> Single AMD Opteron 144
> MSI Master-F dual Opteron MB
> 1GB ECC PC2700
> 320GB RAID5 3Ware SATA RAID
> Win 2K Pro running file shares
>
> 3181 Integer MIPS
> 1713 Double MIPS
>
> Try beating that with a $250 chip from Intel.
On E@H, it can be done, no sweat. I have a single 3.0 GHz P4 478 Prescott running at about 3.4 GHz. It cost $189.00 in the retail box. It routinely turns in results in 41,200 sec. Sounds slow? But it is hyperthreading & does two units from one CPU in that same 41,200 sec. So on a per-unit basis, it does 41200/2=20600 sec per unit. For most tasks, hyperthreading is not as valuable, but for E@H it exactly doubles the throughput. This computer is called "yeongw2k" if you want to check. Running with old, slow PC2100 memory too.
ADDMP (I like your Opteron, though. Like to have one myself.)
> > Single AMD Opteron 144
> > MSI Master-F dual Opteron MB
> > 1GB ECC PC2700
> > 320GB RAID5 3Ware SATA RAID
> > Win 2K Pro running file shares
> >
> > 3181 Integer MIPS
> > 1713 Double MIPS
> >
> > Try beating that with a $250 chip from Intel.
>
> On E@H, it can be done, no sweat. I have a single 3.0 GHz P4 478 Prescott
> running at about 3.4 GHz. It cost $189.00 in the retail box. It routinely
> turns in results in 41,200 sec. Sounds slow? But it is hyperthreading &
> does two units from one CPU in that same 41,200 sec. So on a per-unit basis,
> it does 41200/2=20600 sec per unit. For most tasks, hyperthreading is not as
> valuable, but for E@H it exactly doubles the throughput. This computer is
> called "yeongw2k" if you want to check. Running with old, slow PC2100 memory
I should have said that I had looked up how PickCoder's Opteron 144 was doing on E@H & it was running 24800 sec/unit compared to 20600 sec/unit for the 3 GHz P4 with hyperthreading.
And I should not have said that hyperthreading doubled the throughput. I have never run it with hyperthreading turned off & expect it would look quite bad.
> And I should not have said that hyperthreading doubled the throughput. I have
> never run it with hyperthreading turned off & expect it would look quite
> bad.
It does not look that bad. HT does not double the performance. It does give a boost to throughput between 10% to as high as 70% more work done per unit time.
We did fairly extensive tests during the late stages of the Beta test program with HT on and off and it does surprisingly well at improving the total production. And I say surprising in that our usage of the processors is very atypical ... we heavily use the math portions and do not have a very "standard" mix of instruction streams.
I mean, my G5 does more work because of the two CPUs than does any of my HT systems, but it is a cheap way to "see" two CPUs.
> > Single AMD Opteron 144
> > MSI Master-F dual Opteron MB
> > 1GB ECC PC2700
> > 320GB RAID5 3Ware SATA RAID
> > Win 2K Pro running file shares
> >
> > 3181 Integer MIPS
> > 1713 Double MIPS
> >
> > Try beating that with a $250 chip from Intel.
>
> On E@H, it can be done, no sweat. I have a single 3.0 GHz P4 478 Prescott
> running at about 3.4 GHz. It cost $189.00 in the retail box. It routinely
> turns in results in 41,200 sec. Sounds slow? But it is hyperthreading &
> does two units from one CPU in that same 41,200 sec. So on a per-unit basis,
> it does 41200/2=20600 sec per unit. For most tasks, hyperthreading is not as
> valuable, but for E@H it exactly doubles the throughput. This computer is
> called "yeongw2k" if you want to check. Running with old, slow PC2100 memory
> too.
>
> ADDMP (I like your Opteron, though. Like to have one myself.)
>
I just loaded a 3.0Ghz Prescott here and notice that too. You're not really comparing apples to apples, though. That's 2 WU running at a time. I've yet to find a way to run more than one WU, for a project, on a single CPU. When I run E@H, it slightly raises the temp on the chip. I was surprised after a few days of continuous crunching. Task Mgr shows 99% usage, but I know it's not even touching the available CPU bandwidth. The machine doesn't even blink with E@H running. I can run defrag(NTFS) and any number of other applications and they barely reflect, not that I can tell, that E@H is using "99%" of the CPU. I'm eager to find out if multiple WU on a 64-bit capable CPU would reduce the WU time or not. Anyone got any ideas?
> I've yet to
> find a way to run more than one WU, for a project, on a single CPU.>
> Glen
I was interested in your reply, but I do not quite get the sentence above.
