harmony-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Renaud BECHADE" <renaud.bech...@numerix.com>
Subject RE: some ideas
Date Fri, 03 Jun 2005 02:02:38 GMT

I guess this is a question of available means and talents for each part
 (CG, encryption, or even scientific computing & exotic DB, etc). Many
things can be done, conceptually, to get warp performances (intelligent use
of SIMD instructions & GPU for instance) but on the other hand there
probably needs to be some cute, not so arch dependend support from the VM
infrastructure to make it "just scalable" (ex: GPU[1] ops. rescheduling,
generic SIMD support, etc.) because we don't want to write so much specific
code for each architecture...

I admit this vision of the runtime is bit more expanded than the core J2SE,
yet these "extensions" are quite in heavy use today.

RB

[1] See http://www15.in.tum.de/Research/Publications/
 Or http://multires.caltech.edu/pubs/ (look for keyword GPU)
 Basically it has been demonstrated that today's GPU can perform a well as
today's CPU for some well chosen heavy computational ops (linear systems,
linear programming -> some particular DB ops, cryptography of course, etc.),
with a hope for way better perfs tomorrow. Please note too that the new
IBM/Sony Cell processor is somewhat in that trend. I've heard of guys who
mount machines with 5 or 10 GPU-cards to get real cheap mini-computers, too
(for image processing I think. Cf. the OpenVidia project if my memory is
good)

-----Original Message-----
From: Steven Gong [mailto:steven.gong@gmail.com] 
Sent: Thursday, June 02, 2005 11:49 PM
To: harmony-dev@incubator.apache.org
Subject: Re: some ideas

On 6/2/05, Sven de Marothy <sven@physto.se> wrote:
> 
> On Thu, 2005-06-02 at 10:42 +0900, Renaud BECHADE wrote:
> > I guess this is a schema that happens very often.
> > Lol
> 
> It does happen quite a bit though. I'll admit that there are plenty of
> parts in Classpath which are pretty poorly implemented. And there are
> parts of Sun's implementation which seem very good.


Is there any fundamental reason that Classpath can't surpass Sun's 
implementations? Sorry, I am just new in this area. Is it that we can't get 
enough test cases?


-- 
Best Regards
Steven Gong


Mime
View raw message