|
AltiVec Fractal Carbon is a numerically-intensive parallel
graphics application that uses the Velocity Engine, a.k.a. AltiVec for
its computations. Version 1.3 was used in this experiment.
Results
Each row in the following table shows the average MegaFlops (out
of 3 runs) obtained on each system. A score is then assigned to each
platform, with a reference score of 1.00 for a run under MacOS 9.0.4
within SheepShaver for MacOS X. All guest systems were MacOS 9.0.4
except PearPC which run MacOS 10.2.
| Host Operating System |
CPU |
Speed |
AltiVec |
MegaFlops |
Score |
| MacOS X 10.2.8 |
G4 |
400 |
Yes |
1300.7 |
1.52 |
| MacOS X 10.2.8 [SheepShaver] |
G4 |
400 |
Yes |
854.3 |
1.00 |
| Linux 2.6.8.1 [SheepShaver] |
Xeon |
3200 |
Yes |
1013.7 |
1.18 |
| Windows XP SP1 [PearPC] |
Xeon |
3200 |
Yes |
517.7 |
0.60 |
| MacOS X 10.2.8 |
G4 |
400 |
No |
298.3 |
0.35 |
| MacOS X 10.2.8 [SheepShaver] |
G4 |
400 |
No |
199.5 |
0.23 |
| Linux 2.6.8.1 [SheepShaver] |
Xeon |
3200 |
No |
108.3 |
0.13 |
|
Comments
SheepShaver AltiVec results were achieved on Linux/x86-64 with a
snapshot from 2005/03/14.
PearPC results were obtained with a 0.4pre build with AltiVec
support on Windows. That's the prasys
build as of 2005/03/16.
The program displays some glitches while computing a
scene. However, this only happens when AltiVec is enabled and the
program is running under SheepShaver for MacOS X.
There is no specific optimization in the SheepShaver AltiVec
emulation code but some SSE. Even though the current JIT is very basic
and does not allocate SSE registers, performance is again on par with
a run within SheepShaver on a PowerPC G4/400.
Contrary to what people usually claim on certain forums, PearPC is
clearly not the fastest PowerPC G4 emulator, and definitely not
the first.. The AltiVec emulation is currently twice as slow as
SheepShaver, even though it performs register allocation of SSE
registers. This is reasonnable though.
|
