Re the PPT PCMCIA Presenter, exactly. The Newton 2000/2001 and eMate's hardware was designed to allow for some awesome improvements. Too bad Apple wasn't able to exploit them.
Fwiw, the Newton 2000/2100 had a 480x320 16-shade greyscale screen. (What did the eMate have, the same?) Internally it ran at 8bit color as well (there were video out cards which took advantage of this).
This is more than a printer, it is a revolution!
on
Books on Demand
·
· Score: 5
"Working from a digital file, it can print, bind, and trim a book of any size in a matter of minutes."
Wow. The scalability of this things is awesome! To be able to make books at the nanolevel all the way up to ones with enough pages to bridge the earth and moon! The implications of this new technology reaches far beyond any printing system, that's for sure.
To be able to cut paper of *any* size, whether having a length of the sun's or an atom's diameter. Amazing.
And talk about strong bindings!
A case can be brought up against a person only if the offense has been in the past year.
A couple fwiws:
The honor system is a single sanction system. If you are found guilty of lying, cheating, or stealing, found to have intended the act, and if it is determined that similiar acts would be detremental to the coummunity of trust, the student is asked to leave the university.
Otherwise, all records of the trial are destroyed.
Students (and alumns) are only formally bound by the honor system while in the county in which UVa is or when the person identifies themself as being from UVa.
One of the primary reasons I chose to be here at UVa was because of the honor system, and while any system certainly has issues, I've found the honor system to work amazingly well. I'm able to leave during finals and come back if I please, take home tests are abundant, I can leave my things laying about outside and they're still there when I come back, etc, etc. Sure, you have to use common sense, but it really is amazing how nice it is to be able to pretty much trust everyone!
As long as your main bus is fast enough and you have memory, you could always buffer a *whole* lot of data and be fine, but that's not a very pretty solution.
Plextor now has cd writers (and I assume rw) which stop writing when they hit about 10% of their buffer and can go back and start back form where they stopped within spec limits. Very cool.
And this doesn't include the older big machines either like Onyx, the Challenge series (which had somewhere around a 1 or 2GB/s transfer rate), Crimson, etc.
And then of course sun does well too, though I don't know as much about their machines (and someone already mentioned alphas).
Indy (as does Indigo2, both out in 1993) had a 400MB/s memory bandwidth and a 266MB/s gio-64 bandwidth. The R4k Indigo (circa 91 I think) had these same limits. So if an earlier post is right that the p3 has a 400MB/s bandwidth, sgi's had that about a decade ago.
The O2 (which you can get for under $1500 now) has a 2GB/s memory bandwidth, and the Octane has it's switch at around 1.6GB/s (memory bandwidth is dependent on cpu clock). Origin's memory bandwidth scales linarly with the number of nodes you have (each node having up to two processors and its own ram). So when the pentiumpro had just started coming out, with what intel pushed as its really fast l2 cache -> cpu bandwidth (1.2GB/s), there were computers from sgi with faster system memory (though more latency of course). Interesting.
Actually, I said Indigo2 *Impact*, huge difference (at least 2-3x for 3d, 2-6x 2d, and ~infinite for texturing). Impact graphics came out in 1995 and are one *huge* leap over the previous generation. So, do the comparison again with an Impact with texture momory (High Impact, or MaxImpact (2x HighImpact)). Also, a semi-recent cpu (like a 200 or 250Mhz R4k with 2MB L2 cache, or a 175/195MHz R10k (2-6x 250MHz R4k)) will help out a lot. (the above configured box is faster at bzflag than a friends 600MHz p3 with a g400).
After doing a comparison with a machine more like the above, also keep in mind quake[all versions] was designed with fairly basic pc 3d cards in mind, thus it doesn't try to do anything normally expensive (time wise), which on an Impact card may not even change the speed at all. As for the cost of an Indigo2, I recently purchased a 250MHz L2 cache I2 with MaxImpact graphics, 256MB ram, 4.5GB drive, dds-2 drive, cdrom, 20" monitor, and all the standard for less than $600. Not too bad. You could buy a pc with a faster cpu for less, but try to buy a pc with the same amount of ram and a 20" monitor, you'll be pressed to do so.
