[Arm-netbook] Should we support libre.computer's efforts at promoting lima?

[Luke Kenneth Casson Leighton]

---
crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68

On Thu, Oct 18, 2018 at 12:44 AM David Niklas <doark at mail.com> wrote:
>
> On Wed, 17 Oct 2018 16:30:17 +0100
> Luke Kenneth Casson Leighton <lkcl at lkcl.net> wrote:
> > ---
> > crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68
> >
> > On Wed, Oct 17, 2018 at 2:58 PM David Niklas <doark at mail.com> wrote:
> > >
> > > On Wed, 17 Oct 2018 08:43:29 +0100
> > > Luke Kenneth Casson Leighton <lkcl at lkcl.net> wrote:
> > >
> > > > USD $1m would fund the Libre RISC-V 3D GPU effort through to 100%
> > > > completion including quite likely actual test silicon in a smaller
> > > > geometry like 65 / 55nm (700mhz or so, which for a GPU would be damn
> > > > good).
> > > >
> > > > l.
> > >
> > > Wait a sec! Last year you said:
> > >
> > > On Mon, 8 May 2017 16:43:07 +0100
> > > Luke Kenneth Casson Leighton <lkcl at lkcl.net> wrote:
> > > > On Mon, May 8, 2017 at 4:27 PM,  <ronwirring at safe-mail.net> wrote:
> > > > > Best knowledge is, that new intel and amd processors cannot be
> > > > > reverse engineered. What in regard of the latest mali gpus?
> > > > > If you have the money, they can be reverse engineered?
> > > >
> > > >  yes.  about $150k would do it.   but the question is, really: what
> > > > would happen if you did?  and, what else could you do with the same
> > > > money?
> > > >
> > > >  well, with the same money it would be possible to make our own
> > > > libre processor, with enough extensions to be able to do 3D graphics
> > > > *without* paying anyone a cent.  any company tries to claim patent
> > > > royalties, all that happens is a search is made on their "claims",
> > > > for anything similar that has prior art.
> > > >
> > > >  if it's another company, guess what?  we notify that other company
> > > > and watch the fireworks...
> > > >
> > > > l.
> > >
> > > 1M - 150k = 750K!!!
> > > Therefore you're tripling your price estimate on us!
> >
> >  it depends on what you take into account.  if someone else pays the
> > NREs to the foundry, i.e. a university agrees to collaborate and is
> > offered access to a foundry for either free or at reduced rates,
> > $250k-$500k comes off that amount, straight away.
>
> My statement was based on: 1M - (3 x 150K == 450K [for the HDL]) = 550K
> for the silicon, and that's huge!

 relatively speaking it's tiny.

> What is the cost here?

 it was a throwaway comment, david, written in about 30 seconds, with
about 5 seconds thought.  the focus was more "a much higher bang per
buck can be achieved with that kind of money" than "here is a detailed
statement of work".

 it would take me a *lot* more time to specifically answer that very
general question without specific information.

> Silicon as an
> element is inexpensive and even 32nm foundries should have their HW paid
> for by now. It's been what? 8 years since 32nm and the half node, 28nm
> started? And the big buyers (AMD, Intel, Nvidia) were on it for 3+ years.

 28nm mask charges - the lithographic stencils onto glass - for 28nm
are i think USD $2m.  it's an exponential curve.  45nm is $1m.  it
drops to around $250,000 for 180nm.

> It's not like an OSH project can purchase a license for closed source
> HW something-or-another libraries,

 the industry term is "HDL" - i think it's just "hardware definition
language" or something.

