Cet article existe également en français.
Introduction
I bought a shuttle X27D (330€) that has got a 64 bit compatible, 2 core hyper-threaded (4 core with Linux, 512K cache for each one), ultra-low voltage (8W TDP) Atom 330, the one used in the SGI molecule supercomputer concept (10,000 cores in a rack), sadly, the North bridge has a greater TDP (around 23W) than the CPU. I choose the shuttle, because of the DVI connector. Every Atom 330 motherboards has only VGA and there are problem of detection in VGA between computers and my LCD screen.
As there wasn’t any 64 bit mode Atom test, I do one using Ubuntu. Ubuntu-8.10 liveCD crash at boot on atom 330, but Ubuntu 9.04 alpha works perfectly.
I used two tests, Hardinfo, mono-thread test included in Ubuntu and Blender well known bench using multithreading (4 threads for the 4 cores was used for this test).
Test Hardinfo MONOCORE
* Zlib compression (integers). greater is better
* Fibonacci suite (integers). lower is better
MD5 sum, used for file integrity (integers). greater is better
SHA1 encryption (integers). greater is better
Blowfish encryption (integers). littler is better
There are great difference between 32 bits and other in this test.
3D raytracing (floats). littler is better
The 64 bit mode is better than 32 bit mode for floats
The 32 bit mode is better than 64 bit mode for integers
As all these test are mono-threaded, performance of the CPU should be far better if multi-threaded, as this is the case on lot of application on GNU/Linux today.
Test multicore with Blender rendering
This test uses eofw.org benchmark/
This test uses default Ubuntu or getdeb Blender, these version are not optimised for sse and so…
The atom330 has got : sse sse2 and ssse3 but no sse3 (could need patch for Hackintosh).
As Blender is perfectly multithreaded, it will use multicore. On Turion CPU, Getdeb version of Blender is used. Don’t know if Ubuntu Jaunty version is more optimized ???
Ubuntu 2.46 32bits version is faster than official Blender.org one.
I can see on the window displaying the rendering that the 1024KB cache of Turion help him gain time by compute bigger blocks of the picture. The 512KB caches of Atom330 reduce the sizes of the block and the FPU performances of the CPU is lower than the Turion. The memory management that is externalized on Atom is also a bad point for him. Anyway Atom works really better in 64 bits mode than in 32 bits modes for floating numbers computation. Next generation of Atom including memory management included in CPU this year will probably has better results.