[Arm-netbook] Mali at fosdem

[Gordan Bobic]

On 02/10/2012 11:03 PM, lkcl luke wrote:
> On Fri, Feb 10, 2012 at 10:43 PM, Alec Ari<neotheuser at ymail.com>  wrote:
>>> Message: 7
>>> Date: Fri, 10 Feb 2012 13:04:26 +0000
>>> From: lkcl luke<luke.leighton at gmail.com>
>>> Subject: Re: [Arm-netbook] Mali at fosdem
>>> To: Linux on small ARM machines<arm-netbook at lists.phcomp.co.uk>
>>> Message-ID:
>>
>>> On Fri, Feb 10, 2012 at 8:39 AM, Alejandro Mery<amery at geeks.cl>
>>> wrote:
>>>
>>>> code released! https://gitorious.org/lima/lima/trees/master
>>>
>>>   suuupeeeeerrrrb.
>>>
>>
>> I read that the ultimate goal for Lima is to work and plug into Gallium. I'm excited to see the code published, but out of plain curiosity, why wasn't the MALI driver written with Mesa/Gallium in mind at the start?
>
>   i've seen a brief explanation on #lima - the architecture of gallium
> simply doesn't comprehend shaders, specifically shader compilers
> (whatever they are).  thus it just... doesn't fit, and therefore the
> simplest thing to do initially is bypass the whole lot, and come back
> to gallium later.
>
>   btw although i understood what you said about shaders being "mini
> ALUs/CPUs", gordan, i still don't grok what they're actually _for_.
> why do you need a group of parallel "shader" processors in the first
> place (as opposed to a general-purpose SMP CPU with a stonking-great
> DSP up its arse)?  what the hell are the damn things doing? :)

Well, your view was shared by Intel when they were cooking up Larrabee. 
The problem is that a general purpose CPU has to be able to do a lot of 
other stuff that affects the power budget and thus affects how many 
cores you can have on a die. The GeForce 5 series has something like 512 
shaders, and the latest Radeons several times more (but they are 
specialized rather than unified, so each one is simpler and smaller). 
Intel gave up because within the same power budget they could get 
somewhere around 64-128 cores (IIRC based on Pentium 1 with SSE added) 
in a package while staying within a vaguely sane power budget, and since 
with that few cores they couldn't compete with the throughput of GeForce 
and Radeon, they shelved the project.

You get better performance/watt out of a GPU-style architecture with 
lots of simple specialized cores, but to do so you have to sacrifice 
features on those cores that a general purpose CPU requires.

Gordan