Hardware gurus delight:

I've got 3 'basket crunchers' based on AMD XP 2000 CPUs. The data below if from Sandra Lite 2005, CPU-Z v1.27 and BOINC CC v4.26. One of them is clocked at 2062.5 and reports as a XP2600. Not a part of this situation.

But here's the data on the other two:

Host---------Freq----------DDRAM------Sandra-----------BOINC----------Sandra
crunch1___2011.3____2.5-3-2-7__3244 8409____1801 3073___3104 18.4
crunch4___1985.7____2.5-3-2-7__3127 8216____1847 3126___3065 18.4

The 1st pair of Sandra numbers are from the CPU benchmark. Followed by BOINC benchmark numbers and then the Sandra Cache/Mem Bandwidth numbers.

These numbers are produced with nothing else running, CA EZ firewall locked, AV suspended and Task manager reports the same processes running except TCPSVCS is running on crunch4 (it is using the mobo based lan port, Crunch1 is using a PCI slot based one with the mobo one disabled). These are both running the same mobo, same XP Pro SP2. They both have 1 256Mb stick of PC3200 in slot 2.

Crunch4 has a much older, smaller, slower HDD.
Crunch4 has the FSB set to 159.
Crunch1 has the FSB set to 161.

The sandra numbers show Crunch1 to be about 2.4% faster... seems about right. It's only the BOINC benchmark that seems backwards. I've run all these benchmarks multiple times and the results are consistant. I've reinstalled v4.26 on Crunch1, no change.

Now here is one difference between the two of 'em... One is a THORTON core and one is a THOROUGHBRED core. But I've never been able to measure any differences in them before.

I've already spent more time on this than is sane, but if anyone has any thoughts or if the BOINC Dev guys can think of any of the v4.2x benchmark changes done that would prefer the T-Bred core, lemme know.

Thanx, Skip

---

- da shu @ HeliOS,
"A child's exposure to technology should never be predicated on an ability to afford it."

> Hardware gurus delight:
> I've already spent more time on this than is sane, but if anyone has any
> thoughts or if the BOINC Dev guys can think of any of the v4.2x benchmark
> changes done that would prefer the T-Bred core, lemme know.
>
> Thanx, Skip
>
The problem is in part that everyone that writes a benchmarking program does it with their own idea of how to do it. In that mind set Berkeley came along and decided to test your computer for what it decided was important to it. No other program tests your computer in the same way.

---

> I've already spent more time on this than is sane, but if anyone has any
> thoughts or if the BOINC Dev guys can think of any of the v4.2x benchmark
> changes done that would prefer the T-Bred core, lemme know.
>
> Thanx, Skip

It is said that if the Aborigine drafted an I.Q. test, all of Western civilization would presumably flunk it.

Same thing with benchmarks: it is hard to design one that is 100% neutral.

---

Also bare in mind that SiSoft Sandra has become completely useless after 2001TE Version.
Everything it turns out CPU-wise is completely unrealistic (way too high, complete nonsense). That Program actually doesn't seem to even really Benchmark some Paramaters, it rather adds as much as it can to achieve marketing numbers :p

Add to that, that the SETI Client relies heavily on RAM bandwidth (especially on AthlonXP Systems) and also amount of Cache, while the Sandra/BOINC Benchmarks hardly even consider RAM bandwidth.

That leads to the "synthetic" Results vs. "real" Results.
(I have Athlon64 Systems synthetically benched slower than AthlonXP's, but the real Results are of course different)

Chances are, your 2 Systems are simply not 100.0% identically configured (as you wrote), thus one is working a tiny bit slower.
A single Driver, a single PCI Card or Enabled Onboard-Devices and an additional IRQ to share can already make the difference.

PS.
8 GFlops on an AthlonXP according to Sandra?
That about tells something about the "quality" of this "Benchmark"

My advice :
Have them run SETI Workunits (either alot of them to get a useful average figure, or compare them to have at least identical Angle Range).
Only then you will see the actual Performance of the Systems.

