[wellylug] Sun confirms plans to open source Solaris / hardware junkies!

[Rob Giltrap]

Wood Brent wrote:

>>Having said that, the 'traditional' 64 bit systems (or at least SPARC) 
>>are about to have a major change to CMT design creating chips that have 
>>massive parallel processing capability orders of magnitude more powerful 
>>than current processor design. I can't tell you the numbers other than 
>>to say 'WOW', SPARC definately has a big future in the data centre.
>>    
>>
>
>Potentially, yes. But how much like Intel's hyperthreading & other
>optimisations which are in some contexts as much hype as they are genuine
>performance boosts? 
>  
>
Very little, diminishing returns are the story of current CPU design

>Intel is planning multi-core cpu's RSN, which are in principle not too
>dissimilar to those Sun is developing. So how much better will Sun's offering
>perform compared to other multi-core cpus? If it is more a whole new
>architecture, with better I/O, memory bandwidth, etc, it might have some value.
>A faster cpu on it's lonesome is pretty crippled.
>  
>
This is it, the current architeture does not scale well. Extra cores are 
only useful if you can get data to them. Sun's approach(now)  is very 
different from a philosophical point of view (i.e. A faster cpu on it's 
lonesome is pretty crippled!).

It's really hard to get a big company like Intel, Sun, IBM to 
fundamentally change the way it does things, and this architecture 
change is one example. The new processors designs were not originally 
designed within Sun. Several Sun engineers had some ideas but could not 
ge traction with the old guard so they left Sun and created a small 
company Afara Systems. They then designed the new processors without 
having to conform to any restrictions. Sun then bought the company and 
now it has a really 'elegant' processor design.

>Looking at NIWA's (somewhat dated now) Cray T3E as an example, which offers
>above average parallel processing ability :-), there are many analyses & models
>which do not efficiently parallelise, so run like a (very old, slow) dog. Hence
>the use of a grunty Sparc box as a numeric analysis server for some
>applications. Which for some work is no faster than a good desktop PC.
>  
>
You make a real good point, and now that the industry has rationalised 
down to only a few instruction sets, there is a change in that you are 
seeing different processors for different workloads with the same ISA 
(e.g. Pentium M, Xeon, P4 & Celeron). The new CMT processors are 
designed for massive multi threading in a networked environment. the 
Opterons are designed for really good single thread performance with 
little to moderate parrallelism. No one CPU can be all things to all 
people these days. I'm actually pretty keen to see what is happening to 
IBM's Cell technology... maybe Carl can shed some light?