LHC@home

Well, just for the sake of proving the question of 64-bit processors being "faster" than their 32-bit counterparts ... I went and bought a 64-bit machine ...

My Original Dell Dual Xeon was 3.4 GHz, 2M cache, 1 G RAM, cheapest video card 80G 10,000 RPM disk drive, but was 32-bit Xeon chips.

Yesterday afternoon I got a Dell Dual Xeon 3.4 GHz, 2M cache, 1 G RAM, cheapest video card 40G 7,200 RPM disk drive, but is 64-bit Xeon chips.

After 16 hours I only have a little bit of work done and for comparison purposes the bulk of the work completed to this point has been LHC@Home which is useless for measuring as the work units can take any amount of time to process ...

But of the 9 LHC@Home work units, 1 failed download (for all of us) 6 have validated and have been granted credit and 2 are pending.

For the other projects I have 2 Trickles for CPDN (I attached there first and have 4 CPDN work units ... be interesting if it maintains 4 in the buffer or not - my older Xeon has two in flight right now, so I AM curious) and one Einstein@Home work unit which is pending at this moment.

I still have to attach it to Predictor@Home and SETI@Home ...

---

Paul,

Are you running XP64, *nix, XP32?

Even in a 64-bit computing environment there wouldn't be a significant speed benefit if the program is still 32-bit code. You'll definately notice a difference when the projects eventually migrate to 64-bit code in the next few years. Making the BOINC client 64-bit wouldn't make any impact on the project running. And even though the processor is 64 bit, assuming it's dual channel, will still only access 128 bits of data at a time though if it's DDR2 667 or faster, it will outperfrom (finally) DDR400.

Several benchmarks have been run - and over at hardocp I do believe they had an exchaustive review of XP64 running Doom3 amongst other games. They then compared their 32-bit program on XP64 to 32-bit on XP32 and although performance was slightly degraded on XP64 compared to XP32, the difference is nearly statistically insignificant. The changes won't be noticed until the programs run 64-bit natively. Same goes for BOINC, sixtrack, hadsm3, amongst others. Though I think you'll notice a difference for the better if it has DDR2 667 or better. :o)

---

Which BOINC Client are you using? Does Sixtrak Client run in 32 or 64 bit? Which OS?

Cheers,

Sysfried

---

64 bit or 32 bit makes very little difference for Sixtrack.
What those numbers refer to are the address bus and integer register size. Since Sixtrack should easily fit in 4 GB of ram the larger addressing range might even be a drawback. This is because an address (or pointer) will now take up twice the space in cache, which is a much more a bottle-neck for Sixtrack than RAM.

Almost all the numbers calculated on in Sixtrack are decimal numbers that are handled by the floating point register, it therefore makes no difference that the integer registers now can hold larger numbers.

Amd 64 bit chips should be able to run 32 bit code without any trouble so you can just run the 32-bit version, at least in our case.

---

Chrulle
Research Assistant & Ex-LHC@home developer
Niels Bohr Institute

> 64 bit or 32 bit makes very little difference for Sixtrack.

that's sad. I expected the cpu's to gain... due to double precition or other things.
>
> Amd 64 bit chips should be able to run 32 bit code without any trouble so you
> can just run the 32-bit version, at least in our case.
>

yeah, they run it fine... check my hosts... 2 amd64 cpu's... but I run 32 bit OS since running lhc 32 bit client would slow down performance (due to WoW wrapper)

still, I'd love to see a 64 bit compiled client!!!!

---

I came with Win XP-64 gotta keep the test even.

I know all about the 64 bit-"ness" and how it affects the probable real world. This is just for those that claim that a 64-bit processor, in and of itself, will be better than its 32-bit counterpart. I guess I was not clear enough that my expectation would be that the difference will be insignificant... or strangely enough that the 64 bit machine will be SLOWER (which is what I expect to be honest).

So, this is just another of Paul's expiriments ... I was allowed to buy one computer this year and this is it (the other Xeon was from my Mother's estate so it does not count, of course that is not what my spouse says)...

Then again, I spent the better part of the week end doing financial stuff (ugh!) and it turns out we are still nearly 4 years ahead of schedule for retirement and we are not even having to put in the counterpart funds I had projected. So, the finances are on auto-pilot and we are just riding on returns ... of course if we also add in my wifes bank account we are 5 years ahead of schedule ...

But I digress ...

As I said, it is like those other expiriments on the old site (which are still updated) showing average processing time are still showing interesting information... I am going to have to update the queries to give me a separate table for the second Xeon now ...

{edit}
I changed the page, but don't udate stats till Sunday.. and I have to relpad them all as my local copy and the site copy are out of synch some how ...

---

>
> As I said, it is like those other expiriments on the old site (which are still
> updated) showing average processing time are still showing interesting
> information... I am going to have to update the queries to give me a separate
> table for the second Xeon now ...
>

Go for AMD64 hardware!!!!!

