AGP vs PCI graphics

Building a new PC requires deciding on an AGP or PCI motherboard.
Nvidia's Quadro FX adaptors seem to be great performers but is one (AGP)
better than the other (PCI)?
All responses appreciated.

 

Mike,

AGP would be much faster then PCI. Unless you meant ExpressPCI. Then
they would be about equal today, but in the future ExpressPCI will have
the edge.

Paul

 

PCI Express is the next generation graphics bus and is the natural
choice when bying a new PC.

 

PCI Express x 16 is available and already faster than AGPx8

 

"better than the other" for what? You mean to run Pro/e? If so, the answer is no,
it doesn't matter. It is technically so irrelevant that it's never even mentioned
in the definition of PTC's Certified/Supported systems or graphics cards. If you
can save $100 on an nVidia Quadro FX3000 with the AGP bus, get that one. I doubt
that it would even make a difference to Pro/e whether it was AGP 4 or 8. The right
driver for your card based on the processor, OS and date code of Pro/e is a MUCH
bigger issue. I wouldn't buy (or put together) a system without consulting this
list:
http://www.ptc.com/WCMS/files/30892en_file1.pdf

Even if you're not buying an HP or Dell workstation, you should still make yours
as close to a known, certified/supported hardware configuration as you can. This
list also tells what the appropriate drivers are for a particular setup. Now
that's not to say that Pro/e won't run on the $1000 gaming machine with a $300
video card. The problem with such hardware is that the motherboards don't support
as much memory or as fast processors as you'd like and the GPUs have hardware
support for DirectX 9 which does Pro/e no good and little or no support for OpenGL
which hurts Pro/e, causing freezes, CTDs and BSODs. nVidia actually sells plenty
of those gaming type cards too. What graphics cards and hardware configurations
are you considering? While I've thought for a while about putting one together, I
get lost just trying to sort out motherboards.

 

The point is that AGP is beeing replaced by PCI Express, currently most
graphic cards comes in both versions but this might not be the case in
two or three years.

 

and doesn't come with any type of support in case you run into any trouble.

That's not true. First, _all_ Quadro GPUs are basically Geforce GPUs (yes,
the cores are identical). The same is valid btw for all ATI FireGL cards
(except the ancient FireGL 2/3/4 and the older cards made by Diamond) which
are basically Radeon cards. The only difference is the cards BIOS, and a few
resistors that decide if the GPU IDs itself as Geforce/Radeon or
Quadro/FireGL (and of course other addons like more memory and stereo out on
the professional boards). With Nvidia, even the drivers between Geforce and
Quadro are identical (the certified Quadro drivers are just plain ForceWare
drivers that have been certified). ATIs FireGL driver set shares the
Direct3D part with the Catalyst drivers for Radeons but have a different
OpenGL driver than the Radeons (the Radeon Mobility also uses the same
OpenGL drivers like the FireGL Mobility).

We tried Pro/E and other MCAD applications on several Geforce cards. Except
some problems which were caused by the drivers (the problems also were
pertinent when running these drivers on a Quadro) the Geforce cards runs
these apps as good as a more expensive Quadro card. Of course the Geforce
cards lack certain features like AA lines, there is no certification, and of
course if one experiences problems there won't be any support. But say for a
student who wants to make his first steps a Geforce card is a valid
alternative.

As to Radeon cards (desktop, non-mobility versions), well, we experienced
several problems with them. The OpenGL driver of the Catalyst driver isn't
very good. On notebooks, the experience is much better. Our Radeon Mobility
systems had no problems with OpenGL apps.

If you want to build a real workstation from generic parts, I'd say avoid
that. Especially since complete workstations (i.e. HP xw8200 and xw9300)
usually can be had for the same price - with real service...

Benjamin

 

Ahh, I forgot about softquadroing. Good you mentioned this. Any of the tecnical
stuff (what cards you can do this with, techniques, etc) is MOST welcome. Maybe I
can finally get my Geforce2 Go card running right.

Drivers drivers drivers... that's always a big deal. And, yes, nice distinction:
softquadroed gaming cards for students. However, that leaves a MUCH narrower range
of cards for the professionals

Good point, service is a big deal. That and the decent price of a baseline
workstation from Dell or Compaq is certainly the reason I started there. But, by
the time I got to the system I really wanted, I was suffering sticker shock ~
$3500-$4500 dollars!!! Now, admittedly, I haven't really dug into this yet, but
I'm guessing I could get a comparable system, built from parts, for about $1000 to
$1500 less.