I started up BOINC-E@H on the hyperthreading P4 & it immediately began running two WUs of E@H simultaneously. It is still running two WUs simultaneously.
That is what it does on a real-dual-CPU machine too.
Are you saying that is different from your experience?
Average time about 29 500 seconds.
1536 Whetstone
2633 Dhrystone
PC "custom-built" by a small firm.
AMD Athlon XP 2000+ (1.7 GHz)
512 MB RAM (2x256)
I don't know if it matters, but the graphics card is GeForce FX 5500 (128/128). No problems with the screen-saver or the "show graphics" option.
No errors up to now. (I'm running only Einstein and SETI.)
Have fun!
Vladimir
Oh, BTW - BOINC 4.25 splashed over 4.19 (I forgot to uninstall it).
The H/T fakes the O/S into thinking there's two cores and two sets of registers. The CPU can only execute one instruction at a time, so it has to switch between the instruction streams, much like thread execution under an O/S. Seeing the CPU throughput double on a P4 H/T makes me wonder how much I could squeeze out of a single Opteron. I would like to be able to run multiple computation threads on a single CPU, to see how much bandwidth the Opteron really is capable of. Would 2 BOINC threads kill the Opteron's performance or would it run both threads with the same WU time(or less) as a single thread? Unfortunately, I've yet to find a way to make BOINC run multiple WU threads when a single CPU is found. The CPU temp goes up a small amount, but the heat sink isn't even warm to the touch so I know it's not running @ max operating temp. I think the Opteron is being underused when running BOINC. Taskmgr shows 98-99% usage, but I seriously doubt the CPU is really loaded @ 99%. I've considered loading something like VMWare on it and running BOINC in several VME's. I'm not willing to buy it, though, just to test multiple threads. I'll check to see if there's a temp demo that'll lemme try it.
Glen
> I was interested in your reply, but I do not quite get the sentence above.
>
> I started up BOINC-E@H on the hyperthreading P4 & it immediately began
> running two WUs of E@H simultaneously. It is still running two WUs
> simultaneously.
>
> That is what it does on a real-dual-CPU machine too.
>
> Are you saying that is different from your experience?
>
> ADDMP
>
> ... I think the Opteron is being underused when running BOINC. Taskmgr shows
> 98-99% usage, but I seriously doubt the CPU is really loaded @ 99%. I've
> considered loading something like VMWare on it and running BOINC in several
> VME's. I'm not willing to buy it, though, just to test multiple threads. I'll
> check to see if there's a temp demo that'll lemme try it.
Why do you think that 99% CPU usage is not high enough?
If you judge it by the fact that if you start anything else up and it does not seem "slow" that there is remaining CPU cycles; well, it is not a true understanding.
BOINC and most other DC programs run at a low priority so they will "naturally" fade into the background and your computer will not seem to be burdened with any work at all. If you do manage to force it to run more than one thread, all you will accomplish is to slow down the processing ...
> The H/T fakes the O/S into thinking there's two cores and two sets of
> registers. The CPU can only execute one instruction at a time, so it has to
> switch between the instruction streams,
No, I don't think that's quite right. As I understand it, the P4 has parallel computation capabilities for SOME PARTS of the calculation. It can actually do more than one calculation at the exact same time. Hyperthreading was devised by Intel do make more efficient use of these duplicated sections of the CPU.
That's entirely different from time-slicing, which has been around much longer, & only produces the illusion of two or more actions happening at the same time, by switching rapidly from one to the other. Time-slicing does not increase through-put. It somewhat decreases through-put because there is some overhead in repeatedly stopping one calculation & starting up the other.
The MMX & 3DNOW & SSE & SSE2 & SSE3 sections also allow more than one calculation in the same clock cycle, but the E@H programers apparently haven't gotten around to putting them to work. But in Folding@Home, they greatly increase the throughput.
I don't know why AMD hasn't developed hyperthreading yet. Maybe someone else knows. Maybe Intel holds some patents.
But both Intel & AMD are said to be working on single chips that will contain two complete CPUs for late this year or early next year. That goes further than the P4 which only has duplication for some parts. Maybe AMD believes it has more to gain by concentrating its design force on this further development. I have seen reports that AMD might beat Intel into production.
> Single AMD Opteron 144 >
)
> Single AMD Opteron 144
> MSI Master-F dual Opteron MB
> 1GB ECC PC2700
> 320GB RAID5 3Ware SATA RAID
> Win 2K Pro running file shares
>
> 3181 Integer MIPS
> 1713 Double MIPS
>
> Try beating that with a $250 chip from Intel.