Not to get into a "this box is better" war , but some sgi workstations are very well priced against pcs, you just have to look for the right ones, and know what you want (hint: if you want the fastest dnet scores, get a pc with 8mb of ram). Oh, and keep in mind that the Indigo2 did come out in 1993! And Indigo2 Impact in 1995. In 1993 the pentium pro was of course non-existant, in 1995 the pentium pro was just hitting production lines (just to give you something to compare it with).
The O2 wasn't designed to excel at 3d graphics, rather for streaming image manipulation (which is why the O2 uses Unified Memory Arch at 3.2GB/s). If you want great OpenGL speed, get an Indigo2 Impact or Octane I or E or Octane2 with vpro graphics.
With regard to irix support/devel tapering off, have you seen sgi's roadmap? (http://www.sgi.com/developers/feature/2000/irix_m ips.html) How many other companies have posted their plans through 2006? Not many... SGI is going strong on irix, and will be for a long, long time. All that is changing is they are hiring *new* people to work on linux.
Of course, you also have to keep in mind that the hardare in the O2 came out about four years ago, that's a pretty long time in hardware-time. PCs still don't have the memory bandwidth of the O2 I don't think.
Since the R4000 came out mips cpus have been fully 64bit capable. How do you define if a machine is 64bit? If the bus size is 64bit? Ints are? Shorts? Longs? Pointers? It's hard to say "a machine" is 64bit.
Anyway, the O2 (using R5k) runs n32 (and older binaries), which are the same as n64 with only one exception: memory-addressing is 32bit rather than 64bit. The assumption is that since you have only 1gig of ram, needing to address more than 2gigs of memory isn't really a problem. It also means pointers and such use *half* the memory they would otherwise. The R4k Indigo2 is the same way, but if you use an R10k in it, you have 64bit pointers (I would have thought R10k and R12k O2s would do the same, but according to you they don't, oh well).
And the O2 does have some nice hardware, check out the docs to ICE on sgi's website sometime. The O2 was designed to do multiple streaming-videos on many objects, and to be an inexpensive sgi workstation. Thus, certain features have to be left out (if you need fast 3d, get an Indigo2 MaxImpact for less than the O2 costs, or an Octane if you need more memory bandwidth or faster cpus).
It's great to see an additional os run on sgi hardware, but at least for a while you likely won't see many switch over. Why?
Sgi does offer maintaince releases of irix for free on their site (several hundred megs, but at least they do offer it). Machines all the way back to R4k Indigo are still supported, so if you have the ram for Irix 6.5 you'll do well (at least 64MB, 96MB being much better on most hardware).
But mainly, sgi boxes are so fast because of their special hardware features! *Esp* boxes like the Indy and O2. The O2 R5k is a pretty slow box cpu-wise, but with crime graphics and ice it is **fast**. Take image manipulation for example: ont he main cpu you have probably pentium-similar performance. Recode your image manip tool to use OpenGL (and ice if you can) and you'll speed most everything up many times. ICE lets you resize, scale, re-color-code, color-space-convert, etc, etc all in *real time* on pretty large images. Its features like this that make sgi's so nice.
Now of course, netbsd once matured on sgimips may be more stable than irix (though they've done wonderfully with 6.5, my I2 easily goes hundreds of days before our house has a poweroutage, or it would go longer), and you have a current os as long as someone else (or yourself) is interested in keep the hardware current with the rest of the os. And then there's the ability to change your kernel and add your own changes, which you can't do with irix.
So...I'm happy (really happy) to see people taking interest in this port (and of course I'd love to see the Indigo^2 be supported too, right now I think linux boots on Indigo^2s (non-Impact?)), hopefully those who were so generous to work this far will also be interested in taking advantage of the machines, that'd be awesome to see!
A company Time Domain is presuing will likely near-completly change the way we use wireless communications. Several groups have been working on similiar technology over the years, but TD is the closest to a great implementation that I know of (though I know of them mostly because they are in my hometown, Huntsville, AL).
