Raspberry Pi vs Power Mac G5

[AphoticD]

Out of interest, I just bought a Raspberry Pi 3 Model B+ to play around with. The new Pi system has a number of improvements over the previous model, including improved cooling and a bumped up CPU which is now:

• SoC: Broadcom BCM2837B0 quad-core A53 (ARMv8) 64-bit @ 1.4GHz

I haven't had any prior experience with these mini computers, but I think it performs incredibly well for web browsing and Youtube with it's default Raspbian OS and it handles UI graphics smooth and responsive when connected over HDMI to my 27" display.

Having onboard wifi means you can just connect it to power (5V, 2.5A) after initial config and wirelessly login via SSH or VNC without using a USB dongle or any other connected devices. This makes it a tidy little box for a media server or a portable little sandbox server for PHP/python development.

To disable the automatic GUI / login on boot, run raspi-config and change Boot Options > B1 Desktop / CLI to "B1 Console".

After reading some benchmarks, I tried replicating some of them with both the 2018, passive-cooled ARM based Quad Core Pi @ 1.4Ghz running Raspbian and the late 2005, air-cooled Dual Core PowerPC (970mp) G5 @ 2.3Ghz running Ubuntu Mate.

pi:~ $ uname -a
Linux 4.9.80-v7+ #1098 SMP Fri Mar 9 19:11:42 GMT 2018 armv7l GNU/Linux

g5:~ $ uname -a
Linux 4.4.0-116-powerpc64-smp #140-Ubuntu SMP Mon Feb 12 22:00:40 UTC 2018 ppc64 ppc64 ppc64 GNU/Linux

Sysbench (0.4.12) CPU Benchmark
(Thread count has been matched to the number of cores)

pi:~ $ sysbench --test=cpu --num-threads=4 run
...
Test execution summary:
    total time:                          33.8234s
    total number of events:              10000
    total time taken by event execution: 135.2127
...

g5:~ $ sysbench --test=cpu --num-threads=2 run
...
Test execution summary:
    total time:                          31.9690s
    total number of events:              10000
    total time taken by event execution: 63.9280
...

CPU Benchmark Results:
Pi CPU Test: 33.82 seconds
----------
G5 CPU Test: 31.96 seconds

Lower results are better. As you can see, the mighty G5 is only slightly quicker (~5.5%) than the little Pi!


Sysbench Memory Benchmarks (READ/WRITE)
Pi Memory Read

pi:~ $ sysbench --test=memory --memory-block-size=1K --memory-total-size=10G --memory-oper=read run
...
Operations performed: 2097152 (870860.93 ops/sec)

2048.00 MB transferred (850.45 MB/sec)

Test execution summary:
    total time:                          2.4081s
    total number of events:              2097152
    total time taken by event execution: 1.7553
...

Pi Memory Write

pi:~ $ sysbench --test=memory --memory-block-size=1K --memory-total-size=10G --memory-oper=write run
...
Operations performed: 2097152 (652977.86 ops/sec)

2048.00 MB transferred (637.67 MB/sec)

Test execution summary:
    total time:                          3.2117s
    total number of events:              2097152
    total time taken by event execution: 2.5592
...

G5 Memory Read

g5:~ $ sysbench --test=memory --memory-block-size=1K --memory-total-size=10G --memory-oper=read run
...
Operations performed: 2097152 (1202025.79 ops/sec)

2048.00 MB transferred (1173.85 MB/sec)

Test execution summary:
    total time:                          1.7447s
    total number of events:              2097152
    total time taken by event execution: 1.1654
...

G5 Memory Write

g5:~$ sysbench --test=memory --memory-block-size=1K --memory-total-size=10G --memory-oper=write run
...
Operations performed: 2097152 (856017.41 ops/sec)

2048.00 MB transferred (835.95 MB/sec)

Test execution summary:
    total time:                          2.4499s
    total number of events:              2097152
    total time taken by event execution: 1.8698
    per-request statistics:
...

Memory Benchmark Results:
Pi Memory Read: 850.45 MB/sec
Pi Memory Write: 637.67 MB/sec
EDIT: Tests re-run with --num-threads=4
Pi *Multithreaded* Memory Read: 2749.16MB/sec
Pi *Multithreaded* Memory Write: 2271.43MB/sec
----------
G5 Memory Read: 1173.85 MB/sec
G5 Memory Write: 835.95 MB/sec
EDIT: Tests re-run with --num-threads=2
G5 *Multithreaded* Memory Read: 1288.34MB/sec
G5 *Multithreaded* Memory Write: 852.00MB/sec

Ok, so [single-threaded] memory performance of the Pi isn't quite as good as the G5 and it is massively limited at only 1GB of RAM, as opposed to the G5's ceiling of 16GB.