On E@H, it can be done, no sweat. I have a single 3.0 GHz P4 478 Prescott running at about 3.4 GHz. It cost $189.00 in the retail box. It routinely turns in results in 41,200 sec. Sounds slow? But it is hyperthreading & does two units from one CPU in that same 41,200 sec. So on a per-unit basis, it does 41200/2=20600 sec per unit. For most tasks, hyperthreading is not as valuable, but for E@H it exactly doubles the throughput. This computer is called "yeongw2k" if you want to check. Running with old, slow PC2100 memory too.
ADDMP (I like your Opteron, though. Like to have one myself.)
> > Single AMD Opteron 144 >
)
> > Single AMD Opteron 144
> > MSI Master-F dual Opteron MB
> > 1GB ECC PC2700
> > 320GB RAID5 3Ware SATA RAID
> > Win 2K Pro running file shares
> >
> > 3181 Integer MIPS
> > 1713 Double MIPS
> >
> > Try beating that with a $250 chip from Intel.
>
> On E@H, it can be done, no sweat. I have a single 3.0 GHz P4 478 Prescott
> running at about 3.4 GHz. It cost $189.00 in the retail box. It routinely
> turns in results in 41,200 sec. Sounds slow? But it is hyperthreading &
> does two units from one CPU in that same 41,200 sec. So on a per-unit basis,
> it does 41200/2=20600 sec per unit. For most tasks, hyperthreading is not as
> valuable, but for E@H it exactly doubles the throughput. This computer is
> called "yeongw2k" if you want to check. Running with old, slow PC2100 memory
I should have said that I had looked up how PickCoder's Opteron 144 was doing on E@H & it was running 24800 sec/unit compared to 20600 sec/unit for the 3 GHz P4 with hyperthreading.
And I should not have said that hyperthreading doubled the throughput. I have never run it with hyperthreading turned off & expect it would look quite bad.
[This is correct-the-blunders night.]
ADDMP
> And I should not have said
)
> And I should not have said that hyperthreading doubled the throughput. I have
> never run it with hyperthreading turned off & expect it would look quite
> bad.
It does not look that bad. HT does not double the performance. It does give a boost to throughput between 10% to as high as 70% more work done per unit time.
We did fairly extensive tests during the late stages of the Beta test program with HT on and off and it does surprisingly well at improving the total production. And I say surprising in that our usage of the processors is very atypical ... we heavily use the math portions and do not have a very "standard" mix of instruction streams.
I mean, my G5 does more work because of the two CPUs than does any of my HT systems, but it is a cheap way to "see" two CPUs.
> > Single AMD Opteron 144 >
)
> > Single AMD Opteron 144
> > MSI Master-F dual Opteron MB
> > 1GB ECC PC2700
> > 320GB RAID5 3Ware SATA RAID
> > Win 2K Pro running file shares
> >
> > 3181 Integer MIPS
> > 1713 Double MIPS
> >
> > Try beating that with a $250 chip from Intel.
>
> On E@H, it can be done, no sweat. I have a single 3.0 GHz P4 478 Prescott
> running at about 3.4 GHz. It cost $189.00 in the retail box. It routinely
> turns in results in 41,200 sec. Sounds slow? But it is hyperthreading &
> does two units from one CPU in that same 41,200 sec. So on a per-unit basis,
> it does 41200/2=20600 sec per unit. For most tasks, hyperthreading is not as
> valuable, but for E@H it exactly doubles the throughput. This computer is
> called "yeongw2k" if you want to check. Running with old, slow PC2100 memory
> too.
>
> ADDMP (I like your Opteron, though. Like to have one myself.)
>
I just loaded a 3.0Ghz Prescott here and notice that too. You're not really comparing apples to apples, though. That's 2 WU running at a time. I've yet to find a way to run more than one WU, for a project, on a single CPU. When I run E@H, it slightly raises the temp on the chip. I was surprised after a few days of continuous crunching. Task Mgr shows 99% usage, but I know it's not even touching the available CPU bandwidth. The machine doesn't even blink with E@H running. I can run defrag(NTFS) and any number of other applications and they barely reflect, not that I can tell, that E@H is using "99%" of the CPU. I'm eager to find out if multiple WU on a 64-bit capable CPU would reduce the WU time or not. Anyone got any ideas?
Glen
22,000 sec. +/- 500
)
22,000 sec. +/- 500 sec.
AMD 2400+ Mobile at 2.08 ghz on a Biostar M7NCG 400 MB with 512m RAM
$77 for the CPU - $63 for the MB - $76 RAM
*I still know CRAP when I see it.
