Benchmarking using simplebench script
A while ago I used a simple benchmark to very roughly compare performance on multiple platforms: PC-6002 vs 80s Computers Benchmark and got some interesting results.
Recently I got a Raspberry Pi 4 and wanted to figure out how its new CPU compare to other platforms so I went back to that simple benchmark I used, scaled it up by 1000x and used it in many different ways on many different devices and platforms. I think the results are noteworthy :) but it’s still just for fun, this is by no means a benchmark that should be taken seriously.
The simple bench I used looks like this:
import math import time import os all_primes =  t1=time.time() skip=False for i in range(2, 100000): skip=False k = math.floor(math.sqrt(float(i))) + 1.0 for j in range(2, int(k)): k1=i/float(j) k2=int(k1) if k1==k2: skip=True break if skip: continue all_primes.append(i) elapsed=(time.time() - t1) print("Prime count = " + str(len(all_primes))) print("Python Time=" + str(elapsed))
Here are the systems I ran simplebench on:
- Raspberry Pi 4: Cortex A72 1.5GHz, Raspbian
- Desktop PC: AMD Ryzen 5 1600 3.2GHz, Windows 10
- Desktop PC: AMD Ryzen 7 3700X 3.6GHz, Windows 10
- Laptop: Intel Core i5-4258U 2.4GHz, Kubuntu 18.04
- Laptop: Intel Core i7-8550U 1.8GHz, Kubuntu 18.04
- Laptop: Intel Pentium 4415Y 1.6GHz, Windows 10
- Laptop: AMD E-450 1.6GHz, Debian
- Mobile: Snapdragon 855 2.84GHz+1.78GHz, Android 9
- ShieldTV: nVIDIA Tegra X1 2GHz, Android 8
Here are the runtimes I used to run the benchmarks on:
- GCC 7+ (alternatively clang) with -std=c++14 and -O3 flags
- Lua 5.1+ and luajit
- NodeJS 8
- Godot 3.1
- Python 2.7
- termux for Android devices
I wanted to put minimal time into this so I didn’t try to run everything on every platforms, just what’s easily doable.
For each measured time, I ran the simplebench script/binary more than 10 times and took the shortest achieved time:
|RaspberryPi4/Cortex A72 1.6GHz||0.61s||4.3s||0.14s||0.061s||0.045s||0.136s||0.087s|
|Laptop/Core i5 4258U 2.4GHz||0.22s||1.2s||0.041s||0.018s||0.015s||0.045s||0.025s||0.637s|
|Laptop/AMD E-450 1.65GHz||1.17s||7.2s||0.210s||0.110s||0.079s||0.233s||0.152s|
|Mobile/Snapdragon 855 2.84GHz||0.36s||1.68s||0.013s||0.020s||0.758s|
|Laptop/Core i7 8550U 1.8GHz||0.18s||1.0s||0.037s||0.012s||0.011s||0.053s||0.018s|
|Laptop/Pentium 4415Y 1.6GHz||0.46s||3.65s||0.094s||0.020s||0.030s||1.411s|
|Desktop/AMD Ryzen 5 1600 3.2GHz||0.22s||1.38s||0.031s||0.008s||0.011s||0.011s||0.725s|
|Desktop/AMD Ryzen 7 3700X 3.6GHz||0.156s||1.26s||0.026s||0.007s||0.008s||0.014s||0.496s|
|ShieldTV/Cortex A57 2.01GHz||0.745s||4.73s||0.049s||0.012s||1.496s|
Mobile phone processors are catching up to laptop processors very quickly. I’ve read that Snapdragon 855 is similar in performance to a current gen core i3, and these results confirm that. It’s especially impressive that Snapdragon 855 is almost exactly matching a desktop PC Ryzen 5 1600 in Godot/GDScript and very close in Lua and Python.
CPython is so incredibly slower than everything else which comes as a no surprise, some of my old CPython/PyGame games struggled to hit 60 fps on laptop processors at the time without some sort of just-in-time compilation thrown in (back then I used Psyco)
Pypy runtime (which is a decendent of psyco, full jit compilation) is impressively quick in comparison to CPython, it almost matches mono actually, it’s a surprise pypy hasn’t become the dominant python runtime yet! I think it ought to be.
Lua and LuaJIT
As expected, lua is very fast for a fully interpreted language. I wish CPython was closer to that.
LuaJIT is damn impressive. Actually really close to native C++ performance! Which is insane.
The anomaly here is NodeJS which uses the excellently optimized Google V8 engine. Not only did it match but actually surpass native C++ performance on several platforms. I have no explanation other than blaming it on timer precision? but I’m not surprised its performance is that good as it not only runs 100% of the web, but a growing list of desktop/mobile applications like this very editor I’m using to write these words now.
It’s worth noting that the startup time when running
node simplebench.js is almost as slow as compiling the C++ version. This indicates some hardcore jit compilation taking place before actual execution of the script starts.
Godot’s GDScript is not as bad as I thought it’ll be, about 2x CPython. Alone it would make Godot a terrible solution for bigger games, but luckily Godot allows C++ modules to be used for critical bits and Mono/C# support is almost ready for prime time. I’m hoping at some point in the future they decide to reimplement GDScript to compile to Mono in the future.
blog comments powered by Disqus