BTW, let me say I'm happy to finally meet someone here who knows the intimate
details of soft quadroing gaming cards (geforce to quadro fx). I've studied this
for a while, trying to improve my Toshiba and got nowhere. It is supposedly
possible. This would make a MOST USEFUL ADDITION to the only official FAQ for
comp.cad.pro-engineer, a graphics card survey by a Columbia prof named Blair
McIntyre. He is still listed as the maintainer but the documents haven't been
updated since the mid '90s. Interesting presentation, though. And an issue that's
certainly worthy of being kept up to date.
http://www.faqs.org/faqs/by-newsgroup/comp/comp.cad.pro-engineer.html

BTW2, you said nothing on this hot AGP v PCIe "controversy". Seems, in fact, like
you avoided it, Benjamin. How come? I'm sticking to my contention that it's a
nonissue regarding Pro/e performance. What's your take on it. Technical details
would be a welcomed breath of fresh air in contrast to all the stale marketing
hype we've heard so far. Someone needs to deal with the fact that, in head to head
tests of systems with AGP 8 v PCI-e 16 cards, the reputed doubling of the PCI-e
cards bore no fruit: the cards have shown comparable results for the two years
that PCI-e has been out. Disappointing for PCI-e; no wonder everyone points to the
future.

 

Be careful I think you are mixing something up. In the first place, this all
has nothing to do with tools like SoftQuadro (these tools just use the
cirumstances I described for their function).

Let me go back in history:
Years ago, there was a clear separation between consumer and professional
gfx. Consumer cards either were 2D only (with great new features like video
playback acceleration) or had very basic 3D capabilities with a very limited
function set which was enough for games. These consumer cards already were
single chip designs where the complete GPU (exclusive RAM and on older
models also the RAMDAC) was made as one IC, and these cards used unified
memory (one memory for everything). On the other sides, there were
professional gfx boards for PCs and workstations. These cards were discrete
designs with lots of ICs where every chip was for a certain function (i.e.
Raster processor, texture, geometry etc), often also combined with separate
memory subsystems (i.e. separate memory for frame buffer and textures). And
these cards were hell of expensive which also due to the high production
costs of the dozen or so different processors.
As time went by, the mass market consumer cards got better and better, and
the product cycles got shorter and shorter. On the other side, professional
gfx card development went a lot slower, because of the higher costs and the
lower sales figures (and other things like the demise of the traditional
RISC workstation market). Very soon, cheap consumer card were able to catch
up (or even surpass them!) with professional cards that cost several
thousand dollars performance-wise. Nvidia realized very soon that their TNT2
chipset (which already had a fully working OpenGL driver) does extremely
well against expensive professional cards like the ones from 3DLabs, and
together with Elsa build the first professional gfx card based on a consumer
GPU (Gloria Synergy II, GPU came from Nvidia, drivers from Elsa). Due to the
success Nvidia increased its effort in the professional market. The (at that
time new) Geforce256 consumer GPU with DDR interface was also sold as Quadro
for the professional market (Nvidia had the drivers certified). The GPU was
identical to the Geforce256 GPU. The difference is only a set of resistors
that tell the GPU if it should identify itself as "Geforce" or "Quadro". The
unified driver checks this identification and if the GPU says "I'm a Quadro"
the driver enables additional features (i.e. AA lines), alters it's
performance and quality settings, and offers more settings in the control
panel. That's it. The same is valid for all other Quadros that followed,
they are identical to the corresponding Geforce GPUs. The only difference is
in the layout of a few resistors on the PCB and in newer models also the
BIOS.
BTW: the tool "SoftQuadro" uses this to "convert" an Geforce into a Quadro.
It tricks the driver to think that the gfx processor said "I'm Quadro" when
it really said "I'm Geforce" so that the driver activates the Quadro
features.

When ATI realized that what Nvidia did worked for the professional gfx
market, they bought the gfx part of Diamond Multimedia to get the quite
reputable "FireGL" label. First ATI continued to sell the FireGL 2/3/4 which
were designed by Diamond and still consisted of discrete processors from
IBM. When ATI had success with their Geforce-competition Radeon, they also
decided to use the Radeon GPU for the professional gfx market. The first
FireGL card that was based on a Radeon was the ATI FireGL 8800 (based on the
Radeon 8500). But while Nvidia has had a very good OpenGL driver for several
years already and so used the same drivers for Geforce and Quadro ATI
couldn't do that. Their OpenGL drivers were aimed at Games, supported only a
subset of OpenGL and also already had issues with several games (forget
about OpenGL applications). So they incorporated a new OpenGL driver subset
for the FireGL series.

Until some time ago I modified lots of Geforce cards to Quadros. But today I
don't since Quadros and FireGLs are quite cheap now. I would never invest
any time in modifying a GF2 any more. Used Quadro2 Pro cards go for
sub-30EUR on ebay, and a few bucks more bring you cards like a Quadro4
500XGL which is much faster and also able to use the latest drivers. FireGL
cards like the T2 are also dirt cheap now.

As to the Nvidia Geforce Go, the story is a bit difficult. Unlike the
desktop GPUs for which drivers can be downloaded by Nvidia the drivers for
their mobile GPUs have to come from the notebook manufacturer. Nvidia
provides them with a driver development kit which they can use to create
drivers for their notebooks.
The same was valid with ATI btw, but ATI now supports more and more of their
mobile GPUs with their Catalyst Mobility drivers. Nvidia only offers
drivers for the Geforce Go 7800GTX.