PickCoder wrote: > I've
)
PickCoder wrote:
> I've yet to
> find a way to run more than one WU, for a project, on a single CPU.>
> Glen
I was interested in your reply, but I do not quite get the sentence above.
I started up BOINC-E@H on the hyperthreading P4 & it immediately began running two WUs of E@H simultaneously. It is still running two WUs simultaneously.
That is what it does on a real-dual-CPU machine too.
Are you saying that is different from your experience?
ADDMP
Average time about 29 500
)
Average time about 29 500 seconds.
1536 Whetstone
2633 Dhrystone
PC "custom-built" by a small firm.
AMD Athlon XP 2000+ (1.7 GHz)
512 MB RAM (2x256)
I don't know if it matters, but the graphics card is GeForce FX 5500 (128/128). No problems with the screen-saver or the "show graphics" option.
No errors up to now. (I'm running only Einstein and SETI.)
Have fun!
Vladimir
Oh, BTW - BOINC 4.25 splashed over 4.19 (I forgot to uninstall it).
The H/T fakes the O/S into
)
The H/T fakes the O/S into thinking there's two cores and two sets of registers. The CPU can only execute one instruction at a time, so it has to switch between the instruction streams, much like thread execution under an O/S. Seeing the CPU throughput double on a P4 H/T makes me wonder how much I could squeeze out of a single Opteron. I would like to be able to run multiple computation threads on a single CPU, to see how much bandwidth the Opteron really is capable of. Would 2 BOINC threads kill the Opteron's performance or would it run both threads with the same WU time(or less) as a single thread? Unfortunately, I've yet to find a way to make BOINC run multiple WU threads when a single CPU is found. The CPU temp goes up a small amount, but the heat sink isn't even warm to the touch so I know it's not running @ max operating temp. I think the Opteron is being underused when running BOINC. Taskmgr shows 98-99% usage, but I seriously doubt the CPU is really loaded @ 99%. I've considered loading something like VMWare on it and running BOINC in several VME's. I'm not willing to buy it, though, just to test multiple threads. I'll check to see if there's a temp demo that'll lemme try it.
Glen
> I was interested in your reply, but I do not quite get the sentence above.
>
> I started up BOINC-E@H on the hyperthreading P4 & it immediately began
> running two WUs of E@H simultaneously. It is still running two WUs
> simultaneously.
>
> That is what it does on a real-dual-CPU machine too.
>
> Are you saying that is different from your experience?
>
> ADDMP
>
> ... I think the Opteron is
)
> ... I think the Opteron is being underused when running BOINC. Taskmgr shows
> 98-99% usage, but I seriously doubt the CPU is really loaded @ 99%. I've
> considered loading something like VMWare on it and running BOINC in several
> VME's. I'm not willing to buy it, though, just to test multiple threads. I'll
> check to see if there's a temp demo that'll lemme try it.
Why do you think that 99% CPU usage is not high enough?
If you judge it by the fact that if you start anything else up and it does not seem "slow" that there is remaining CPU cycles; well, it is not a true understanding.
BOINC and most other DC programs run at a low priority so they will "naturally" fade into the background and your computer will not seem to be burdened with any work at all. If you do manage to force it to run more than one thread, all you will accomplish is to slow down the processing ...
PickCoder wrote: > The
)
PickCoder wrote:
> The H/T fakes the O/S into thinking there's two cores and two sets of
> registers. The CPU can only execute one instruction at a time, so it has to
> switch between the instruction streams,
No, I don't think that's quite right. As I understand it, the P4 has parallel computation capabilities for SOME PARTS of the calculation. It can actually do more than one calculation at the exact same time. Hyperthreading was devised by Intel do make more efficient use of these duplicated sections of the CPU.
That's entirely different from time-slicing, which has been around much longer, & only produces the illusion of two or more actions happening at the same time, by switching rapidly from one to the other. Time-slicing does not increase through-put. It somewhat decreases through-put because there is some overhead in repeatedly stopping one calculation & starting up the other.
The MMX & 3DNOW & SSE & SSE2 & SSE3 sections also allow more than one calculation in the same clock cycle, but the E@H programers apparently haven't gotten around to putting them to work. But in Folding@Home, they greatly increase the throughput.
I don't know why AMD hasn't developed hyperthreading yet. Maybe someone else knows. Maybe Intel holds some patents.
But both Intel & AMD are said to be working on single chips that will contain two complete CPUs for late this year or early next year. That goes further than the P4 which only has duplication for some parts. Maybe AMD believes it has more to gain by concentrating its design force on this further development. I have seen reports that AMD might beat Intel into production.
ADDMP