> the PS4 can’t easily be turned into a cog for a supercomputing machine. “There’s nothing novel about the PlayStation 4, it’s just a regular old PC,” Khanna says. “We weren’t really motivated to do anything with the PlayStation 4.”
The article doesn’t really explain in detail the difference that made the ps3 an attractive buy for making super computers over regular PCs. Can someone elaborate?
From Wikipedia about another, similar project (Folding@Home)
> At the time of its inception, its main streaming Cell processor delivered a 20 times speed increase over PCs for some calculations, processing power which could not be found on other systems such as the Xbox 360.[36][190] The PS3's high speed and efficiency introduced other opportunities for worthwhile optimizations according to Amdahl's law
> The PS3's uniform console environment made technical support easier and made Folding@home more user friendly.[36] The PS3 also had the ability to stream data quickly to its GPU, which was used for real-time atomic-level visualizing of the current protein dynamics
Ironically, this is the kind of massive research that got Sony into troubles because it costed a ton and was hard for developers to master (SEGA felt this hard a decade prior), hence the Nintendo U-turn and Sony/Microsoft 90% mainstream hardware follow-up.
The processor in the PS3 was very different. It was called Cell. It had 9 cores: 1 general-purpose processor and 8 stripped-down, hot-rod processors. They were tied together by a fast bus. I think the cores could even talk to cores in other boxes more efficiently than could be done with x86 processors. Making a supercomputer from PS3s was 10 times cheaper and would run 10 times cooler. --- https://www.datacenterdynamics.com/analysis/the-playstation-...
DISCLAIMER: I am not a chip expert. I threw together this summary after reading several articles.
As I recall, a big part of it was that Sony was subsidizing the powerful hardware, selling it slightly below market under the business model of making the money back by selling games to run on the systems (kind of like the printer/ink model)
This meant you could buy much more hardware for the same cost verse buying the parts individually.
The PS3's CPU was a single general-purpose core and 6 SIMD coprocessors.
It was crap for creating games (and a pain in the ass to develop for), but it was great for pure compute.
By comparison the 360 used a trio of a similar general-purpose core without the SIMD processors on the side[1] had significantly less pure compute power.
[1] but with better built-in SIMD: the PS3's general-purpose core had a single internal SIMD unit while each of the 360's had 2, the 360's SIMD units were also improved with games-relevant operations and 4x the number of SIMD registers
> The PS3's CPU was a single general-purpose core and 6 SIMD coprocessors
6 usable for games but 8 physically in the silicon, with one disabled and one reserved for the OS. So a dual thread PPU and 6 game usable SPEs for the advertised "9 thread" (sometimes just referred to as 9-core) CPU.
> 6 usable for games but 8 physically in the silicon, with one disabled and one reserved for the OS.
Indeed, I initially put in the disabled one (for yield) and one reserved (especially as IBM's server version had all 8 enabled) but I figured that was of low enough interest I could just list the actually usable ones.
Semiconductor pricing is largely determined by volume. The Cell processor had mediocre normal performance inline with other Power derivatives but really good vector performance for the target price (this was before GPU computing matured), so people with vector-amenable tasks could theoretically win by adopting a processor shipping by the million rather than using lower-volume chips targeted at traditional HPC uses. Since Cell was a Sony/IBM project, IBM was interested in reselling the same work to new customers and (IIRC) needed the volume to have their chip foundry business be healthy.
The problem was similar to Intel’s flop with IA-64: you could get a few amazing benchmarks but code which couldn’t be perfectly scheduled like that performed poorly, and that tuning was hard even if it could be done. The scientists I supported at the time were interested but the numbers made no sense: some code was inherently branch-heavy and couldn’t be tuned much and for the portions which could it was cheaper to buy more regular hardware using the money which would have gone to a high-end programmer, especially when you looked at the long-term maintenance cost of that heavily optimized code (everyone of that era remembered ripping out previous assembly optimizations which had become worse than C on newer CPU & compilers).
The article doesn’t really explain in detail the difference that made the ps3 an attractive buy for making super computers over regular PCs. Can someone elaborate?