The AthlonXP Thoroughbred core was AMD's improved 0,13 µm Version of the old Palomino Core to achieve higher Clock rates.
The AthlonXP Thorton Core is basically a Barton Core, but with disabled 50% of the L2 Cache (there have been reports of people being able to re-activate the full L2 Cache).
From all I know, those were not meant for the Retail market, but were planned solely for OEMs to allow sales of Barton Cores that did not pass the validation for 512kB of functional Cache after coming from the production wafer.
(still better for AMD to sell as fully functional 256k L2 AthlonXP's than trash them)

Per Default, both CPU's should perform 100% identical, but their Core specifications, DIE Shape/Area, Thermal behaviour and Typical Power Dissipation (consumption) differ :

Athlon XP-2000+ (6-A-0 1.67GHz Thorton) Vcore: 1.6V TPD : 47.4W
Athlon XP-2000+ (6-8-1 1.67GHz Thoroughbred) Vcore: 1.6V TPD: 54.7W

The Thorton, due to larger DIE area, is easier to Cool and saves you a few Watts of Power Consumption as well.

---

> Chances are, your 2 Systems are simply not 100.0% identically configured (as
> you wrote), thus one is working a tiny bit slower.
> A single Driver, a single PCI Card or Enabled Onboard-Devices and an
> additional IRQ to share can already make the difference.

The process tab in task manager shows identicle processes running except the 1 extra one I listed (that's on the BOINC faster CPU) in the original post. What else can I check to ensure equality?

Yes, the SANDRA numbers are silly but useful for comparing = CPUs to each other.
OK, forget the other benchmarks... let's just consider the BOINC v4.26 benchmark numbers. As long as I'm comparing apples to apples I don't think it makes any difference.

The real point here was that by the BOINC benchmark the one runnnin a 2Mhz faster FSB is slower.


> My advice :
> Have them run SETI Workunits (either alot of them to get a useful average
> figure, or compare them to have at least identical Angle Range).
> Only then you will see the actual Performance of the Systems.

I'll see what I can do along these lines

> From all I know, those were not meant for the Retail market, but were planned
> solely for OEMs to allow sales of Barton Cores that did not pass the
> validation for 512kB of functional Cache after coming from the production
> wafer.

Yes, I bought them in China as OEM chips.


> Per Default, both CPU's should perform 100% identical, but their Core

No, Actually the one running 161Mhz FSB should out benchmark the one running 159Mhz FSB. I also think that 2Mhz difference in the FSB should more than compensate for some minor differences in driver versions but I'll see if I can confirm = driver versions on the things.

> specifications, DIE Shape/Area, Thermal behaviour and Typical Power
> Dissipation (consumption) differ :
>
> Athlon XP-2000+ (6-A-0 1.67GHz Thorton) Vcore: 1.6V TPD : 47.4W
> Athlon XP-2000+ (6-8-1 1.67GHz Thoroughbred) Vcore: 1.6V TPD: 54.7W

Actually MBM reports the T-Bred to run cooler but that could be as simple as I did a better job of Artic Silvering the cpu/cooler. But I've read elsewhere that the T-Breds often run a bit cooler than the Thortons. Dunno why.

Thanx for the input, Skip

---

- da shu @ HeliOS,
"A child's exposure to technology should never be predicated on an ability to afford it."

Hm, since the Taskmanager shows only a fraction of the actual Background tasks, more detailed analysis with a 3rd Party Utility might reveal something (e.g. TaskInfo 2000 or alike).

I agree that (normally) the higher FSB should result in a small increase in performance running SETI/BOINC; looks like something in the config eats up this advantage.

Just be aware that the naked BOINC Benchmark has a small inaccuracy (almost every Benchmark has).
Maybe re-running the Benches (under Zero Workload) again reveals different enough Results to explain the different numbers, that would solve the small puzzle rather quick.

---