Nothing to do with Gambit, but I thought I’d compare a Raspberry Pi 2 B (RPi2B) with a RPi3B using the linpack double-precision benchmark.

On Rpi2B compiled with

gcc -mcpu=cortex-a7 -mfpu=neon-vfpv4 -mfloat-abi=hard -Ofast -o linpackc_new linpackc_new.c

I get

cutie-pi:~> ./linpackc_new Enter array size (q to quit) [200]: 200 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 ---------------------------------------------------- 64 0.67 88.31% 2.85% 8.84% 143587.934 128 1.34 88.31% 2.85% 8.84% 143539.628 256 2.69 88.26% 2.88% 8.86% 143610.046 512 5.35 88.28% 2.86% 8.86% 144172.103 1024 10.70 88.28% 2.86% 8.85% 144153.024

Enter array size (q to quit) [200]: 1000 Memory required: 7824K.

LINPACK benchmark, Double precision. Machine precision: 15 digits. Array size 1000 X 1000. Average rolled and unrolled performance:

Reps Time(s) DGEFA DGESL OVERHEAD KFLOPS ---------------------------------------------------- 1 1.62 96.95% 0.59% 2.46% 106370.268 2 3.17 97.78% 0.60% 1.62% 107598.577 4 6.37 97.78% 0.60% 1.61% 107021.647 8 12.70 97.78% 0.60% 1.62% 107340.380

On RPi3 compiled with

gcc -Ofast -mcpu=cortex-a53 -mfpu=neon-vfpv4 -mfloat-abi=hard -Ofast -o linpackc linpackc_new.c

I get

sweety-pi:~> ./linpackc Enter array size (q to quit) [200]: 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 ---------------------------------------------------- 64 0.53 89.82% 2.83% 7.36% 178548.817 128 1.06 89.81% 2.83% 7.37% 178628.648 256 2.13 89.81% 2.82% 7.37% 178550.449 512 4.25 89.81% 2.82% 7.36% 178541.926 1024 8.50 89.81% 2.83% 7.37% 178647.415 2048 17.00 89.81% 2.83% 7.37% 178647.687

Enter array size (q to quit) [200]: 1000 Memory required: 7824K.

LINPACK benchmark, Double precision. Machine precision: 15 digits. Array size 1000 X 1000. Average rolled and unrolled performance:

Reps Time(s) DGEFA DGESL OVERHEAD KFLOPS ---------------------------------------------------- 1 1.02 97.47% 0.61% 1.92% 167558.926 2 2.08 97.42% 0.73% 1.84% 164153.213 4 4.06 97.87% 0.61% 1.52% 167576.385 8 8.13 97.87% 0.61% 1.52% 167558.026 16 16.26 97.87% 0.61% 1.52% 167568.545

Enter array size (q to quit) [200]: q

Both compilers are

sweety-pi:~> gcc --version gcc (Raspbian 4.9.2-10) 4.9.2

This page:

https://www.raspberrypi.org/magpi/raspberry-pi-3-specs-benchmarks/

says of the Rpi3 "the Broadcom BCM2837 system-on-chip (SoC) includes four high-performance ARM Cortex-A53 processing cores running at 1.2GHz with 32kB Level 1 and 512kB Level 2 cache memory, a VideoCore IV graphics processor, and is linked to a 1GB LPDDR2 memory module on the rear of the board.” It also says the RAM runs at 900MHz.

It’s funny, but I can’t find a definitive answer about how much cache the RPi2B has, or what the memory speed is.

At any rate, the ratio of KFLOPS for the “small” (200 x 200) problem is 1.243 and the ratio of KFLOPS for the “large” (1000 x 1000) problem is 1.575.

So the RPi3B must have a much more efficient memory system than the RPi2B.

Brad