------------------------------------------------------------------
Subject: Re: [TA] Amiga MCC CPU = MIPS64 20000 CPU ?
Date: Tue, 17 Aug 1999 15:48:37 -0400
From: Dave Haynie
Reply-To: teamamiga@thule.no
To: teamamiga@thule.no
On Tue, 17 Aug 1999 10:36:55 -0600, "Sami Cokar"
jammed all night, and by sunrise was overheard remarking:
> > Behalf Of Dave Haynie
> > make some basic assumptions. For one, unless you know those making the
> > decisions are stupid people, you have to assume they're smart, that
> > "they'll do what I would do". This is how I make predictions,
> Unless you have 'insider knowledge' or could be considered to be under NDA
> due to your conversations with Jim Collas, what would be your best educated
> guess on the CPU/partner that Amiga Inc will be teaming with?
Well, from my conversation with Jim Collas, what they've said since,
what they have done, etc. I would have to bet on Transmeta. I'm familiar with
some of the technology Transmeta is working on, from their patent
filings (anyone can read these, once granted, IBM has a nice patent server you
can find on the net). Here are some reasons:
1) Basically in our conversation, Jim said that he can't confirm
Transmeta (I hadn't asked yet), but then went on to describe
their planned system architecture, which essentially requires
Transmeta or something very much like it: it would be a total
boondoggle, as a desktop/home general purpose computer, running
on a traditional CPU. This includes running x86 binaries, running
2) They flashed "Sun" and "Transmeta" at the end of Collas'
speech at WOA. Clearly, Sun's involved, just based on their
admitted support of Java and Jini (whether that's "technology
provider" or "active partner" remains to be seen). It's unlikely
they would put "Transmeta" up there unless they were using it AND
Transmeta gave them permission for such a tidbit. This is also
very much in keeping with the aire of mystery Transmeta has been
milking all along (lots of noise about hiring Linus Torvolds,
but check out http://www.transmeta.com).
3) Amiga is working with Linus to some extent. In fact, given all
of the claims of "we'll be the top multimedia platform" that
Collas stuck to, that Amiga [the company] sticks to, they would
need some world-class Linux kernel people to fix the kernel even
for Windows-class multimedia. And there's not much time. On the
other hand, a partner with much of this work already done would
be very hard to resist. And perhaps the only possible reason one
would drop QNX for Linux and still state with confidence they're
on-track for Multimedia, in the AmigaOS/BeOS sense of the word.
Now, of course, the rumor mill is hot an heavy with "MIPS" and "MAJC"
(Sun's new "architecture for the next millennium", which they're
describing in full, for the first time, at Hot Chips today, at least one
would think so, they're definitely on the schedule).
MIPS rumors only started a few months ago, when Amiga posted a jobs
listing that included a position for a MIPS "porting engineer" or
some-such. There are two theories to explain this.
The first, and by far most likely, is simply looking elsewhere in the
system. It was long rumored the AmigaNG would have a "Multimedia Chip"
as a graphics, sound, and possibly even CPU engine. My guess early on was
something from Chromatic, makers of the VLIW-based MPACT line of MMCs,
in which Gateway held a 10-15% stake. Chromatic was dying fast, since
their chip just didn't quite fit into a PC (graphics way too slow, most
all-in-one PCs had system solutions for graphics and sound). ATi bought
up Chromatic last year; more recently, ATi was announced as the provider
of the graphics architecture for the AmigaNG. But it's not their new
128-bit PCI chip, but something else. My bet: whatever Chromatic was
working on.
It's also clear that MIPS and ATi are working together (MIPS, that is,
not SGI, which split from MIPS last year, taking all of the high end
stuff with them). MIPS provides the core, ATi the 3D experience. MIPS
gets a good 3D ISA (they're also covering it today at Hot Chips, but you
can download the specs from http://www.mips.com), ATi gets a general
purpose computing core to run in their 3D chips, ala Hombre. This won't
be fast enough to be a desktop CPU, but might fare well in game consoles
or STBs. And as a stand-alone graphics processor, it's excellent, and
given the right new instructions, it'll handle more of the 3D pipeline
than do the current 3D chips (nVidia, 3Dfx, etc). Again, just what we
had planned for Hombre-as-a-graphics-card.