> which I've read as being the main
> expense from anandtech (No, I'm not in the field).

 so that's why i spent several months tracking down libre equivalents,
and/or tracking down people in open hardware willing to *create*
libre-licensed HDL

> >  if only wanting to reach a target of say producing the design files
> > (HDL), that's a different target entirely as well, and if someone else
> > can take over at that point (and make the actual money selling
> > product), then that's one way to ensure that the goal's reached [but
> > not guarantee financial benefit from it].
>
> I assumed at the time, you were talking HDL and with an FPGA for
> hardware backing with basic OpenGL, OpenCL, and 2D driver. No video
> encode/decode HW or Vulkan driver. Not mass produced, just a few
> prototypes. At about the same speed as a good Intel GPU, say a high end
> HD630 series. If you have a more concrete or different idea I'd be glad
> to hear it.
>
> > so it was a throwaway comment written in about 30 seconds with no
> > strict analysis done.
>
> Really? You sounded so serious! Contrary to popular belief, sound does
> travel across the Internet, it's just that no body has managed to tie the
> Internet into the speakers yet. ;-)

 :)

> > last time i collaborated with an associate to
> > come up with a proper figure it was around USD $5m (for a client) and
> > it took us around five days to put that together. however that
> > particular deal excluded certain resources and had specific
> > requirements.
>
> :(
> I was aiming for 2M minimum. And that is BIG to me.

 if aiming for 28nm (without additional investors) that would barely
cover the production mask charges.  hence why i am interested to work
with IIT Madras.

 if requiring for example a modern DDR4 memory controller, over 50% of
that budget would be taken up.  hence why i am interested in HyperRAM
(upcoming JEDEC xSPI).

 if not using a back-end team such as the people that IIT Madras have
access to, that USD $2m would be entirely eaten by proprietary layout
tool licensing from Mentor Graphics and the engineers who would need
to be hired to do the work.  hence why i am interested in Magic,
alliance2 / coriolis, and libresilicon, all of which are developing
open ASIC layout tools.

 there is a *lot* i simply have not had time to talk about, here, david.

> > > Seriously, when I get BIG money (which will be a while),
> >
> >  cool
>
> Let me be frank, luke, I am under no delusions. No lottery. No gambling.
> No get rich quick schemes. But it will take time, as in ~5 years.
> Therefore, be excited about the OSH, not the money which may or may never
> be there.
>
> I don't always agree with you, but what you write I do pay attention to.
> And I have a terrible tendency of believing you. :)

 :)  always never do that: i am first and foremost an ethical person,
who goes, "that's not good, let's fix that", has absolutely *no idea*
how to go about "fixing that", and persistently chips away (often
randomly) regardless of complete lack of knowledge and expertise,
until success.

 it means that i really *really* need feedback and people to
double-check and triple-check.

> I had assumed that
> some company would opensource or support the OSS Mali drivers,
> eventually. Google has android running off of Mali, Samsung and LG have
> Smart TVs running off of Mali, the list goes on...
> But when you wrote about creating an OSH GPU yourself I realized that
> is spite of every Mali customer, Google et. al. would not support
> opensourcing anything if they could help it.

 YA THINK???

> Chromium's binary blob
> problems are even further proof.
> This became painfully clear when RISC-V came out. I expected *at least*
> the cache to be OSH... :CRY:
>
> I decided that if I could, it was up to me to pull in the funds for OSH,
> and "I think I can"(TM).

 awesome.  well, as usual i have a parallel set of tracks being
investigated, as part of a wider strategy of deals and collaborations
with various people, just bear in mind that your help would be part of
a much larger deal, ok?

> > > I'm planning on calling you on this one
> >
> >  excellent.  well, when that happens, let's talk (on or off list), you
> > let me know what you can raise and i'll put the time in - which will
> > be several days not a few moments - to work out what can be done, and
> > what side-deals will be needed, if necessary.
> >
> > l.
> >
>
> I will tell you, but keep in mind that there are other OSH products to be
> developed that I have my eyes on.
>
> Finally, I'm uncertain that I want to know, but what is a "side deal"?

 ok, so let's say we make a chip.  it's successfully made, it's all
done, and it works, right?

 and then we go, "ok, who wanna buy?"

 and... total absolute silence.

 egg.

 on.

 face.

 so the way to deal with that is: find **MULTIPLE** potential
customers, take their requirements into account, and make damn sure
that the SoC is capable of fulfilling THEIR needs times as many of
such potential customers as can possibly be found.

 and if even 30% of them come through and order 100,000, that's enough
to actually get a serious order with a foundry.

 now, some of those people might be willing to actually invest
up-front, to get the SoC made.

 that's what i mean by "side deals".  deals "on the side" that result
in everybody succeeding in getting what they want.

l.