EDIT: Multithreaded memory tests show memory throughput on the Pi to be more than twice as fast as the G5!

Real World Benchmarks
Now for a real world test. Compiling QEMU 2.11.1 from source, after installing dependancies.
(make is executed with 2 threads per CPU e.g. make -j 8)
Pi configure

pi:~/Downloads/qemu-2.11.1/build$ time ../configure --target-list=ppc64-softmmu,ppc-softmmu,i386-softmmu,x86_64-softmmu --disable-spice
...
real    0m22.329s
user    0m14.970s
sys    0m2.850s

Pi make

pi:~/Downloads/qemu-2.11.1/build$ time make -j 8
...
real    18m29.085s
user    66m50.310s
sys    2m50.230s

G5 configure

g5:~/Downloads/qemu-2.11.1/build$ time ../configure --target-list=ppc64-softmmu,ppc-softmmu,i386-softmmu,x86_64-softmmu --disable-spice
...
real    0m12.802s
user    0m7.136s
sys    0m2.928s

G5 make

g5:~/Downloads/qemu-2.11.1/build$ time make -j 4
...
real    12m26.464s
user    22m22.660s
sys    1m49.316s

QEMU Build Time Results:
Pi configure and build time: ~19 minutes
----------
G5 configure and build time: ~12.5 minutes


Booting The Mac OS 9.2.2 Universal iso in QEMU (PPC, no KVM)

pi:~ $qemu-system-ppc -M mac99 -m 128 -cdrom ~/vm/macos922uni.iso -boot d -vnc :5,password -monitor stdio

Pi Time to boot and show the Mac OS 9.2.2 Read Me: 4 minute 01 seconds

g5:~ $ qemu-system-ppc -M mac99 -m 128 -cdrom ~/vm/macos922uni.iso -boot d

G5 Time to boot and show the Mac OS 9.2.2 Read Me: 1 minute 27 seconds


Booting Debian 9 x86_64 img in QEMU (x86_64, no KVM) [With thanks to @Lastic]

pi:~ $ qemu-system-x86_64 -smp cpus=4 -boot c -hda vm/debian9_x64.img -m 512 -net nic -net user -vnc :5,password -monitor stdio

Pi Time to boot to user login prompt (no GUI): N/A (I gave up after 13 minutes of zero progress)

g5:~ $ qemu-system-x86_64 -smp cpus=2 -boot c -hda vm/debian9_x64.img -m 512 -net nic -net user -vnc :5,password -monitor stdio

G5 Time to boot to user login prompt (no GUI): 3 minutes 52 seconds


Booting KolibriOS in QEMU (i386, no KVM)

pi:~ $ qemu-system-i386 -m 256 -cdrom kolibri.iso -boot d -vnc :5,password -monitor stdio

Pi Boot to Desktop: ~11 seconds

g5:~ $ qemu-system-i386 -m 256 -cdrom kolibri.iso -boot d -vnc :5,password -monitor stdio

G5 Boot to Desktop: ~6 seconds

Both machines could run a smooth game of Doom in an emulated x86 system delivering graphics over VNC, running KolbriOS (the Pi is running over wifi, the G5 is on gigabit).


EDIT: Despite these two systems being capable of performing CPU based multithreaded activities at similar speeds (with other factors causing bottlenecks, e.g read/write of the 16GB micro-SD card vs the G5's 7200rpm SATA 1TB HDD), the clear highlight here is that energy consumption has been the real advancement in the 13 years between these two systems. The G5 has a 1kW capable PSU with a real world draw of approx 150 - 250 watts per hour. The Pi on the other hand will draw ~6w under full load, with an idle of ~2w.

At least the final PowerPC Mac, the Mighty Quad G5 will continue to stand up against a few more revisions of the Pi until it's beat out by this tiny opponent.

 

[Dronecatcher]

I bought the first model Pi when they were released...but took it back when configuring it to do anything was such an uphill struggle...my impatience again. I think the process is a lot slicker now.
Will have to revisit soon but I find the Pi Zero more alluring being an even more reduced form factor...would love to bung one into an old radio chassis and make a portable Software Defined Radio.

