Discussion The Playstation 3's use in supercomputers?

[Order 66]

Why was the PS3 used as in supercomputers, and how powerful truly was it?

 

[Eximo]

In basic terms it had a unique processor and support for Linux out of the box. Console hardware tends to be rather efficient at launch as well compared to other consumer CPUs, it was also pretty inexpensive all things considered. Included networking, storage, etc.

Similar projects were done with the PS2 as well.

Not sure what measurement you could use for performance outside the listed specs. It really would have depended on the application, whether they used the GPU, and so on.

 

[Order 66]

In basic terms it had a unique processor and support for Linux out of the box. Console hardware tends to be rather efficient at launch as well compared to other consumer CPUs, it was also pretty inexpensive all things considered. Included networking, storage, etc.

Similar projects were done with the PS2 as well.

Not sure what measurement you could use for performance outside the listed specs. It really would have depended on the application, whether they used the GPU, and so on.

Would it have been possible to run any standardized benchmark on it?

 

[Eximo]

Well the whole idea for supercomputer usage was clustered processing, so it would scale with the number of PS3.

Only one I know of that might work is LINPACK for a straight FLOPS benchmark.

PlayStation 3 Gravity Grid

PlayStation 3 Gravity Grid

web.uri.edu

"Our 16 PS3 Gravity Grid generates a total performance of 40 GFLOP/s (40 billion calculations per second). It should be noted that this benchmark was run in double-precision and because of the limited RAM on each PS3 we were only able to fit a matrix of size 10K on the entire cluster."

 

[Order 66]

Well the whole idea for supercomputer usage was clustered processing, so it would scale with the number of PS3.

Only one I know of that might work is LINPACK for a straight FLOPS benchmark.

PlayStation 3 Gravity Grid

PlayStation 3 Gravity Grid

web.uri.edu

"Our 16 PS3 Gravity Grid generates a total performance of 40 GFLOP/s (40 billion calculations per second). It should be noted that this benchmark was run in double-precision and because of the limited RAM on each PS3 we were only able to fit a matrix of size 10K on the entire cluster."

Are there any cross-platform benchmarks that would work on the PS3? I would like to test and compare my devices, if possible. by cross-platform, I mean includes support for ChromeOS, or can run in a browser.

 

[Eximo]

Linpack is available from the play store, so it should work on ChromeOS.

I believe Linpack is actually the CPU benchmark in several common benchmark tools, but it is available separately.

 

[USAFRet]

Why was the PS3 used as in supercomputers, and how powerful truly was it?

1,760 PS3's.
https://phys.org/news/2010-12-air-playstation-3s-supercomputer.html

"I know a guy..."

 

[kanewolf]

1,760 PS3's.
https://phys.org/news/2010-12-air-playstation-3s-supercomputer.html

"I know a guy..."

The key was they were cheap, and available. They could be networked and could run standard linux. That made them attractive for clusters.

 

[hotaru.hino]

Why was the PS3 used as in supercomputers, and how powerful truly was it?

The Cell processor's design could be thought of as an earlier version of how modern GPUs are designed: use a more general purpose processor to manage a bunch of simpler processors that do all the number crunching. I recall one of the use cases Sony said you could do with the Cell was assist the RSX GPU with graphics tasks. And indeed, some of them used the Cell to perform what would be done in compute shaders today. Computer shaders are often used today to work on screen space buffer effects like SSAO.

I'd argue the Cell was the progenitor of APUs.

 

[USAFRet]

The key was they were cheap, and available. They could be networked and could run standard linux. That made them attractive for clusters.

Exactly.

And apparently use a lot less power.

 

[Order 66]

The Cell processor's design could be thought of as an earlier version of how modern GPUs are designed: use a more general purpose processor to manage a bunch of simpler processors that do all the number crunching. I recall one of the use cases Sony said you could do with the Cell was assist the RSX GPU with graphics tasks. And indeed, some of them used the Cell to perform what would be done in compute shaders today. Computer shaders are often used today to work on screen space buffer effects like SSAO.

I'd argue the Cell was the progenitor of APUs.

Wasn’t the cell really difficult to develop for?

 

[hotaru.hino]

Exactly.

And apparently use a lot less power.

I wanted to see where the grain of truth in this is, but if we went by the TOP500 list in 2008, the IBM Roadrunner (which uses Cell) was the only one in this list to do something better than 0.25 TFLOPS per kW (it got ~0.44 TFLOPs per kW).

Cost wise, to get that much compute power in PS3s, you'd need around ~16,200 of them. And at $500 a pop for a base model, you're looking at $8.1 million. Roadrunner cost about $100 million. Talk about a bargain.

Wasn’t the cell really difficult to develop for?

It was only difficult in the sense that:

• It was a new paradigm for game programmers to develop on. I don't think anyone except maybe bleeding edge gurus like John Carmack or the guys at Naughty Dog knew how to do multithreading effectively.
• Sony's programming tools and documentation was crap

 

[COLGeek]

More background on the topic: https://en.wikipedia.org/wiki/PlayStation_3_cluster

Note the change that led to the demise (I had one of these. Dang shame when it happened).

 

[Order 66]

More background on the topic: https://en.wikipedia.org/wiki/PlayStation_3_cluster

Note the change that led to the demise (I had one of these. Dang shame when it happened).

Right, the removal of Linux. I remember now. Security concerns were cited as the reason, I believe.

 

[hotaru.hino]

On an aside of game development on the PS3, I found an excellent presentation from Guerilla Games: https://ubm-twvideo01.s3.amazonaws.com/o1/vault/gdc09/slides/GDC2009-vdLeeuw-KZ2SPUsCaseStudy.pdf

Especially note page 55. The SPUs are taking care of a lot of things that today would be shoved onto a compute shader. So basically what would be handled today by a GPU driver, the developers had to take care of themselves, and they had no trailblazers to go off of that could help them figure this out.

 

[Order 66]

Is there any way Cinebench could be run on one? I wonder if it would do well if it cinebench was coded for it since it is basically a GPU and the cell could be used to assist the GPU and potentially score higher. I'm not sure if this is even remotely possible, or how it would be done, just speculating.