The other theory (not mine, but I collect them), is that Transmeta is
indeed building a CPU as described, but their VLIW engine uses some MIPS
technology. If so, clearly, those with MIPS experience would be more
useful working with this chip than those without.
The MAJC idea is perhaps a bit sounder, since from what Sun describes, a
mature MAJC chip will be much faster than today's CPUs. Rather than run
instruction level parallelism like the superscalar RISCs or VLIW
processors today, they're building complete separate CPUs on the same
chip. This, coupled with pervasively threaded code (x86, no; Java,
maybe) would tend to make the machine many times faster than a conventional
CPU, but only if each MAJC processor were comparable to a PIII or some-such.
We'll have to see about this one. And there's no telling when real chips
will show up. Sun has nothing to lose talking about MAJC now, because
it's not taking away from an existing product -- they have nothing in
the home, embedded, or desktop markets. When that's the case, the talk
usually starts sooner, rather than later.
Multiple CPUs per chip is not new. TI has a thing called the 32080,
which sports four 32-bit DSPs and a RISC processor on a single chip. Not
cheap, but it works nicely on some problems (TI also makes a VLIW based DSP,
and that seems to be where they're concentrating for the future).
IBM is doing something similar now, though they're actually building a
"multithreaded" chip, not whole separate chips. What this means it that
each chip has multiple register sets (including PC), but will share MMU
and a pool of execution units between the different threads of execution.
--
Dave Haynie | V.P. Technology, Met@box AG | http://www.metabox.de
My opinions are my own, but sign the right NDAs, pay those
royalties on time and in cash, and they may be yours, too.
-------------------------------------------------------------
Subject: Re: [TA] Amiga MCC CPU = MIPS64 20000 CPU ?
Date: Wed, 18 Aug 1999 13:52:18 -0400
From: Dave Haynie
Reply-To: teamamiga@thule.no
To: teamamiga@thule.no
References: 1 , 2
On Wed, 18 Aug 1999 19:00:58 +0200, "g'o'tz ohnesorge"
jammed all night, and by sunrise was overheard remarking:
> Dave Haynie schrieb:
> > My bet: whatever Chromatic was working on.
> Maybe not, see below. But it's possible that ATi got the MIPS license cheap by
> trading some of the Chromatix stuff in for it, ending up as MIPS 3D ASE .. ?
Or yet again, maybe. The main problem with the original Chromatic (not
Chromatix) architecture, other than "low performance on things PC user
value", was the fact it needed all kinds of handholding by the host CPU.
So it's extremely reasonable that any new graphics thing they did would
get its own general purpose computing engine internally. This would both
offload the host system and allow stand-along game/STB consoles. A MIPS
core would be ideal for this, and that's likely the piece that ATi gets
in return for helping MIPS develop the 3D ISA.
In other words, Hombre, reinvented.
> > It's also clear that MIPS and ATi are working together (MIPS, that is,
> > not SGI, which split from MIPS last year, taking all of the high end
> > stuff with them).
> http://www.mips.com/pressReleases/081699A.htmlhas an announcement from MIPS
> about some extra 3D instructions to their existing SIMD unit.
As I said... this isn't new news, though it'll get more exposure, since
MIPS was doing a paper on it yesterday.
> The new
> instructions are meant for geometrie engine software, they'll do 10-25 million
> polygons per second on a 500MHz processor, needing an extra 0.5mm2 of silicon
> area. That's at least three times as many polygons as you see in today's
> fastest AGP 3D cards!
In throughput, sure. But that's usually bus limited: they simply can't
render that many polygons per second. Most of the 3D cards run in the
10's of billions of operations per second, and can handle their piece of
the 3D pipeline better than either the CPU feeding it, or their bus
architecture, worst-case, will permit.
The problem, primarily, has been the CPU's ability to supply floating
point polygons. Certainly something like this MIPS 3D set, or AltiVec
(which I'm betting is faster still), maybe even SSE, will improve this,
with existing cards of today, before long.
> Geometry calculation (creating the polygon data which is then sent to the 3D
> graphics chip for display) consumes about 70% in today's x86 CPUs in 3D games.
> The next generation gfx chips (those after ATi Rage128, nVidea RIVA TNT2, 3dfx
> VooDoo3, and some more) are all expected to have geometry engines on board, to
> reduce load on the CPU.
