- This topic has 32 replies, 8 voices, and was last updated 16 years, 6 months ago by .
-
Posts
-
Xavier, I see we share the same ideas about discreet/sgi/PC
I totally agree, that the PCI-X is a very fast bus architecture. But I also want to add that AGP bus has some problems fro OpenGL rendering. The 8x bus can write 2.1GB/sec, but can not read with this speed. So you can preview your opengl gfx on your monitor very fast. But when it comes to reading that gfx back to system memory, AGP isn’t optimized for that. But SGI hardware is bidirectional. There are lots of cables going from video to gfx, from gfx to video. Even the OnyxRE2 could have live video textures. And remember, sirius was a seperate board. Even at very old times, 12 yrs ago. SGI hardware could do that.
According to INTEL:
AGP 8x technology is intended to be the last parallel interface step that meets the industry’s requirements before transitioning to a PCI Express-based serial graphics solution in 2004. The PCI Express architecture is a high-speed, general-purpose, serial I/O interconnect that provides a unifying standard, consolidating a number of I/O interconnects within a platform.So the PCI-X seems to be the future, also for gfx.
hello,
seems like here are some numbers missing:
pci-x v2 = 533mb/s
xio full duplex = 1,6gb/s
numaflex = 3.2gb/s each brick
tezro/fuel also have 3,2gb/s
o2000/onyx2: 6.24gb/s i/o bandwidth
o300 and above: 11,2gb/s i/o bandwidth
all above: 48bit rgbaand hard- and software are rock solid 😉
they are really nothing alike
Kuban,
Very good points. I wasn’t aware of those AGP “limitations”…
As for AGP being fast for interaction, but slow for rendering, I say: Who wants to render on their main workstation anyways? Just cluster a bunch of machines to render in the background.
Compositing is not bandwith balanced anyways (bunch of frames come in, a single frame comes out).
It is a very exciting time for graphics hardware and I can’t wait to bench a PC with “proper” software… 🙂
Foetz,
>>Seems like here are some numbers missing:
You know as well as I do that numbers mean nothing without “real-world” results… However, for the argument’s sake, I will admit that SGIs probably still have more “real-world” bandwidth than PCs.
However, 533MB/s is not “slow” by any means. It’s definitely faster than what your disk array can pump out of the disks (usually 200 MB/s for a FibreChannel Stone).
So for applications with huge datasets and “little” disk I/O (like scientific visualisation) SGIs are perfect and still have an edge over PCs. BUT, compositing DOESN’T involve huge datasets and IS VERY MUCH a disk bound process…
For a long time, PCs had fast CPUs that were starved for data because of PCI and other system bottlenecks. Now, that gap is closing.
On the other hand, SGIs are bandwith kings, but those huge pipes are not being filled because their slower CPUs can’t render data fast enough to fill the bandwidth!
>>pci-x v2 = 533mb/s
>>xio full duplex = 1,6gb/s
>>numaflex = 3.2gb/s each brick
>>tezro/fuel also have 3,2gb/s
>>o2000/onyx2: 6.24gb/s i/o bandwidthRAM and disk arrays CAN’T pump 6.24GB/s… so while impressive, these numbers don’t necessarily translate into faster renders or interaction.
For what it’s worth, from Apple’s website:
http://a1280.g.akamai.net/7/1280/51/3aa967f4ca31ab/www.apple.com/powermac/pdf/PowerMacG5_TO_111803.pdf“Frontside bus up to 1GHz. Provides up to 8-GBps bandwidth between the processor and the rest of the system.”
“Dual independent frontside buses up to 1GHz. Provides up to 16-GBps aggregate bandwidth in dual processor systems.”
“133MHz PCI-X expansion. Supports advanced high-performance PCI devices, providing total throughput of up to 2 GBps.”
I know it’s propaganda, but your numbers probably come from SGI propaganda too… 🙂
>>all above: 48bit rgba
Huh? The numbers you quote are system bandwith numbers, not graphic adapter numbers.
BTW, FFI is behind in the bit depth wars. While Shake, Fusion, Nuke and Combustion users can comp at 8, 10, 16 bits and 32 bits floating point, we’re stuck at 8, 10 (barely supported), and 12.
>>and hard- and software are rock solid
Hardware is pretty solid — but has issues like everybody else. I agree that on average SGI hardware is more robust that typical high-end PCs, though.
However, last time I checked, IRIX had patch numbers in the thousands… 🙂
Depending on the version, discreet is very average when it comes to being rock solid. Inferno 4.1 was rock solid. Inferno 4.6 was a total nightmare (do you remember entire framestores turning black?). Inferno 5.5 definitely has stability issues as well, especially when you start to mix bit depth and resolutions.
Eltopo,
What are you gonna use your new system for?
Commercials?
Feature film compositing?
Motion graphics?
With clients attending the session or not?
Multi-hour or multi-month deadlines?
NTSC/PAL only or 2K and HD as well?Depending on your answers, Flame might be well worth the money, or not. Shake is a good value, no doubt. But if it’s not doing what you need it to do, who cares if it’s cheap? iMovie is 49$… is it better than Flame or Shake?
Flame and Shake are *very* different beasts. Don’t be tempted to think they are interchangeable. However, Flame might be overkill for what you need it to do… really depends on what are your needs.