Yes, it is. And a terrible one if you have to run several applications from
which every one requires a different driver. But today we use the latest
ForeWare drivers with all our Quadro cards. Even if they are not certified
they run fine with everything.

A student doesn't need SoftQuadro. Every somewhat current Geforce card
should do everything he wants just fine - without any modification (be it
hardware or software). He usually can live without the Quadro features like
AA lines.

To some extend, yes. Of course if you make your living out of this work I'd
strongly get a certified gfx board (or better a complete system) which means
FireGL or Quadro. But even there the price range is very wide today, entry
level Quadro cards start at ~80EUR (Quadro NVS, mostly for 2D work) or
~160EUR for a Quadro FX 330 (3D card). ATI also starts at ~160EUR with their
FireGL V3100 3D cards. There's something for almost everyone...

When I bought my HP xw8200 a comparable self-made system would have been
around 150-200EUR cheaper (but also would have come without OS and 3yr
onsite service). If I would have gone the RENEW path it would have been even
cheaper.

Don't know about other countries but here in Germany building a workstation
(means: some Opteron or XEON system, not a cheap Athlon64 gaming rig) by
oneself isn't cheaper than buying a system from a reputable vendor like HP.

I only had a short look into it but it looks _very_ outdated. The GPUs
which are mentioned there are all pre-2000 aera and I doubt that they are in
much use today any more - especially for CAD...

But one thing to mind is that at the time this FAQ has been made there were
dozens of different gfx chipsets out there, and choosing the right gfx was
very difficult. Today, the majority of cards use GPUs from either Nvidia or
ATI, and both manufacturers have unified drivers. It's much easier today
than it had been at the time this FAQ is from...

Well, I didn't avoid this theme, I wasn't sure if I should anything to it.
But ok, here I come ;-)

In short, there is basically(!) no real-world performance difference between
AGP and PCIe (PCI Express). There is none in the latest super-gfx highend
games, and there is none for less performance demanding tasks like CAD. So
no need to throw your fast enough and well working AGP machine out to get a
new PCIe system

So why PCIe?

AGP is basically an enhanced PCI bus, but unlike PCI which can have multiple
devices on the bus AGP only has two (gfx card and AGP bridge in the
chipset). Like PCI, AGP is a parallel bus (several parallel signal lines).
AGP has been improved over the year, but still suffers from several
backdraws:
- only one gfx card possible (you can't have more than one card on the AGP
bus; systems like SGIs Prism which have several AGP busses need a _lot_ of
technical effort for being able to do this)
- as bus performance is increased, frequency also increases which leads to
increased noise and other interference radiating from the signal lines. As
they run parallel, every line catches the interference from all the
neighbour lines which leads to signal deformation.
The second point (interference) is also a problem with PCI (and PCI-X).
Another problem with PCI/PCI-X is that the bus has to be shared between all
devices that connect to this bus. Depending on what cards you have PCI
already is a real bottleneck.

Since it was clear that the current parallel busses have too many drawbacks
that prevent them from being improved much more, something new had to be
invented. And this new thing is PCIe.

PCIe is a replacement for AGP and PCI. It's not a parallel bus but consists
of serial two-point connections (called "lanes") from which every lane does
up to 250MB/s. To increase performance it's possible to combine several
lanes 8up to 32) so that the throughput sums up. This for example is done
with gfx cards. PCIe cards use a PCIe 16x connector (also called "PEG" which
means "PCI Express Graphics") which uses 16 lanes to connect the gfx card to
the chipset resulting in up to 4GB/s throughput. You can connect less lanes
to an PCIe 16x slot (i.e. 8 lanes, like it's done by some SLI boards) which
limits throughput but still works. The main advantage of PCIe is that every
device has it's own connection, there is no bus to be shared over several
devices which is a great improvement. And PCIe has much more room for foture
enhancements than PCI and AGP.

So what does this mean for the user?
Not much. If you have a good working AGP system you are not limited in
performance because it's not PCIe. But you should be aware that even when
PCIe has no performance advantage today(!) that this is the system of the
future and that PCI and AGP are obsolete. If buying new, buying an AGP
system today means locking itself out of most of the new hardware that will
come out. So if you buy new go PCIe.

One note to performance: I said that today there is no performance advantage
of PCIe. That is not fully true. There is no performance advantage of PCIe
systems because of PCIe, but since the fastest gfx cards are only available
as PCIe and since also the latest boards with fastest CPUs, memory and
chipsets are also PCIe only, even today PCIe can give you more performance
than any AGP system.

BTW: the same discussion (why a new standard? Why not using the older
system?) happened when computers moved from ISA to VESA Localbus, then from
Vesa Localbus to PCI, and it also happened when AGP was introduced.
Everytime a new slot comes up, it has had no real-world performance benefit
first. But even when the first AGP cards were not faster than the PCI cards
of the same time, AGP fastly was mandantory because PCI got a real
bottleneck. The same will happen with AGP - it's not a bottleneck today, but
it will be in the future.

Benjamin