On Mon, Sep 14, 2009 at 6:07 PM, Nicolas Boulay
<[email protected]> wrote:
>>> So they use a core beside the cpu ?
>>
>> No.  The only special hardware they have is texture decompression.
>
> It's a kind of load/Store unit on steroid ? It's a must to avoid
> stupid bits move.

The determined that this was a worth-while co-processor.  They wanted
to make the hardware as general as possible, only adding specialized
logic if it was a substantial improvement over software.  This turned
out to be the case here.

>>> I think that the challenge is not in the performance of the chip but
>>> how it's easy to use or not.
>>>
>>
>> I'm tempted to make an argument that pushes off complexity into the
>> compiler.  But we've seen the negative effects of that a thousand
>> times over (e.g. Itanium, Common Lisp).
>>
>
> itanium use 2 or 3 bundle of 3 instructions. It's a lot compare to the
> average ipc of 3 for a typical code. A LIW with 2 or 3 instructions
> should not have the same issue. It could be a typical RISC pipeline,
> with the fpu pipeline. The idea is to have the fpu used every cycle.

Itanium's trouble wasn't about the way instructions were packed.  It
was that the processor is in-order and therefore requires the compiler
to order the instructions optimally.  Cache misses, however, make
latencies nondeterministic, making it impossible for the compiler to
do a perfect job of scheduling instructions.  An OOO processor, which
schedules dynamically, will do a better job.  (Although it will
require more hardware.)

Also, Itanium partially relied on predication for its efficiency, but
predication is notoriously difficult to use well by a compiler.  After
all these years, they STILL don't have a compiler that can use
predication well.

> If you use massive SMT technics, aren't you afraid of having more
> muxes than fpu ?

The idea is to round-robin the instructions from different threads.
We don't need MUXes.  Just tags indicating which thread the
instruction is for.

-- 
Timothy Normand Miller
http://www.cse.ohio-state.edu/~millerti
Open Graphics Project
_______________________________________________
Open-graphics mailing list
[email protected]
http://lists.duskglow.com/mailman/listinfo/open-graphics
List service provided by Duskglow Consulting, LLC (www.duskglow.com)

Reply via email to