- Anthony Frausto-Robledo, Editor (afrausto@architosh.com)
- 30 Jul 00 Second Update:
4 Aug, 8:15am [See comments in red]
AltiVec and
multiprocessing in CAD/3D
In
this article:
The strengths of the G4's architecture.
If you read our exclusive
interview of Sean Flaherty, CTO, of Nemetschek
N.A., you may have been surprised to hear him say that AltiVec
doesn't get VectorWorks that much in terms of speed up. The reality
is this comment is somewhat missing leading.
" I think technology
that speeds us up WITHOUT requiring a lot of work as the best
type of improvement." Sean
Flaherty, CTO, Nemetschek NA. |
The PowerPC G4 processor
with the AltiVec instruction set is a superior chip architecturally
for science and 3D visualization than anything coming out of Intel.
Over the next two years, with perhaps particular emphasis on 2001,
we'll see a number of amazing feats in 3D visualization running
on Power Macs with the G4 processor. To borrow words from Billy
boy himself, "the best is yet to come."
Let's Get Down to Facts
Let's talk about Sean's comments for a second. Sean mentioned that
"AltiVec primarily speeds up two major types of operations:
bit array manipulation and single-precision
floating-point." Some of you may be missing the word
'major' in there. That's important. Bit array manipulation is
a 'major deal' in rendering and animations (which are frame by frame
sequences of raster images), as well as image optimization processinglike
the kind done in Photoshop and graphics and video applications geared
toward the Web. That's why Adobe
says the Power Mac G4 is the ultimate Photoshop machine, and
why that little 'G4 AltiVec plug-in' that ships with the latest
version of Photoshop is so important to you if you have a G4.
When Sean mentioned that AltiVec doesn't gain VectorWorks that
much in terms of speed up he was primarily speaking about 2D/3D
CAD processing, not OpenGL
or QD3D-based full color rendering with texture mapping. Although
an important part of any CAD program, rendering in VectorWorks
is just one of many larger segments of the program. Still, I was
a bit confused and asked him to clarify some issues:
AFR: Sean, this makes AltiVec should like
it misses the boat on helping CAD applications and will likely
confuse readers about all of Apple's hoopla about Velocity Engine
[Apple's term for AltiVec] and rendering 3D rendering apps. Did
Apple make a mistake with AltiVec here?
SF: I don't think Apple missed the boat
at all. They sped up many raster techniques such as texture mapping
and image composition, the things that make interactive display
of 3D objects work very quickly. We support AltiVec by using OpenGL.
For some reason, people see this as us not supporting AltiVec
or AltiVec not supporting our needs, but I see it as a new chip
architecture that speeds one of the slowest operations in VectorWorks.
I think technology that speeds us up WITHOUT requiring a lot of
work as the best type of improvement.
Sean mentioned that AltiVec provides a tremendous speed up in VectorWork's
rendering, especially with texture mapping, a very important part
of high-end rendering and animation. He also said that many may
confuse a chip that accelerates floating-point with accelerating
all scientific apps across the board. That just isn't true, as these
programs use different programming techniques.
The G4: weaknesses and strengths
The G4 with AltiVec is not
perfect. One current draw back, which Motorola is reportedly working
on, is the single precision floating-point processing. Double-precision
is what the G4 should provide and will help tremendously with luring
back high-precision CAD/CAD/CAID programs to OS X. [This
paraphraph needs clarification. The G4 chip does support double-precision
floating point. Only AltiVec is limited to single precision vectors.]
{see other comments below}
Another weak area currently (and will supposedly be remedied this
Fall) is the processor cycle speeds on the G4. The G4 is losing
the megahertz war to the Pentium III's and AMD Athlons. It's amazing
how quickly this can change. It was not too long ago that the G3
was ahead and equal in terms of megahertz alone. Currently the G3
in the mobile market is not that far behind Intel in terms of mehahertz
alone.
