "GFLOPS" and "GFLOPS per watt" in various cards specs, because that data looks like no straight chain with real card performance, even in crunching. gt630 has about 692 GFLOPS processing power, GTX650 has about 802 GFLOPS - minimal gap. price is about 48 eur vs 90 eur. performance on GTX650 looks like about 3 - 4x better. very significant.
try to small test on games too. x2: the threat benchmark on 1024x768x32 on all on ( shadows too) on GT630 (gk208) shows about 26 FPS, on GTX650 - 52 fps.
Nexuiz on GT630 shows about 47 FPS average, on GTX650 is about 115 average.
looks like from all aspects, GTX650 is more preferable, if you can afford active cooling for videocard.
be good to see a GTX750 tests too...


Oh yes - small "FYI" for folks, who think, he can search over SETI users, find machine with interesting GPU and compare based on that - if you look at "GFLOPS" performance for seti tasks here, you must be sure, on that host is used only that videocard. because, if you swap videocard, and all mother stays the same, that fact is no declared anywhere. for example, i put in my computer GTX780 card, do a 100wu, then swap that GTX780 with GT210.
now you see my "computer details" in SETI site, and see my computer have GT210 card, with very fantastic results...it has been polluted with previously GTX780 results, but that fact is hidden.
be a better, if changed vidocard automatically make a new entry in "my computers" view, but....

---

yes, also found out "glxgears" stranginess.
putting gtx650 in my home computer ) slackware64 14.1, core2duo 2.2 ghz, i965 chipset. i get bit better numbers in glxgears - about 15 000 fps, bit better than on gt630 on the same machine, and significantly better, than tyhe same gtx650 on work computer ( core2duo 1.8 ghz, slackware64 14.1, i945 based chipset.). GFLOPS on MB seti tasks on work computer too is a significantly lower - about 63 gflops versus 73 on home computer.
but tendency, in any way, is easy to see anyway.

---

There's some talk on the Internet saying glxgears isn't good as a benchmarking tool.