— Xavier
eltopo wrote:Here are the specs of a G5 (soon to be upgraded)
2 Ghz 64 Bit PowerPC 970
1 GHZ bus from each processor. 9Gbps of bandwith per processor
133 mhz PCI-X, 2GBps
AGP 8X. 2.1 Gbps
Firewire 800, Firewire 400, Gigabit Ethermet, USB 2 connected through
16bit bidirectional Hypertransport. 3.2 Gbps
Superdrive DVD-RW (4X)
400 Mhz ECC DDR SDRAM. 6.4 GBpsThose number are Tzero crushing not just Pentiums 4’s
So please GOD have Inferno 6 (Toxik 1) come in a Xserve Cluster… 😉
hehe, maybe you fell for the good old trick.
in general ‘gbps’ or ‘mbps’ as used for compression settings stand for megaBIT and not BYTE,
so you might have to divide the above values over 8!Foetz,
>>hehe, maybe you fell for the good old trick.
>>in general ‘gbps’ or ‘mbps’ as used for compression settings stand for megaBIT >>and not BYTE, so you might have to divide the above values over 8!Nope. These numbers are GigaBYTES not GigaBITS. (A GigaByte is defined as 1 billion bytes in Apple’s doc., to be precise).
Use the link I posted to see the PDF for yourself. It’s an interesting read anyways.
You’ll thank us for the reality check later, I swear. 🙂
— Xavier
did some research and it looks like apple is neatly regarding techspecs.
Quote:133 mhz PCI-X, 2GBpsonly 64bit, 133mhz pci-x can reach 1gb/s but not 2.
Quote:AGP 8X. 2.1 Gbpstrue.
Quote:Firewire 800, Firewire 400,the firewire numbers are megabit. so we have 100mb/s and 50mb/s.
Quote:USB 2 connected through
16bit bidirectional Hypertransport. 3.2 Gbpsusb2 has 480mbit/s => 60mb/s
so they would have to bundle 53 usb2 lines to get 3,2gb/s 😉Quote:400 Mhz ECC DDR SDRAM. 6.4 GBps400mhz ddr specs are 3,2gb/s or in some configs 3,4gb/s.
Quote:Those number are Tzero crushing not just Pentiums 4’snot looking like…
ok.
seems like this is not leading to a reasonable result.however, i worked with nearly all except for cray in many areas and for me sgi still leads.
not as much as in 1996 but still.i see that x86 and mac are coming closer but they still have their way to go.
like i said there will always be big irons, low end stuff and stuff between but as long as i can i
will go with the big irons.
if they wouldn’t be better nobody would buy them 😉maybe one day this will change and you can be sure i will be there to get the max. availible then.
but for now nothing reached the onyx experience and as long as this is the case i’ll wear my sgi
cap every day and enjoy… 😀one last technical thing.
i guess we all know that there’s much more technical stuff behind simple values.
like the agp example from kuban or how ram is used within smp systems and so on.
i’m no engineer so this is beyond my knowledge.and like we said before reality often talks another language.
work with it and this will be the best spec at all!x86 and apple don’t offer more than 2/4 cpus. all others are some kind of cluster. but sometimes
one needs more within one machine.other thing: reliability.
just one pass lost due to a simple error – and this happens with x86(i guess you know) – and
all headstart is lost even if the intel would have double the peak. so you’re not done in time and
fuc… 😉btw: for the last 2 years i earned 60% of my money by beeing hired by companies to fix linux
x86 problems. from this point of view i love linux 😀one could say so much more but enough for now.
i guess it’s all said and we should enjoy the weekend.
as long as we provide stunning clips who cares how we did it?
xavier will hope for inferno/osx, kuban will wait and i will stay with sgi as long as i find something
better.stay well
In all honesty, I wouldn’t mind seeing Flame, or any other Discreet software in the Mac, Linux, and Windows; it would have to offer me what I get from SGI, and therein lies the rub. For years we see the non-SGI world producing ever faster components, that SGI will be outperformed on all fronts for video/film work;but it doesn’t happen. Several factors come into play for this to not work so smoothly; they suffer from hardware combinations problems, no matter how much bandwidth a gfx board can do, or the video card, the motherboard seems to be the issue in most cases; and the software issue. It’s not so bad on the surface, but just watching how bad something like Combustion can be on the Mac, compared to the Windows version, tells me about the problem of which platform they will focus on. Then there’s the real world situation, might be different in a year or so, but we’ll see.
Another thing, wherever the Macs, or PCs will be in a couple of years, it might still be way behind the SGI’s.
In any case, I’m pretty open to migrating to whatever Discreet moves to, all ~I need is for everything to go as smoothly as possible. 🙂
>>> Foetz
SGI seems to give the bandwidth numbers as a total. Especially on the onyx series. Onyx series are multibus machines. Multibus seems to affect compositing speed a little, but mainly the communication speed between system memory, and texture memory is very important. So about the onyx2, we might talk about XIO bus speed. Yes, XIO is faster then PCI-X. But how expensive is a PCI-X bus?Also on onyx series, SGI likes to tell, all the added raster managers, gfx systems, etc… And add them, and tell a max. texture fill rate. Which is a gigantic texture fill. But none of the onyx owners configure such an onyx.
kuban wrote:>>> Foetz
SGI seems to give the bandwidth numbers as a total. Especially on the onyx series. Onyx series are multibus machines. Multibus seems to affect compositing speed a little, but mainly the communication speed between system memory, and texture memory is very important. So about the onyx2, we might talk about XIO bus speed. Yes, XIO is faster then PCI-X. But how expensive is a PCI-X bus?sure, as always. money is a factor…
Quote:Also on onyx series, SGI likes to tell, all the added raster managers, gfx systems, etc… And add them, and tell a max. texture fill rate. Which is a gigantic texture fill. But none of the onyx owners configure such an onyx.don’t worry, i know the exact values.
btw: would be interested in your dvd vol.3. are they out now?
- You must be logged in to reply to this topic.