 

[AphoticD]

I bought the first model Pi when they were released...but took it back when configuring it to do anything was such an uphill struggle...my impatience again. I think the process is a lot slicker now.
Will have to revisit soon but I find the Pi Zero more alluring being even an more reduced form factor...would love to bung one into an old radio chassis and make a portable Software Defined Radio.

I can imagine the early revisions would have been challenging. But they have come a long way in terms of ease of use.

When I first plugged it in, it took maybe 30 seconds on initial boot to sort itself out and then I was staring at a ready to use desktop UI, with a very responsive Chromium as the default browser. I'm impressed with this little device and have been stepping through some of their projects on the Pi site for fun. I can see how good these can be for schools to teach the kids about "building" a computer, identifying what's what on the system board, and even learning to code with Python, etc

It also made me think about how second-hand lower spec PowerPC Macs could easily fulfil the same need at roughly the same price point (and often less). Something like an iBook G4 (with Airport Extreme) would be perfect for kids to learn on without too many distractions.

 

[AphoticD]

For reference, here are the same benchmarks from my 2008 Mac Pro 3,1 (2x Quad Core 3.2Ghz Xeon).

Mac Pro CPU Test

cMP:~ $ sysbench --test=cpu --num-threads=8 run
...
CPU speed:
    events per second:  8268.55

General statistics:
    total time:                          10.0010s
    total number of events:              82718
...

Mac Pro Memory Read *Multithreaded*

cMP:~ $ sysbench --test=memory --threads=8 --memory-block-size=1K --memory-total-size=10G --memory-oper=read run
...
Number of threads: 8
...
Total operations: 10485760 (23677316.36 per second)

10240.00 MiB transferred (23122.38 MiB/sec)

General statistics:
    total time:                          0.4401s
    total number of events:              10485760
...

Mac Pro Memory Write *Multithreaded*

cMP:~ $ sysbench --test=memory --threads=8 --memory-block-size=1K --memory-total-size=10G --memory-oper=write run
...
Number of threads: 8
...
Total operations: 10485760 (4491974.58 per second)

10240.00 MiB transferred (4386.69 MiB/sec)

General statistics:
    total time:                          2.3316s
    total number of events:              10485760
...

Memory tests also benefit from being multithreaded. I'll re-run the Pi and the G5 memory tests again to reflect multithreaded performance.

Compiling QEMU benchmarks:

cMP configure

cMP:~ $ time ../configure --target-list=ppc64-softmmu,ppc-softmmu,i386-softmmu,x86_64-softmmu --disable-spice --python=/usr/bin/python
...
real    0m9.141s
user    0m4.102s
sys    0m4.809s

cMP build

cMP:~ $ time make -j 16
...
real    1m46.595s
user    8m29.961s
sys    1m47.083s

Total configure and build time: less than 2 minutes.

 

[Lastic]

Interesting comparison.

I once coupled an Odroid-W to a Motorola lapdock and was also impressed of it running Qemu and mplayer quite good.

However fried it trying to add a flat USB connector and now the lapdock serves as a HDMI KVM for my Xserve G4

 

[marcofranchimoretti]

Welcome at all. I will post my contribute.
Libre Computer Le Potato
[https://libre.computer/products/boards/aml-s905x-cc/]

Linux lepotato 4.14.52-meson64 #6 SMP PREEMPT Thu Jun 28 11:17:29 UTC 2018 aarch64 GNU/Linux. [Armbian]

sysbench 0.4.12: multi-threaded system evaluation benchmark
Running the test with following options:
Number of threads: 4
Doing CPU performance benchmark
Threads started!
Done.
Maximum prime number checked in CPU test: 10000
Test execution summary:
total time: 2.6804s
total number of events: 10000
total time taken by event execution: 10.7122
per-request statistics:
min: 1.03ms
avg: 1.07ms
max: 33.78ms
approx. 95 percentile: 1.05ms
Threads fairness:
events (avg/stddev): 2500.0000/24.69
execution time (avg/stddev): 2.6780/0.00

