Quote from jamesh on September 8, 2011, 09:25
Well, if anyone has a Sheeva plug, can they run the linpack bench mark on it? We have results for the Arm11 on the Raspi so it would be interesting to compare. I couldn't find any results using Google.
This is from my (original model) Sheevaplug running Debian Squeeze
Linux sheeva 2.6.32-5-kirkwood #1 Tue Mar 8 10:56:14 UTC 2011 armv5tel GNU/Linux
and compiled with gcc -O3
Enter array size (q to quit) :
Memory required: 315K.
LINPACK benchmark, Double precision.
Machine precision: 15 digits.
Array size 200 X 200.
Average rolled and unrolled performance:
Reps Time(s) DGEFA DGESL OVERHEAD KFLOPS
4 0.50 88.00% 6.00% 6.00% 11687.943
8 0.99 93.94% 2.02% 4.04% 11564.912
16 2.01 92.04% 2.99% 4.98% 11504.363
32 4.01 92.02% 3.49% 4.49% 11474.326
64 8.02 92.27% 2.87% 4.86% 11519.441
128 16.05 92.27% 2.62% 5.11% 11542.132
So about twice as fast as the R-Pi's software FP, but half as fast as the hardware floating point quoted in the other thread.
I tried encoding a 3 minute MP3 from an uncompressed WAV using lame . The Sheevaplug encodes at about 0.33x realtime - i.e. it takes 9 minutes to encode a 3 minute track. It'd be interesting to see how the R-Pi compares with the faster floating point, faster ram but slower processor. Flac encoding the same WAV file takes about 40s on the Sheeva. I don't think there's much floating point math involved in that, though.
Also, is response the the Ram is Ram comment above.....well, yes, but the memory subsystem (ie the silicon used to handle access to the memory) is very important in how well that RAM is used. The one in the Raspi SoC is very (very) good.
Sounds great, and would make up for the fewer mhz in the processor when compared with the Sheeva's ARM. It's good as long as you stay within the bounds of the RAM, and don't hit the swap, though.
If the base system is fairly lean, and care is taken about Ram usage in the programs you're running (or programming) it should be OK. It'll be interesting to see how the little R-Pi performs generally. I'm not sure how much high-end photo-editing I'd be doing on it, though!
What would be great if there was an easy way for the system (by extension the user and programmer) to make use of the GFLOPS that the GPU is reported to have - it would make the little R-Pi card a mathematical powerhouse.