Re: [f-cpu] x86-64 long article

[nico <nicolas.boulay@ifrance.com>]

On Sat, 25 Jan 2003 10:30:23 +0100
Yann Guidon <whygee@f-cpu.org> wrote:

> hi !
> 
> nico wrote:
> 
> >On Fri, 24 Jan 2003 22:00:56 +0100 (CET)
> >devik <devik@cdi.cz> wrote:
> >
> >>>:) If you want more than 8 fcpu in a single machine use multicore
> >>>chip ! :)
> >>>      
> >>>
> >>Hmm .. Their Opteron or NUMA in general is good because
> >>of independent memory access.
> >>With multicore and limited pin count what to do ?
> >>
> >
> >Share the L2/ L3 caches. It's a whole new way of doing optimising
> >(tight multithreading) but it could work nice.
> >  
> >
> 
> Alright, but then, when the core needs to communicate, all the pins' 
> bandwidth will be saturated.
> Look at the Opteron's 1000 pins and ask yourself how much this will
> cost ....

And half are Vcc/Gnd pin... (1A/pin max i beleive)

> It's the OLD problem of a system's "cut" and the non linear ratio 
> between the gate count
> and the pin count .... this is my major problem.

yep.

> I believe that this will put a limit to 4 FC0s on a single chip.

Where come this number ??

> Looking at other numers, i think that half of the die's surface will 
> contain L2 and L3 and it's
> good, despite the high fabrication cost. But adding too many cores
> will suffocate the whole chip.

What do you comparre ?? What is better : 2 cpus with it's own memory
bank (NUMA as the poteron) or 2 cpu on the same die (thight SMP)?

The first system will be much much costly than the second. Because you
need to split the memory bank and you have 2 cpu on your main board
(even if the second cpu is bigger, the cost of the complete system
will be much higher).

Which one is the fastest system ? You try to explain that the second
solution will be the fastest because you have the double of memory
bandwith. But it will depend on the load. 

In the second system, you could have more cache so a single thread could
run faster. In a NUMA system, the communication between node are a
nightmarre and could kill your system.

So it's not evident at all for the same amount of money which one is the
best. If this amount is quite low, mono cpu design is much more
interresting. After it depend if you can't double the bandwith of the
single die (using 256 bits data path instead of 128).

> On top of that, a lot of software will need major rewrites to include 
> finer multithreading.

That's true. But thight SMP or SMT (P4 3.06 Ghz for example) have the
same problem. And maybe we could find compiler that could handle 2 cpu
at the same time. If the memory are shared is much easier to do. There
is no communication penalty.

> 
> Furthermore, even if we run at only 100MHz on an old technology, this 
> will be an incredible
> success for freedom, for the "alternate computing" world and for us in
> general. Extreme
> performance will come naturally later, we only have to make one core 
> work and others
> will want to join. So let's concentrate on the critical things, as we 
> will be able
> to play with top technology later :-)
> 

yep.

nicO

> bon week-end,
> 
> >>devik
> >>    
> >>
> YG
> 
> *************************************************************
> To unsubscribe, send an e-mail to majordomo@seul.org with
> unsubscribe f-cpu       in the body. http://f-cpu.seul.org/
> 
*************************************************************
To unsubscribe, send an e-mail to majordomo@seul.org with
unsubscribe f-cpu       in the body. http://f-cpu.seul.org/