Ok, quick explanation: Instead of sending out a constant signal technology like what is being worked on at TD transmits on a range of frequencies (say 700MHz through 1.3GHz just as an example) with sharp (non-sine wave), semi-randomly timed pulses (the timing is known at both the recv and trans ends) at levels under the level of noise in dB. Now both the trans and recv ends know when to open up and listen or send data, and then shut down very quickly (application for OGR here for those interested in dnet). This does several things for you: it is now *really* hard to "fox-hunt" transmissions. Why? Because the signal is so extremely low that its very difficult to try to watch for it. It looks like plain old noise at this threshold. Secondly, since the transmission is so short and random, you can't monitor it for any real period of time.
Through this method you also gain the ability to have large numbers of users all on the "same freq" (not really as I said before, but sort of) because of the differing times which each pair (doesn't have to be just one recv of course) knows to listen for.
Right now they are beginning to market their "PusleOn technology", which emphazies the lower powered nature of this technology ("Megabits at microwatts I believe is their quote).
Might want to check out http://www.time-domain.com/
There are three primary levels of "classifiedness" for government work (in increasing order): confidential, which is almost never used, secret, which is most everything, and top secret, which is almost never used. They are rated by the degree to which the nation could possibly be hurt should the document become public knowledge: from "little," to "serious," to "grave."
http://www.ostgate.com/classification.html is a good summary.
Back in the mid-1800s the then-head of the USPO said that everything worth inventing had been invented, and wanted to (or did?) close the Patent Office. Similiar concept here. Digital computing has only been around for about fifty years, how someone can say that the human race has perfected systems-level-research as much as we ever will I don't know. We will keep having new ideas for at least a *very* *long* time to come.
Actually, almost all of the NewtonOS (everything except for basic hardware interfaces, and apple's hwr) is in newtonscript. There haven't been any implementations of newtonscript anywhere except on arm cpus, but there is no reason for this (the two are completely unrelated). Apple's HWR is in C++ actually (and designed and coded under unix).
GNUton is a project to implement the NewtonOS in python, check out http://archive.dstc.edu.au/AU/staff/david-arnold/n ewton/gnuton/
I haven't had a chance to read much about XF4 yet, but I thought one of the primary points of 4.0 was the code dedicated which gives XFree its own loader, allowing XFree to use the same binary driver under linux/i386, openbsd/sparc, etc.
Now I would much rather have a free[speech] driver, but at least (or maybe this is a bad thing...) it is usable on all platforms now.
Do you mean that it's not possible, or that the person to whom you were replying hasn't?
Just in case it is the former, I have used several Origin2000s through text terminals and Indigo2s w/o the windowing system (you get a console instead of XDM and X). AFAIK, any sgi box can have a console over the first serial port.
I haven't used AIX, but I would assume it's pretty much the same.
Does anyone know why there have not been any Irix builds for *any* of the milestones?
I have tried many of the nightly builds over the past couple months, but they all have problems mapping the gtk library for some reason (Irix 6.5.6m, R4k; gtk 1.2.5 and 1.2.6).
I have tried build from nightly and milestone releases using gcc 2.8 and 2.95 several times, but it seems I need sgi's cc to bulid nspr.
I did find a page detailing how to build mozilla under Irix (didn't work), but the person who wrote that page has since left sgi, and after I was able to get in contact with him, he said that he couldn't really help.:(
Could anyone point me in a direction to an Irix build? Or even better maybe some pointers on why nspr doesn't build with Irix/gcc (but it does with Linux/gcc or Irix/cc)?
I'm sure there are technologies like this for other platforms, but . . .
It's name is XIO. The interconnect system used on SGI's Origin, Onyx2, and Octane systems. With a slower cpu (250MHz R10k) you get around 720MB/s peak. That means, with a Origin2100 (the lowest-end of SGI's "supercomputers", others can do much better, maybe even by an order of magnitude or so) you would get around 5.0GB/s sustained (or 6.24GB/s peak). As you speed up the cpu (the 300+ MHz R12k), the bandwidth is increased.
This of course means that even a desktop Octane is able to do 1.6GB/s/port (six ports of course). With only two cpus the Octane would begin to run out of steam after a little while, but that still blows the pants off of standard pcs.
Of course, this has been around since 1996; sgi should be releasing a new line to replace the current in the coming months, so it will be exciting to see what *it* will be capable of.
cray does sell the te3-1200e of course (http://www.sgi.com/t3e/tech_info.html), with a 122Gb/sec max bisection bandwidth and up to 2.4TFLOPS peak cpu performance. And up to 2048 (when liquid-cooled) processors.
