[Arm-netbook] FSF-Endorseable Processor

[luke.leighton]

On Tue, Dec 4, 2012 at 12:36 PM, Dr. David Alan Gilbert
<dave at treblig.org> wrote:
> * luke.leighton (luke.leighton at gmail.com) wrote:
>> On Mon, Dec 3, 2012 at 11:45 PM, Dr. David Alan Gilbert <dave at treblig.org>wrote:
>>
>> yes.  process shrink and a modest increase in clock rate from 600mhz (65nm)
>> to 800mhz (40nm).
>
> Hmm I guess that's low enough not to be a push.

 love to go to 28nm - bit risky though.

>> the processor *is* the GPU, that's what's beautiful about it.  to put that
>> another way: the CPU has been extended with instructions that massively
>> speed up the kinds of operations that a GPU does.
>
> Hmm - I'll stay my normal pessimistic self on that; got an instruction set
> manual to show ?

 *lol* not yet.  have been asking for the compiler+kernel+android
source code as a priority over that.  they grok the "open source"
thing, but hadn't heard of "zhiyou", all got a bit confusing but it's
sorted now: just a misunderstanding.

they've told me they're committed to being open right from the start
[unlike many companies i've spoken to].  i'm happy to believe them
until proven otherwise, and even then am happy to give them good
reasons to steer in the desired direction.

>> performance at 600mhz is comparable to many of the embedded GPUs running at
>> 300mhz, so power consumption will, clearly, be a bit higher, but that's ok:
>> it's not running at 1ghz or 1.5ghz
>
> Hmm have they got any benchmarks for real apps - things like page
> scroll times, javascript benchmarks etc?

 give it a minute!!  they only got the EVB android demo done a few
months ago! :)

> The thing is that your A10 has a faster clocked core *and* the GPU;
> are you sure what you're suggesting isn't possible using a standard ARM cored
> SOC, and a dumb CRTC ?

 yes.

 ignoring the performance of GPUs, they've implemented H.264 (ok,
compiled up the standard libx264 and ffmpeg etc. etc.) and shown that
by using only 6 of the 8 hardware threads they can do 720p @ 30fps...
in 600mhz.  look up how well a 1.5ghz ARM Cortex A9 does: you'll find
it's only 15fps.

 once you bring in the GPU however, any hardware-dedicated GPU is
going to win on power-performance numbers alone *but* not on software
complexity.

>> > If they need to develop anything new then the timescale will be much
>> > higher.
>> >
>>
>> there are absolutely no plans to develop anything new.  the only minor
>> change needed would be to the 2nd level cache if doing an 8-core SMP chip.
>
> What have they got so far?

 dual-core SMP 65nm, 600mhz.  with not enough modern interfaces to
make it saleable except for... say... as a truly open alternative to
the rbpi.  the current chip would do well there... *if* competitively
priced.

l.