[Eug]

Code:
System                     L           XL          XXL
--------------------------------------------------------
Quad 1.6GHz Opteron       0.33        1.5         17.98
Dual 3.06GHz Xeon         0.28        1.33        14.48
3.2GHz P4                 0.28        1.33        10.37*
2.66GHz P4                0.33        1.85        17.58
Dual 1.25GHz G4           0.44        2.46        29.81
1.8GHz G5                 0.26        1.44        13.02
Dual 2GHz G5              0.23*       1.29*       11.66

* = fastest

L and XL cases are single precision
XXL case is double precision
(XXL only runs properly in double precision).

PGF compiler was used on Opteron
PGF and IFC on P4/Xeon
VAST/gcc-3.3/XLF on G4/G5

Fastest score posted in each case, 
so some P4/Xeon results are with IFC, some with PGF.

Well, the single 1.8 thrashes the dual G4 1.25 mainly because this app is not dual aware. However, it's interesting to note that the single G5 1.8 is 2.3X as fast as a single G4 1.25. So even if it were dual aware, it'd thrash the dual G4 1.25, and probably the dual 1.42 too. And while it thrashes the Opteron 1.6 too, it only merely beats the Xeon 3 GHz.

Actually, I'm more interested in the fact that the G5 2.0 completely dominates the competition in this bench, once the xlf compiler is used. Data is from here.

[MindFad]

I'm a bit confused by the numbers, but I see it now. Smaller = faster, right?But why didn't the *dual* Xeon do any better than the P4, and why did the P4 beat the G5 in XXL. Keep in mind, I have no friggin' idea what any of that actually is. It's cool to see the G5 competitive in the least and kicking ass most of the time, according to NASA.

[alien]

MindFad, this bench is not dual processor aware, so I think the 3.06 Xeon is about where it should be; between the 2.66 and 3.2 P4s.

For the L and XL problems, AltiVec kicks in on the G4 and G5 (VAST is used to vectorize). For the XXL problem, double precision is used, so now AltiVec can't be used anymore and the G4 performance gets very poor. (It could have to do with memory bandwidth too I guess.)

But the G5 has two floating point units (each capable of multiplication and addition), and plenty of bandwidth.

The P4 on the other hand uses SSE/SSE2 all the way for floating point, and while this is no match for the G5 with AltiVec, the 1.2GHz difference gets a tad too much on the XXL problem size (G5 w/o AltiVec).

That's my (speculative) understanding, anyway...

[MindFad]

*Not* dual aware? Lordy, the dual 2.0 G5 *is* a beast. Man, I want one. Thanks for th info.

[olePigeon]

Originally posted by alien:
For the L and XL problems, AltiVec kicks in on the G4 and G5 (VAST is used to vectorize). For the XXL problem, double precision is used, so now AltiVec can't be used anymore and the G4 performance gets very poor. (It could have to do with memory bandwidth too I guess.)

G4 and G5 both don't have very impressive interger marks. But if you're running apps that take full advantage of AltiVec and are reliant on floatingpoint, the G5 just slaughters everything. Examples are Maya 3D, LightWave, Genom and DNA sequencing, etc. Dual G5 has some serious ownage in those areas.

[Eug]

Keep in mind, I have no friggin' idea what any of that actually is.

Computational Fluid Dynamics.

I don't do any of this either, but I just found the numbers interesting...

One reason I found this so interesting is because the fast compilers are used with the Intel stuff as well. ie. the test was on even ground for both the P4/Xeon and for the G5.

[idyll]

Careful when you post topics like this, you might make awcopus cty .

[hyteckit]

Oh course the G5 1.8GHz trashes the Dual 1.25GHz G4. It's more like G5 1.8GHz trashes the Single 1.25GHz G4 which is no big deal. Nasa speed test is only for single processor.

Look at the Dual 3.0GHz Xeon and Single 3.2GHz P4. The Dual 3.0GHz Xeon is slower than single 3.2GHz P4? Hehe...

[Eug]

Originally posted by hyteckit:
Oh course the G5 1.8GHz trashes the Dual 1.25GHz G4. It's more like G5 1.8GHz trashes the Single 1.25GHz G4 which is no big deal. Nasa speed test is only for single processor.

Look at the Dual 3.0GHz Xeon and Single 3.2GHz P4. The Dual 3.0GHz Xeon is slower than single 3.2GHz P4? Hehe...

Obviously you didn't read everything in my post because I already said that.

But I also said the single 1.8 is up to 2.3X as fast as the single 1.25 (despite only having 1.4X the GHz).

[blackwind]

Originally posted by hyteckit:
Oh course the G5 1.8GHz trashes the Dual 1.25GHz G4. It's more like G5 1.8GHz trashes the Single 1.25GHz G4 which is no big deal. Nasa speed test is only for single processor.

It is a fairly big deal for per-MHz comparisons between the G4 and the G5, considering that many typical tests seem to show the G4 having an advantage.

Theoretical 1.8-GHz G4 with Perfect Linear Scaling:

L: 0.30 s
XL: 1.71 s
XXL: 20.70 s

The Altivec-enabled L and XL tests still show the G5 having better per-MHz performance despite a weaker Altivec implementation (thank you, FSB! ), and the double-precision XXL test shows the G5 having a 59% per-MHz advantage over the G4.

This is even more impressive:

Theoretical 2-GHz Opteron with Perfect Linear Scaling:

L: 0.26 s
XL: 1.20 s
XXL: 14.38 s

The fact that the fastest G5 is faster than the fastest Pentium 4C, the fastest Pentium 4 Xeon, and by linear extrapolation, the fastest Opteron at 2-GHz is even more impressive since it shows how competitive the G5 is with other PC's in at least computational fluid dynamics. (Thank you, fused add-multiply! )

Ay, I am hyper this morning. I need to do some depressing work.