The crays they compare to are pretty old beasts, and they only tested with a few processors (Cray's SV1 for example can take advantage of over 1200 cpus!).
Drop by http://www.sgi.com/sv1/tech_info.html (or http://www.cray.com/) to see info on the SV1 if you're interested.
Now, don't get me wrong; this is a very nice cluster, but them seem to unfairly compare it to a cray (the t3e-900 is not even a recent machine!). I'm sure someone else will explain where computers such as crays and sgis come into real use (high-throughoutput work), but for distributed systems requiring less than gigatnic amounts of communication bandwidth, beowulfs do handle many kinds of tasks very well (and cheaply!).
Just didn't want eveyone to think a 16-node g4/g3 cluster was faster than a cray (actually, the sv1 can use cpus/each/ capable of 4.88GFLOPS).
Sorry about posting "KDE smokes X...", that was a bit of sarcasm on my part in response to the previous poster. I went on to explain the various levels of the X windowing system in hopes of clearing this topic up for a few people. I'll try to either not use sarcasm in my next post, or be sure to include a winking face or some such.
Slashdot Mirror
Re the PPT PCMCIA Presenter, exactly.
The Newton 2000/2001 and eMate's hardware was designed to allow for some awesome improvements. Too bad Apple wasn't able to exploit them.
Fwiw, the Newton 2000/2100 had a 480x320 16-shade greyscale screen. (What did the eMate have, the same?) Internally it ran at 8bit color as well (there were video out cards which took advantage of this).
"Working from a digital file, it can print, bind, and trim a book of any size in a matter of minutes."
Wow. The scalability of this things is awesome! To be able to make books at the nanolevel all the way up to ones with enough pages to bridge the earth and moon! The implications of this new technology reaches far beyond any printing system, that's for sure.
To be able to cut paper of *any* size, whether having a length of the sun's or an atom's diameter. Amazing.
And talk about strong bindings!
SGA is Student Government Association, AKA Student Council.
UAH is the university this article is about, the Unversity of Alabama in Huntsville.
A case can be brought up against a person only if the offense has been in the past year.
A couple fwiws:
The honor system is a single sanction system. If you are found guilty of lying, cheating, or stealing, found to have intended the act, and if it is determined that similiar acts would be detremental to the coummunity of trust, the student is asked to leave the university.
Otherwise, all records of the trial are destroyed.
Students (and alumns) are only formally bound by the honor system while in the county in which UVa is or when the person identifies themself as being from UVa.
One of the primary reasons I chose to be here at UVa was because of the honor system, and while any system certainly has issues, I've found the honor system to work amazingly well. I'm able to leave during finals and come back if I please, take home tests are abundant, I can leave my things laying about outside and they're still there when I come back, etc, etc. Sure, you have to use common sense, but it really is amazing how nice it is to be able to pretty much trust everyone!
Fwiw, SGI uses Rambus memory in their IMPACT graphics systems (IMPACT for the Indigo2 came out in 1995).
As long as your main bus is fast enough and you have memory, you could always buffer a *whole* lot of data and be fine, but that's not a very pretty solution.
Plextor now has cd writers (and I assume rw) which stop writing when they hit about 10% of their buffer and can go back and start back form where they stopped within spec limits. Very cool.
And this doesn't include the older big machines either like Onyx, the Challenge series (which had somewhere around a 1 or 2GB/s transfer rate), Crimson, etc.
And then of course sun does well too, though I don't know as much about their machines (and someone already mentioned alphas).
Indy (as does Indigo2, both out in 1993) had a 400MB/s memory bandwidth and a 266MB/s gio-64 bandwidth. The R4k Indigo (circa 91 I think) had these same limits. So if an earlier post is right that the p3 has a 400MB/s bandwidth, sgi's had that about a decade ago.
The O2 (which you can get for under $1500 now) has a 2GB/s memory bandwidth, and the Octane has it's switch at around 1.6GB/s (memory bandwidth is dependent on cpu clock). Origin's memory bandwidth scales linarly with the number of nodes you have (each node having up to two processors and its own ram). So when the pentiumpro had just started coming out, with what intel pushed as its really fast l2 cache -> cpu bandwidth (1.2GB/s), there were computers from sgi with faster system memory (though more latency of course). Interesting.
