Very Computer

When running software that doesn't use the new instructions (i.e. Linux),
the K6-2 has nearly identical performance to a K6 running at the same
clock speed. I've used both a K6-300 and a K6-2-300 (both at 3 x 100)
and I couldn't tell them apart.

The performance is very nice. A full kernel compile on my old P166
took about 8 minutes. On the K6, it took just over 3 (3:18, to be exact).

In my limited experience, the K6 overclocks better (maybe something to do
with fewer transistors), if that floats your boat.

If you play Quake 2, paying only an extra $30 or so for the K6-2, for a
potentially huge performance boost is hard to resist. Otherwise, you're
probably best off with a K6.

-Greg

Even without the 3d stuff my K6-2 300 overall (cpu,fpu & ram speed),
is within 4% of it's PII 300 counterpart. With slight  over clocking,
it's faster for ~1/2 the price. The only kicker is the relatively
expensive PC100 ram. Under linux, the K6-2 300 runs >600 bogomips.

Quote:>(I'll prolly be plugging it into a super-7 mobo with 32Mb RAM min)

>Cheers now,
>Chris

http://members.home.net/bigrex/

Do you know if there are any projects underway to support the K6-2 3D
instruction set?  Is AMD going to work with Linux developers to ensure
that we can support their chip with open-source kernel support?

I am going to be uprading from a Pentium 100MHz this week and want to
know whether I should bother paying the extra for the K6-2 over the
standard K6.

Quote:> If you play Quake 2, paying only an extra $30 or so for the K6-2, for a
> potentially huge performance boost is hard to resist. Otherwise, you're
> probably best off with a K6.

True, the game-specific support sounds like it could be nice but I'd
really like to see it supported in the kernel or at least through
XFree86.  Sorry, I'm no coder otherwise I'd stop asking and start
doing.  :-/

K6-2 also has a dual-issue MMX unit, whereas the original K6 has only one.
This places it at parity with Pentium and Pentium II.

Of course there's no MMX in the Linux kernel either...

--
Richard Krehbiel, Kastle Systems, Arlington VA USA

Check out <URL:http://www.slashdot.org>.  A few days ago, there was
mention of some optimizations to Mesa, and a request for people
wanting to help with K6-2 optimizations.

Quote:> Is AMD going to work with Linux developers to ensure that we can
> support their chip with open-source kernel support?

Their web site has a bunch of technical documents which anybody can
download.  In PDF, true, but gv and gs can grok that.  We're not
talking about NDAs or anything.

Quote:> I am going to be uprading from a Pentium 100MHz this week and want
> to know whether I should bother paying the extra for the K6-2 over
> the standard K6.

I'm in a similar situation: I've got a P120 and a motherboard which
can't take dual-voltage chips.  I think even the potential of being
able to use 3DNow is worth the relatively small extra.  Tie that with
the 100MHz FSB, and it looks definitely worth the extra cost.
(Doubtless you can run a K6 at 100MHz, too, but the K6-2 is supported
at that.)

Quote:> True, the game-specific support sounds like it could be nice but I'd
> really like to see it supported in the kernel or at least through
> XFree86.

3DNow is about single-precision floating point, as far as I can see.
It's an application thing, not something that the kernel or XFree86
could provide.  It's something that would be an obvious fit in Mesa.

(I'm less sure about one instruction: one of the AMD docs mentions
some kind of fast way of saving/restoring the floating point state.
Assuming I'm not completely misreading it, that would fit nicely into
the kernel's context-switching code, when a process has used the
floating-point unit.  I suspect I've misread it, though---surely
they'd speed up the existing instruction rather than provide a special
new fast one.)

Stupid question #2: what is mesa?  Be gentle as I am a recent convert
from the NT camp.  :-)

Quote:> > Is AMD going to work with Linux developers to ensure that we can
> > support their chip with open-source kernel support?

> Their web site has a bunch of technical documents which anybody can
> download.  In PDF, true, but gv and gs can grok that.  We're not
> talking about NDAs or anything.

Excellent.

Quote:> > I am going to be uprading from a Pentium 100MHz this week and want
> > to know whether I should bother paying the extra for the K6-2 over
> > the standard K6.

