[erlang-questions] Tilera 64-core chip - ...

G Bulmer gbulmer@REDACTED
Thu Sep 6 23:56:18 CEST 2007


On 6 Sep 2007, at 16:46, Thomas Lindgren wrote:
> --- G Bulmer <gbulmer@REDACTED> wrote:
>> IMHO, getting a book as good as Joe Armstrong's will
>> take significant
>> time (which is precious right now for Tilera); it'd
>> almost be worth
>> supporting an Erlang port just to leverage Joe's
>> book and the new
>> interest in Erlang stimulated by it. Further, if
>> Tilera can do it
>> right, they could cover there costs. Intel have
>> benefits to gain from
>> a good Erlang port too, and HiPE may even work on
>> the i80-core!
>> What's not to like?
>
> I would strongly suspect the chip is intended for
> performance-conscious embedded C/C++ applications
> (like so many other similar chips).
Yes, I believe that is how they are describing it too.

On the other hand, 80% of the performance is in 20% of the code (or  
96% in 4% :-), so using Erlang for the 80% of the code, and 20% in  
their own C-ish language may be fine. With this approach, I get some  
of Erlang's benefits for those less performance-critical parts. My  
perception is many of Erlangs benefits are expensive to write for one- 
off solutions. For example, I might write e.g. codecs in Tilera-C,  
but do all of the server monitoring, management, availability, etc in  
Erlang.

> Surely Tilera's
> main marketing and sales effort has to be in that
> direction, and turning from that mainstream to
> emphasizing Erlang would then seem a pretty risky
> move. What if the customers shrug and turn elsewhere?

I', not advocating Tilera spend much effort on it. I am a believer in  
Moores "Crossing the Chasm"; I don't expect them to distract themselves.

I am suggesting a small investment on a 'spread bet'. A small amount  
of support for self-motivated external communities, with no hard  
deadlines.

Specifically, I am suggesting Tilera support a few external open  
source groups working with different, but applicable technologies,  
and do it at a relatively low cost. The goal isn't main-stream,  
business critical outcomes. It is growing a skill base, getting more  
brain-power applied to their products, and hence improve their near  
term appeal, and broaden their long term market base. Even Intel 80- 
core may benefit from this too.
>
> (I'm not convinced this view is entirely correct, but
> some strong examples are probably needed to start
> addressing that market. As a startup, you can't vary
> too many parameters at the same time. Better let other
> people take most of the bullets :-)
I would agree with you unequivocally *except* Tilera don't have a  
clear field. They are playing against established companies like  
Intel, Sun, IBM, NVIDIA, AMD/ATI, MIPS licensees, ... who already  
have commercial and technology e-co systems. After all, we have to  
agree x86 was not the best CPU for much of the 90's, but Intel puled  
through to their dominant position.
I feel that working with a few, well chosen, Open Source communities  
makes sense for many-core chip producers, while there is a window of  
opportunity.
I think I'm suggesting initial steps (documentation, cheap SDK's)  
that they likely need to go through anyway.
>
> Regarding the advancement of Erlang as a development
> environment for these chips, I do think it would be
> interesting to see the people in this thread getting
> their PCI cards and start hacking away. But it's
> probably best seen as cutting-edge R&D at this point.
Yes. That's a great way to look at it.
They won't get full attention or control, but they avoid the major  
problem of building and retaining talented R&D groups; Hackers who  
take part are 'self selecting'.
I could justify a few hundred $'s for a few months interesting  
hacking (just on the basis of not eating out, or going down the pub  
as often :-)
Success may be very useful to the smarter many-core companies!

GB




More information about the erlang-questions mailing list