Thankfully, megahertz isn't the only and major factor affecting
chip performance. If that was the case Pixar,
Disney and Lucas' ILM wouldn't be making movies like Star Wars and
Bug's Life on processors currently running at half the speed of
Intel's Pentium III. But they are. Those RISC-based MIPS and other
UNIX processors providing the firepower behind their high-end workstations,
ganged together in render-farms of hundreds of machines, not only
run at speeds approximate to the G4's current limit, they, like
the G4, are true RISC (reduced instruction set computing) microprocessors.
One of the most overlooked advantages of the G4 processor is
its RISC architecture. "In the industry", according
to previously published quote by Sean Flaherty, "the G3/G4
chip architecture is known for its incredible math performance,
both integer and floating-point." It is not surprising then
to learn about the parallel
supercomputing G4 clustering projects at leading universities
like UCLA. And it should not surprise you to learn that more scientific
and advanced
3D visualization applications are coming back to the Mac platform,
and in full
force.
Multiprocessing with the G4
With Apple's introduction of the multiprocessor G4 Power Macs at
Expo, we should continue to see advances in the technical market
for G4-enabled applications ready to take advantage of both AltiVec
and multiple processors. At Macworld Expo,
Black Lab Linux demonstrated parallel
computing systems with eight (8) G4 Power Macs running AltiVec-enabled
Linux apps in excess of 10 times faster (1,000%). The PowerPC architecture
not only offers scalable, parallel multiprocessing but offers the
highest performance-lowest power consumption per watt of any comparable
Pentium system.
When will all of this parallel, multiprocessing power come to our
finger tips?
At SIGGRAPH last week Newtek was featuring 13
Power Mac G4s in a render farm running an advanced beta of LightWave
6 for Mac OS X. This was an impressive sign of things to come. Also,
Maya for OS X is well advanced at this point and promises to deliver
Mac OS X only features in a powerful, multiprocessor ready package.
Over the next year expect to see more high-end 3D visualization
and CAD applications come to OS X, complete with both AltiVec and
multiprocessing in mind.
[Reader Comments]
According to Larry Rosenstein, who helped
point out the correction above, "it's unlikely that we'll see
AltiVec operating on double precision vectors anytime soon. That
would probably require Motorola extending the AltiVec register size
of 256 bits, which would be a significant change."
Larry also mentioned that AltiVec should
be useful for matrix multiplication, something that can help 3D
applications as well.
Edwin van Spronsen, of Amsterdam, wrote in
about emphasizing the difference between the rendering of RT (real
time) previews in 3D applications compared to print-ready, high
resolution rendering (which takes many hours, or even days, to complete).
Edwin mentioned that the latter type of rendering done by every
3D app uses 64-bit floating-point processing (something the G4 itself
does quite well without AltiVec). In his opinion, the future support
of 64-bit AltiVec was doubtful. Rather, Edwin feels Motorola will
concentrate on the the G5.
Should Motorola work hard with Apple to support
64-bit AltiVec? Does it matter
more then clock speed? What do you think? When the G4 first
came out, test run here at Architosh showed that the floating-point
performance of the chip was second to no Pentium chip. However,
things have changed quite a bit on the Intel side (and all in just
11 months), first with the push to 700 MHz, then with the climb
to 1 GHz, and with enhanced architectures. Now Intel has introduced
1.13 GHz (or is it 1.12 GHz?) Pentium III's.
While the thesis of this article is the G4s
superior architecture, we can't deny that the G4 chip has made little
progress in the last yearwhile Intel has made great progress,
at least in terms of MHz. Someone needs to seriously kick Motorola's
you-know-what into gear. Before this year is over Apple needs to
ship G4's running at at least 700 MHz. If it can't, Apple is going
to be perceived as making a mistake with choosing Motorola over
IBM's PPC roadmap. While AltiVec holds great promise, so too does
MHz speed, something we have not seen change with the G4 in too
many months.]
Related Articles
Feature: Interview
with Sean Flaherty, CTO, Nemetschek NA.
LightWave
Featured on Apple G4 Render Farm
Maya for
Mac OS X: No Illusion
Apple
Debuts New Power Mac G4s with Dual Processors
|