---

I plan to buy another PC. So what am I suppose to do : buy a Pentium D or an AMD x2 ?

---

Depends on what you want. If you can afford it, an Intel dual with HT still outproduces the AMD chips... if you can't afford it ... then the AMD x2 is better for production, worse for heat and power consumption ...

---

What do you mean by afford? An Pentium D ( well in Europe) is far far far cheaper than an AMD X2. And the Pentium D has not HT technology. So, I think, there's two single bi-core. One is faster.... which one?

---

> I plan to buy another PC. So what am I suppose to do : buy a Pentium D or an
> AMD x2 ?
>
>
>
check my computers. I have 1 Dual cpu 1.8 GHz Opteron (Opteron 244)
Also, 1 DUAL CORE 2.2 GHz 512k L2 Cache Athlon X2 (4200+)
AND 1 Intel 3.43 GHz Machine (it's a P4 EE with 2 MB L3)

having HT on that Intel, my Dual Core 2.2 GHz outperforms it without a problem!
Pentium D have 2 full cores... forget everything you learned about HT on them.

AMD will introduce cheaper X2's this summer (August, IIRC)... so pricing shouldn't be a prob.

I don't have Intel P4 D power data handy, but I doubt that it beats the 110W TDP on the X2 (which also supports cool'n'quiet.. but you don't really use that if you have BOINC running)

so.. anyone interested in buying the Intel P4 I own? I'd be happy to sell it...
and when I sell it, I'm gonna get more AMD64 hardware!

good night,

sysfried

---

OK. I'm slowly convinced to buy an AMD. But the price here is around 1100 € for a x2 4800....
Thanks for your comments.

---

OK. I'm slowly convinced to buy an AMD. But the price here is around 1100 € for a x2 4800....
Thanks for your comments.

---

Thierry,

Tom's hardware has a wonderful in-depth article detailing the real-world differences between the PD and X2.

Summary:

X2 is up to 30% more energy efficient than PD
PD is 130w TDP, X2 is 110w TDP, however the X2's real-world temperatures suggest 110 watts is more than the heat energy it actually generates
In single applications the X2 is up to 30% faster than PD
In multiple applications, with HT enabled, the PD runs more threads faster than the X2
In multiple applications, with HT off, the X2 runs more threads faster than PD
The Socket 939 (AMD) platform is more stable and experiences fewer problems in regard to processor compatibility than Intel's 2nd generation Socket 775 (LGA)
(sidenote: if a user purchased an ASUS A8N-SLi series motherboard when it debuted over a year and a half ago the motherboard would be X2 compatible with the 1008 bios update, however if a user purchased a socket 775 intel-based motherboard around the same date, it would *not* be compatible and would require a new motherboard because of hardware differences between the "old" and "new" 775)
If you desire SLi capability, the nForce4 product for AMD is more mature and reliable than that for Intel

On a final and barely-related note, AMD is getting ready to obsolete socket 754 next year and make 939 its value line when their new socket comes out in 2006. It will support DDR2 memory and faster HT bus speeds. Since AMD has a tradition of using the same socket for a long time, expect the new one to offer as much upgrade potential as the Socket-A & 939 have.

---

For how much are you wanting to sell it? Also, Where are you located? Just want to make sure that shipping isn't going to cost more than the hardware. :)

Jim

> so.. anyone interested in buying the Intel P4 I own? I'd be happy to sell
> it...
> and when I sell it, I'm gonna get more AMD64 hardware!
>
> good night,
>
> sysfried
>

---

> > 64 bit or 32 bit makes very little difference for Sixtrack.
>
> that's sad. I expected the cpu's to gain... due to double precition or other
> things.

I think you are adding a 32bit CPU to a 32bit CPU, thinking it'll do double precision. It doesn't. 2 times 32bit isn't equal to 64bit.
2 to the power of 32 = 4294967296 .. 2 times that is 8589934592
2 to the power of 64 = 18446744073709551616 ..

Where does 2 come from? It's your bit.

What does a 64bit CPU do extra that a 32bit CPU doesn't? Address more memory.
A 32bit CPU can address at most 4GB of memory, whereas a 64bit CPU can address up to 16 exabytes. If you haven't got a clue how much that is, look at the number above again.

Now, since most PCs, bar servers, out there don't have motherboards yet that can pack more than 4GB of memory, or their Windows don't allow them to have a pagefile larger than 4GB, it's useless to call your 64bit CPU truelly a 64bit CPU, since it's hampered by the OS, the motherboard and the amount of memory.