sysbench --test=memory --memory-block-size=1K --memory-total-size=10G --memory-oper=read run
sysbench 0.4.12: multi-threaded system evaluation benchmark
Running the test with following options:
Number of threads: 1
Doing memory operations speed test
Memory block size: 1K
Memory transfer size: 10240M
Memory operations type: read
Memory scope type: global
Threads started!
Done.
Operations performed: 10485760 (1259478.29 ops/sec)
10240.00 MB transferred (1229.96 MB/sec)
Test execution summary:
total time: 8.3255s
total number of events: 10485760
total time taken by event execution: 5.9370
per-request statistics:
min: 0.00ms
avg: 0.00ms
max: 0.17ms
approx. 95 percentile: 0.00ms
Threads fairness:
events (avg/stddev): 10485760.0000/0.00
execution time (avg/stddev): 5.9370/0.00

sysbench --test=memory --memory-block-size=1K --memory-total-size=10G --memory-oper=write run
sysbench 0.4.12: multi-threaded system evaluation benchmark
Running the test with following options:
Number of threads: 1
Doing memory operations speed test
Memory block size: 1K
Memory transfer size: 10240M
Memory operations type: write
Memory scope type: global
Threads started!
Done.
Operations performed: 10485760 (859664.11 ops/sec)
10240.00 MB transferred (839.52 MB/sec)
Test execution summary:
total time: 12.1975s
total number of events: 10485760
total time taken by event execution: 9.8195
per-request statistics:
min: 0.00ms
avg: 0.00ms
max: 0.14ms
approx. 95 percentile: 0.00ms
Threads fairness:
events (avg/stddev): 10485760.0000/0.00
execution time (avg/stddev): 9.8195/0.00

P.S. All for terminal emulator under XFCE with browser Chromium open.

 

[llama233]

Dusting off this ancient thread, I hope that's OK...

I used to own a Dual 800 G4 Quicksilver PPC, then used a PowerBook 12" 1.5 from 2006 until I finally got on the Intel bandwagon in 2013 with a Retina MacBook Pro. I also ran two G5 Xserves for a few years at work.

Recently I was starting to get the itch to buy a PowerMac G5 to tinker with and that sent me down the rabbit hole of available software etc these days, and how slowly it now ran, and some Googling led me to this thread!

I have a bunch of Raspberry Pi's doing various things in my home so thought I'd try and replicate the QEMU compile above to see just how far we've come in ~20 years. The first problem was getting QEMU 2.11.1 compiling as it's pretty old now, the easiest way I found was the spin up an Ubuntu 14.04 LXC container as that had the necessary (older) support software needed for setting up the build environment and compilation.

Anyway, enough waffle - my set up is a Raspberry Pi 4 with 8GB RAM and a Raspberry Pi 5 with 4GB RAM, both running on external SSD drives in USB enclosures. The actual software build is running in an LXC container running on top of the Pis.

G5 configure and build time: ~12.5 minutes (taken from the post above)
Pi 4 configure and build time: ~11.5 minutes (configure 0:26, build 11:09)
Pi 5 configure and build time: ~4 minutes (configure 0:08, build 3:51)

So a Pi 4 is *marginally* faster than a Quad G5 at compiling that one app, a Pi 5 smokes both of them (at least in this configuration - I think the speed may be down more to the USB interface to the disk than anything else? It would be interesting to try an NVMe SSD drive).

It has made me rethink the whole nostalgia aspect of running a noisy, large G5 consuming ~250 watts when I get way more power out of a board that fits in my hand using ~8 watts (and can power 4K displays!), but nostalgia isn't ever rational 😀

 

[AphoticD]

This is great! Thanks for bumping an old thread. The Pi 5 is certainly impressive in compile time, and well done matching the qemu version on this almost 8 year old thread.

I still have that Pi 3 Model B+ running on my desk. It's spent almost every day of its life silently serving music on my local network and streaming internet radio. The G5 is still here too along with 3 of its G5 brothers, 4 G4 towers, a couple cMPs, and a beige G3 from '97. Amazing much of this old tech still runs. Even at 250w/hr or more, but that's just a small part of the price of the retro Mac bug.

But its more than just the cost and heat produced by these things. There's a great deal of joy in going down the nostalgia rabbit hole, and there's good they can still be used for, so long as you're not expecting to browse modern youtube or social media sites.

There is no rational reason to spend money on this old tech, and I've been thinking lately that anyone observing how much time and money I have put into these old macs would think there was something mentally wrong with me. (My wife included)

So in saying all that, maybe rethinking your G5 hunt could save you some time, money and likely a number of headaches... but then again, maybe you'd have some fun? 😎

 

[Doq]

running a noisy, large G5 consuming ~250 watts

250 is a pretty optimistic value, especially for a Quad, especially especially at full tilt.

