Post a Comment
'Twas on Slashdot yesterday
http://linux.slashdot.org/linux/05/08/23/2032218.shtml?tid=174&tid=...
Well I will be going to a conference on parallel computing in a few weeks (cpa2005). There has been some interest from that community to explore if the cell architecture might be useful as a Transputer replacement (I think not).
Anyway a couple of fellas from IBM will be right in front of me talking about the communications support for concurrency.
So I can ask any questions I want if I want!!
Transputer guy
> Don't compilers take care of most of the complexity ?
Newer compilers have autovectorization which allows the compiler to convert floating point code to sse/altivec code. The problem is that cell has 8 vector units. This pretty much makes it difficult to utilise without writing code especially for them.
As for the features that give cell a lot of power (ability to chain the SPUs), I don't think compilers can generate code for this from your standard app.
Finally the main core is rumored to be cut down, thus less suitable for general purpose computations, then the 3.2ghz would appear. However even if this is just a rumor, it is a 3.2ghz PPC processor. Within 12 months I would expect AMD & Intel to have cpus which easily surpass this in power.
The big market for cell processors will be dedicated appliances, where the spus can replace existing DSPs, such as digital tvs, sound equipment and scientific equipment. Should there become a Military Spec version of the chip, then this could be possibly be used as a major chip in radar / sonar / ELINT / EWAR systems.
As a CEO of a high performance computing company involved in the oil and gas and other industries, I certainly welcome diversity. As is the case with a number of specific vertical industry, diversity *is* actually accepted (sometimes) at the expense of 'common the shelf'. Mercury's RACE architecture is a good (but expensive) example thereof. Why would you even bother including PowerPC controllers of stantard off the shelf PC's? Defense knows why. Oil and gas, on the other hand, don't want that.
Back to the point - the Cell:
Unfortunately, my sources at IBM Deep Computing claim that their tests of the Cell processors for typical HPC tasks proved the Cell processor an unlikely candidate. The Cell really didn't perform.
I am sure we will see highly specialized use of the Cell, but I take it as a big 'bump' when IBM themselves aren't very impressed.
Time will tell. Anyhow, good luck !
> to try and shake the image that it's only good for gaming.
IBM's technical specialist also tried to emphasise this at a recent Amiga fair in Sweden:
http://amigaworld.net/modules/features/index.php?op=r&cat_id=1&rev_...
IMO the kind of information spread around on the internet by some feels very similar to me as was the case with regard to classic Amigas during the 80s/90s. All kinds of people (including respected computer press) treated this innovative mulitmedia system as just a dedicated games machine..




"