Actually, I said Indigo2 *Impact*, huge difference (at least 2-3x for 3d, 2-6x 2d, and ~infinite for texturing). Impact graphics came out in 1995 and are one *huge* leap over the previous generation. So, do the comparison again with an Impact with texture momory (High Impact, or MaxImpact (2x HighImpact)). Also, a semi-recent cpu (like a 200 or 250Mhz R4k with 2MB L2 cache, or a 175/195MHz R10k (2-6x 250MHz R4k)) will help out a lot. (the above configured box is faster at bzflag than a friends 600MHz p3 with a g400).
After doing a comparison with a machine more like the above, also keep in mind quake[all versions] was designed with fairly basic pc 3d cards in mind, thus it doesn't try to do anything normally expensive (time wise), which on an Impact card may not even change the speed at all. As for the cost of an Indigo2, I recently purchased a 250MHz L2 cache I2 with MaxImpact graphics, 256MB ram, 4.5GB drive, dds-2 drive, cdrom, 20" monitor, and all the standard for less than $600. Not too bad. You could buy a pc with a faster cpu for less, but try to buy a pc with the same amount of ram and a 20" monitor, you'll be pressed to do so.
Not to get into a "this box is better" war , but some sgi workstations are very well priced against pcs, you just have to look for the right ones, and know what you want (hint: if you want the fastest dnet scores, get a pc with 8mb of ram). Oh, and keep in mind that the Indigo2 did come out in 1993! And Indigo2 Impact in 1995. In 1993 the pentium pro was of course non-existant, in 1995 the pentium pro was just hitting production lines (just to give you something to compare it with).
The O2 wasn't designed to excel at 3d graphics, rather for streaming image manipulation (which is why the O2 uses Unified Memory Arch at 3.2GB/s). If you want great OpenGL speed, get an Indigo2 Impact or Octane I or E or Octane2 with vpro graphics.
m ips.html) How many other companies have posted their plans through 2006? Not many...
With regard to irix support/devel tapering off, have you seen sgi's roadmap? (http://www.sgi.com/developers/feature/2000/irix_
SGI is going strong on irix, and will be for a long, long time. All that is changing is they are hiring *new* people to work on linux.
Of course, you also have to keep in mind that the hardare in the O2 came out about four years ago, that's a pretty long time in hardware-time. PCs still don't have the memory bandwidth of the O2 I don't think.
Saying the O2 is 32bit isn't really correct.
Since the R4000 came out mips cpus have been fully 64bit capable. How do you define if a machine is 64bit? If the bus size is 64bit? Ints are? Shorts? Longs? Pointers? It's hard to say "a machine" is 64bit.
Anyway, the O2 (using R5k) runs n32 (and older binaries), which are the same as n64 with only one exception: memory-addressing is 32bit rather than 64bit. The assumption is that since you have only 1gig of ram, needing to address more than 2gigs of memory isn't really a problem. It also means pointers and such use *half* the memory they would otherwise. The R4k Indigo2 is the same way, but if you use an R10k in it, you have 64bit pointers (I would have thought R10k and R12k O2s would do the same, but according to you they don't, oh well).
And the O2 does have some nice hardware, check out the docs to ICE on sgi's website sometime. The O2 was designed to do multiple streaming-videos on many objects, and to be an inexpensive sgi workstation. Thus, certain features have to be left out (if you need fast 3d, get an Indigo2 MaxImpact for less than the O2 costs, or an Octane if you need more memory bandwidth or faster cpus).
Hope that clears up any misconceptions!
It's great to see an additional os run on sgi hardware, but at least for a while you likely won't see many switch over. Why?
Sgi does offer maintaince releases of irix for free on their site (several hundred megs, but at least they do offer it). Machines all the way back to R4k Indigo are still supported, so if you have the ram for Irix 6.5 you'll do well (at least 64MB, 96MB being much better on most hardware).