> I'm in a similar situation: I've got a P120 and a motherboard which
> can't take dual-voltage chips.  I think even the potential of being
> able to use 3DNow is worth the relatively small extra.  Tie that with
> the 100MHz FSB, and it looks definitely worth the extra cost.
> (Doubtless you can run a K6 at 100MHz, too, but the K6-2 is supported
> at that.)

Well I am also getting a new motherboard so the dual-voltage isn't
going to be a problem.  One of my other machines is a Tyan dual
processor motherboard and it would be really nice if I could throw a
pair of K6's on that but I know I am stuck with Intel here.  Right now
it's only got a single 120MHz CPU on that particular box but with 80MB
RAM it chugs along nicely.  Still, I've gotta play around with SMP.
It seems to work well with some types of services on NT (I've used SMP
at work) but I'd like to see how Linux handles it.

Quote:> 3DNow is about single-precision floating point, as far as I can see.
> It's an application thing, not something that the kernel or XFree86
> could provide.  It's something that would be an obvious fit in Mesa.

Could this potentially help with non-graphical applications as well,
such as crypto applications?

Quote:> (I'm less sure about one instruction: one of the AMD docs mentions
> some kind of fast way of saving/restoring the floating point state.
> Assuming I'm not completely misreading it, that would fit nicely into
> the kernel's context-switching code, when a process has used the
> floating-point unit.  I suspect I've misread it, though---surely
> they'd speed up the existing instruction rather than provide a special
> new fast one.)

Can anyone else comment on this?

Oops.  That should be <URL:http://slashdot.org>, sorry.  Mesa is an
OpenGL(TM) type library.  (So it's an API for 3D things.  It's what
Quake2 uses, for example.  It's one of the options, anyway.)

Quote:> > 3DNow is about single-precision floating point, as far as I can see.
> > It's an application thing, not something that the kernel or XFree86
> > could provide.  It's something that would be an obvious fit in Mesa.

> Could this potentially help with non-graphical applications as well,
> such as crypto applications?

I wouldn't rule it out.  I suspect the population of K6-2 owners is
going to be sufficiently small that nobody will bother coding it, but
I could be wrong.

Quote:> > (I'm less sure about one instruction: one of the AMD docs mentions
> > some kind of fast way of saving/restoring the floating point state.
> > Assuming I'm not completely misreading it, that would fit nicely into
> > the kernel's context-switching code, when a process has used the
> > floating-point unit.  I suspect I've misread it, though---surely
> > they'd speed up the existing instruction rather than provide a special
> > new fast one.)

> Can anyone else comment on this?

Yeah, I'm curious too.  It just looked strange to me.  I suppose I
should have made a note of where I'd read it...

Ah, found it.  It's in the 3DNow!(TM) Technology Manual, and I was
obviously reading too quickly.  

The instruction is FEMMS, which provides a quick way of switching
between MMX and floating point, as far as I understand it, by clearing
some flag but leaving the floating point/MMX registers undefined
afterwards.  Could be useful, but I just don't know enough to be able
to tell.  PREFETCH looks potentially useful too (specifically in
optimized 3DNow! code, like what might be chucked in Mesa).  Hmm, it
looks a though I was wrong: there are a few integer SIMD operations
too, which are more obviously potentially useful for cryptography (I'd
guess, anyway).

On Thu, 16 Jul 1998 16:23:04 -0400, Chris Hedemark

Doubtful.  Most crypto is designed to work only on integers, as that
is the easiest to put into hardware.  There may be some use for
elliptical curves, but even that is generally though of as integer
operations.  

Theoretically, one could use 3DNow to fake integer operations.  I
believe Quake 1 did something similar.  It used the floating point
processor to parallelize some rendering code.  It wasn't really using
floating point, just taking advantage of the second processor.  There
might be a similar use for 3DNow.  I don't know.  This is beyond my
capabilities.

OTOH, there definitely are applications that use floating point that
are not graphical, particularly in the scientific arena.  These could
definitely use 3DNow.

No faking required.  3DNow provides some integer operations,
apparently.  They may not be useful ones, of course.

Or you can get the Linux version of Acrobat Reader from Adobe.

--
Jim Zubb

Or you can get the Linux version of Acrobat Reader from Adobe.

--
Jim Zubb