My Quad peaks out at almost double this. Still far less than the 1000W power supply inside of it, but if it was 250 I might have actually considered using it more than I currently do.

 

[Certificate of Excellence]

I always wanted to try this awesome Pi use case - and I will, eventually. I mean why use 8 when you can use 258w lol 😀

 

[Matias_]

One of my favorite benchmarks is Thrawn's x264 video encoding suite. You can (and often will) use hand crafted assembly and/or SIMD optimizations for your particular platform so it's really a best case scenario and video encoding scales nicely. On PPC AltiVec can be fully leveraged. It's a very long benchmark (runtime on older system is expressed in hours or days) so I think it provides a decent average of performance.

Anyway, here you can see that a G5 Quad beats many (but not all) Raspberry Pi 4 setups and gets absolutely smoked by a Raspberry Pi 5: http://www.xin.at/x264/index-en.php?comp_id=8&search="powerpc"+"raspberry"&systemclass=allsystems

 

[llama233]

One of my favorite benchmarks is Thrawn's x264 video encoding suite. You can (and often will) use hand crafted assembly and/or SIMD optimizations for your particular platform so it's really a best case scenario and video encoding scales nicely. On PPC AltiVec can be fully leveraged. It's a very long benchmark (runtime on older system is expressed in hours or days) so I think it provides a decent average of performance.

Anyway, here you can see that a G5 Quad beats many (but not all) Raspberry Pi 4 setups and gets absolutely smoked by a Raspberry Pi 5: http://www.xin.at/x264/index-en.php?comp_id=8&search="powerpc"+"raspberry"&systemclass=allsystems

That's really interesting, and in line with what I was seeing (Pi 4 basically matches G5, Pi 5 is way faster than both, though I'll admit I only did one run on one bit of software, not a whole suit of tests).

This is great! Thanks for bumping an old thread. The Pi 5 is certainly impressive in compile time, and well done matching the qemu version on this almost 8 year old thread.

I still have that Pi 3 Model B+ running on my desk. It's spent almost every day of its life silently serving music on my local network and streaming internet radio. The G5 is still here too along with 3 of its G5 brothers, 4 G4 towers, a couple cMPs, and a beige G3 from '97. Amazing much of this old tech still runs. Even at 250w/hr or more, but that's just a small part of the price of the retro Mac bug.

But its more than just the cost and heat produced by these things. There's a great deal of joy in going down the nostalgia rabbit hole, and there's good they can still be used for, so long as you're not expecting to browse modern youtube or social media sites.

There is no rational reason to spend money on this old tech, and I've been thinking lately that anyone observing how much time and money I have put into these old macs would think there was something mentally wrong with me. (My wife included)

So in saying all that, maybe rethinking your G5 hunt could save you some time, money and likely a number of headaches... but then again, maybe you'd have some fun? 😎

That fun factor can't be underestimated! I've spent many hours and a stupid amount of money over the years just tinkering with old systems that really had no reason to be running at all, and enjoyed (mostly!) every moment.

As well as the mentioned G4 tower and G4 PB 12" I also had an early intel Mac Mini, Sun Ultra 5 (maxed out the RAM to 512MB!), Ultra 10, 2x SGI O2's, an SGI Octane and a whole slew of old Intel workstations (early HPs etc). Yeah, you could say I'm a sucker for a purpose build Unix workstation 😁

I slowly got rid of them all about a decade ago (sold some, donated others to keen collectors) and now I'm again feeling the pull of the (once!) high powered workstation calling out, SGI and old Sun kit is silly money these days, G5's still seem to be reasonable in New Zealand and I did always like the look of those G5 towers, so that's where I'm leaning. It's also nice that the G5 is still reasonably usable day to day unlike the old Sun and SGI kit.

 

[llama233]

I always wanted to try this awesome Pi use case - and I will, eventually. I mean why use 8 when you can use 258w lol 😀

Oh man that's awesome! Right up my alley, I'll have to look into something like this.

I love building Pi's into enclosures, here's an early prototype of what would eventually be called "The Bakery" - an old Cisco switch chassis with 4x Pi 2's, an ethernet switch, PSU, and eventually a 4TB 2.5" external drive. It was used for hosting small websites and offsite backups, I put a cool mesh over the front and had "BAKERY" printed on metal letters so it would look cool in the rack (I worked for a web hosting outfit back then so had a few U of free rack space). Note one of my O2's making a cameo in the background.

Why bakery? I was trying to make a pun on pies inside bakeries 😉