My thoughts on Playstation 3(PS 3) Cell Processor and benchmarking results…
With the recent purchase of twenty one Playstation 3 console by my school, we were allowed to run programs (via linux) to make use of the massive computational potentials of it. Naturally, we wanted to find if it lived to the 204 billion floating point operations (Gigaflop) IBM claims it is capable of. The following charts show the results of various tests run on the PS3 by a third party and our own matrix computation tests in which matrices of sizes ranging from 100 to 100,000 was calculated:
From the chart we see the PS3 will excel in computations that require little memory use but high computation. Applications requiring high network data transfer rates might seek alternative solutions elsewhere also.
While the Sony Playstation 3 might not be known for its good line up of games, its introduction has led to the possibility of end users to own what might very well be a cheap supercomputer. It has been hailed as a true next generation multicore computer for its wonderful integration of Eight(8) Synergistic Processor Element(SPE) and Power PC(PPC) processors. Users can now exploit its massive power to write programs that are highly parallel achieving this for a fraction of the usual cost. The PS3 might seem like the ultimate machine when its potentials in the field of supercomputing is considered but it is not without its draw backs. The most notable of these being its memory limitations; it comes with 256MB of system memory and only 256KB of cache on each of its SPE cores. If these limitations are overcome, then its gigabit ethernet port might deter those seeking to string together a cluster of PS3s. A detailed report on the limitations of the PS 3 is available here. IBM has a good list of tutorials on programming the PS3. My source code for the PS3 benchmarking is available here
In a future post, I will compare the performance of other potential supercomputing machines like the Nvidia 8800GT and Tesla GPGPU which both use CUDA to achieve parallel programming.
Hello i am a student of sciencie computers from Carabobo university in Venezuela (www.uc.edu.ve), and i’m working in a cluster of ps3, i’m interested in benckmarking the performance of this cluster. I enter here by google looking for something to do the benchmarking…I want to know if is posible, that you explain me, how do you do the tests.
Hi, there are several ways this can be done. I believe there are a few packages out there that contain benchmarking software for your linux cluster irrespective of the underneath hardware. If you want to write your own benchmarking software, I would recommend using matrix computation, you try to compute a very large matrix utilizing all the cores and viola! you have your benchmark.