Yup. The CPU has been the bottleneck for quite some time. It still will
be, with these new systems with geometry engines, in many cases. But the
trend is clear -- more and more of the 3D pipeline is moving to the
graphics chips, since they can explot new (to the PC, most of this stuff
comes from high-end graphics systems) parallel and pipelined
architectures to work on 50-100 stages at once, something a CPU can't
dream about.
> Putting geometry engines on those chips would put Intel in trouble, as gamers
> wouldn't need any new processors above 300MHz any more; Windoze will support
> the new functionality from DirectX 7.x.
And it's already handled in OpenGL. But it won't hurt Intel that much,
I don't believe for a minute that games programmers won't find ways to
use up the whole CPU.
But this has been a recurring theme. Back when PCs didn't have 3D
engines, adding a faster CPU was the only way to go. Then 3D came out,
and you could clear get faster performance (with acceleration, which
wasn't always there) by adding a relatively low-cost 3D card. Then games
caught up with these, and the CPU was once again an issue. Then the pixel
set up engine went on-chip, and the graphics card was the issue. Once you
had the graphics engine improved, the CPU was again the issue. Now it's
going flop once again, for awhile.
> Since processors sold today are already
> double that fast, that frees an awful lot of extra power to make games more
> interesting.
Yup. Funny how the PC has become such a "game machine" that virtually all
new technology is driven by games these days :-)
> I assume that ATi will use the MIPS processor with those new instructions on
> their chips for this purpose; it could also run as a standalone chip for set
> top boxes then as a side effect.
That's what I've been saying for weeks. Originally, before the Linux
announcement, someone spied a request for MIPS support in an Amiga
want-ad, and threads abounded about how the mystery CPU might be MIPS
based. The ATi thing makes that much less likely. Gateway owns a big
chunk of Chromatic; Chromatic folds and gets sucked up by ATi, Amiga
announces they're using a new ATi chip that doesn't come from ATi's
normal Rage, etc. line. What else could that be but their further
development of the stuff they got from Chromatic? They need an embedded CPU
for this, IMHO, to get around the problems with the old Chromatic
designs. Lo and behold, ATi has a MIPS licence, and they're working
closely with MIPS on a number of things. MIPS isn't making graphics
engines, they're making CPUs, and in fact, they need to tap some else's
3D expertise. Good match, perfect sense, etc. But if this is the only CPU
in the AmigaNG, they're doomed, at least beyond the STB level.
--
Dave Haynie | V.P. Technology, Met@box AG | http://www.metabox.de
My opinions are my own, but sign the right NDAs, pay those royalties
on time and in cash, and they may be yours, too.
------------------------------------------------------------------
Subject: Re: [TA] Info on Amiga Clones?
Date: Mon, 16 Aug 1999 16:52:06 -0400
From: Dave Haynie
Reply-To: teamamiga@thule.no
To: teamamiga@thule.no
References: <002601bee816$63eaa840$93611c26@amiga1.earthlink.net>
<37B87266.F12CACD9@lh-computertechnik.de>
On Mon, 16 Aug 1999 22:19:51 +0200, "g'o'tz ohnesorge"
jammed all night, and by sunrise was
overheard remarking:
> Iwin Corporation DOES NOT use any of the copyrighted technology
> or patents by Amiga Inc. All Iwin computers compatible with Amiga computers
> ar based upon our own technology which - in fact - is compatible with Amiga
> computers.
Highly unlikely. It's virtually impossible to be Amiga compatible
without violating Amiga patents. Yes, you can get around the software patents by
sending people to get the AmigaOS elsewhere, but if the blitter blits
and copper cops, there's practically no chance that you're not violating the
blitter patent, and perhaps others. And, of course, it doesn't matter
even a tiny bit if this is done in hardware or software -- patents cover
the invention, not the specific implementation of the invention (and
yes, UAE certainly does violate the Amiga patents).
It seems rather odd these guys would be making such claims without,
apparently, understanding the first thing about patent law. Not that it
matters in practice -- I can't imagine Gateway going after someone for
making a new Classic Amiga clone. But if they don't understand this,
what else are they confused about?
--
Dave Haynie | V.P. Technology, Met@box AG | http://www.metabox.de
My opinions are my own, but sign the right NDAs, pay those
royalties on time and in cash, and they may be yours, too.