Nice try. :)

---

Jord

BOINC FAQ Service

*g* Double Precition in terms of calculations.... if you do math and your result has twice as much digits than if it is done on 32 bit cpu's....

this link is hardcore for developers with math background.

http://www.amd.com/us-en/assets/content_type/DownloadableAssets/dwamd_CW2004.pdf

Taking the ACML AMD provides I would expect to see quite some improvements on any application that has lots of calculations done. On the other hand, Chrulle pointed that Sixtrack does mostly integer stuff on the FPU, so that shows me that he knows quite a bit about 32 vs 64. Yet, still I'd like to see a 64 bit build. Cause the speed decrease on LHC is the only thing that keeps me from switching to 64 bit OS.

Cheers.
>
> I think you are adding a 32bit CPU to a 32bit CPU, thinking it'll do double
> precision. It doesn't. 2 times 32bit isn't equal to 64bit.
> 2 to the power of 32 = 4294967296 .. 2 times that is 8589934592
> 2 to the power of 64 = 18446744073709551616 ..
>
> Where does 2 come from? It's your bit.
>
> What does a 64bit CPU do extra that a 32bit CPU doesn't? Address more memory.
>
> A 32bit CPU can address at most 4GB of memory, whereas a 64bit CPU can address
> up to 16 exabytes. If you haven't got a clue how much that is, look at the
> number above again.
>
> Now, since most PCs, bar servers, out there don't have motherboards yet that
> can pack more than 4GB of memory, or their Windows don't allow them to have a
> pagefile larger than 4GB, it's useless to call your 64bit CPU truelly a 64bit
> CPU, since it's hampered by the OS, the motherboard and the amount of memory.
>
>
> Nice try. :)
>

---

Located in Germany.

Somewhat 350 € for the cpu with cooler (boxed cooler)

If you want, we can talk about DDR2 RAM and a Mobo as well.

To contact me, check the website of my team. Look for the contacts and then write me a note.

Cheers,

Thorsten

aka. Sysfried.


> For how much are you wanting to sell it? Also, Where are you located? Just
> want to make sure that shipping isn't going to cost more than the hardware.
> :)
>
> Jim
>
> > so.. anyone interested in buying the Intel P4 I own? I'd be happy to
> sell
> > it...
> > and when I sell it, I'm gonna get more AMD64 hardware!
> >
> > good night,
> >
> > sysfried
> >
>

---

After double checking (and fyi, I'm not a programmer but I believe I understand this enough to talk about it)... what Ageless is saying, is even on a 32-bit cpu, all floating point data is stored in one of eight 80-bit registers. So double precision numbers do fit in the registers properly on a 32-bit cpu.

Furthermore, the AMD64 has 16 instead of 8 general purpose registers (read: integer data) when it's in 64-bit mode, and they're 64-bits long instead of 32. However the x87 (floating point) registers are still 80-bits long. According to Chrulle, sixtrack uses fpu (x87) instructions for the most part when it's computing. He's right stating there would be no performance gain on 64-bit architecture.

From what I can tell, sixtrack only runs on x86 architecture between *nix and Win32 operating systems. However I don't know if sixtrack currently uses SSE or if it could even benefit from using SSE instructions in the first place. If SSE is presently used or can be used then there most likely would be a compelling reason to release a 64-bit client. In 64-bit mode, 16 128-bit SSE registers are available, which is double the number of registers than 32-bit mode offers and is 48 bits longer than x87 registers. However if sixtrack doesn't use SSE there is no benefit for x87 or MMX instructions on a 64-bit chip compared to a 32-bit CPU. The same quantity of registers which hold the same quantity of data are present on both architectures.

If you'd like a good example of FPU math and a performance comparison between MMX and SSE, try running LAME (MP3 encoder) 3.96.1 in "-q0" mode and then run LAME 3.97 alpha 10 in "-q0" mode. The preset on the former only uses MMX instructions to encode and the latter has a new SSE optimized encoding method which is 3.5x faster than MMX. It gets the same job done with identical quality, just makes better effective use of the 128-bit registers and instructions available to SSE (MMX only has 64-bit registers). :o)

Sysfriend - I believed the FPU does floating point calculations, not integer calculations, doesn't the ALU does integer functions? Chrulle said the FPU handles decimal numbers which are not to be confused with integers. His statement says floating point data (which is 80 bits long at most) will not be helped by increasing the amount of integer data can which fit in a general purpose register (the 64-bit part).

---

I'm afraid that the shipping charges would be too much to ship to the states. :( Oh well.... I"ll just keep my eyes open for something closer. :)

Good luck selling your equipement.

Jim

> Located in Germany.
>
> Somewhat 350 € for the cpu with cooler (boxed cooler)
>
> If you want, we can talk about DDR2 RAM and a Mobo as well.
>
> To contact me, check the website of my team. Look for the contacts and then
> write me a note.
>
> Cheers,
>
> Thorsten
>
> aka. Sysfried.
>
>

---