But mainly, sgi boxes are so fast because of their special hardware features! *Esp* boxes like the Indy and O2. The O2 R5k is a pretty slow box cpu-wise, but with crime graphics and ice it is **fast**. Take image manipulation for example: ont he main cpu you have probably pentium-similar performance. Recode your image manip tool to use OpenGL (and ice if you can) and you'll speed most everything up many times. ICE lets you resize, scale, re-color-code, color-space-convert, etc, etc all in *real time* on pretty large images. Its features like this that make sgi's so nice.
Now of course, netbsd once matured on sgimips may be more stable than irix (though they've done wonderfully with 6.5, my I2 easily goes hundreds of days before our house has a poweroutage, or it would go longer), and you have a current os as long as someone else (or yourself) is interested in keep the hardware current with the rest of the os. And then there's the ability to change your kernel and add your own changes, which you can't do with irix.
So...I'm happy (really happy) to see people taking interest in this port (and of course I'd love to see the Indigo^2 be supported too, right now I think linux boots on Indigo^2s (non-Impact?)), hopefully those who were so generous to work this far will also be interested in taking advantage of the machines, that'd be awesome to see!
A company Time Domain is presuing will likely near-completly change the way we use wireless communications. Several groups have been working on similiar technology over the years, but TD is the closest to a great implementation that I know of (though I know of them mostly because they are in my hometown, Huntsville, AL).
Ok, quick explanation:
Instead of sending out a constant signal technology like what is being worked on at TD transmits on a range of frequencies (say 700MHz through 1.3GHz just as an example) with sharp (non-sine wave), semi-randomly timed pulses (the timing is known at both the recv and trans ends) at levels under the level of noise in dB. Now both the trans and recv ends know when to open up and listen or send data, and then shut down very quickly (application for OGR here for those interested in dnet). This does several things for you: it is now *really* hard to "fox-hunt" transmissions. Why? Because the signal is so extremely low that its very difficult to try to watch for it. It looks like plain old noise at this threshold. Secondly, since the transmission is so short and random, you can't monitor it for any real period of time.
Through this method you also gain the ability to have large numbers of users all on the "same freq" (not really as I said before, but sort of) because of the differing times which each pair (doesn't have to be just one recv of course) knows to listen for.
Right now they are beginning to market their "PusleOn technology", which emphazies the lower powered nature of this technology ("Megabits at microwatts I believe is their quote).
Might want to check out http://www.time-domain.com/
There are three primary levels of "classifiedness" for government work (in increasing order): confidential, which is almost never used, secret, which is most everything, and top secret, which is almost never used. They are rated by the degree to which the nation could possibly be hurt should the document become public knowledge: from "little," to "serious," to "grave."
http://www.ostgate.com/classification.html is a good summary.
Back in the mid-1800s the then-head of the USPO said that everything worth inventing had been invented, and wanted to (or did?) close the Patent Office. Similiar concept here. Digital computing has only been around for about fifty years, how someone can say that the human race has perfected systems-level-research as much as we ever will I don't know. We will keep having new ideas for at least a *very* *long* time to come.
"all ARM native"
n ewton/gnuton/
Actually, almost all of the NewtonOS (everything except for basic hardware interfaces, and apple's hwr) is in newtonscript. There haven't been any implementations of newtonscript anywhere except on arm cpus, but there is no reason for this (the two are completely unrelated). Apple's HWR is in C++ actually (and designed and coded under unix).
GNUton is a project to implement the NewtonOS in python, check out http://archive.dstc.edu.au/AU/staff/david-arnold/
I haven't had a chance to read much about XF4 yet, but I thought one of the primary points of 4.0 was the code dedicated which gives XFree its own loader, allowing XFree to use the same binary driver under linux/i386, openbsd/sparc, etc.
Now I would much rather have a free[speech] driver, but at least (or maybe this is a bad thing...) it is usable on all platforms now.
Do you mean that it's not possible, or that the person to whom you were replying hasn't?
Just in case it is the former, I have used several Origin2000s through text terminals and Indigo2s w/o the windowing system (you get a console instead of XDM and X). AFAIK, any sgi box can have a console over the first serial port.
I haven't used AIX, but I would assume it's pretty much the same.
Does anyone know why there have not been any Irix builds for *any* of the milestones?
:(
I have tried many of the nightly builds over the past couple months, but they all have problems mapping the gtk library for some reason (Irix 6.5.6m, R4k; gtk 1.2.5 and 1.2.6).
I have tried build from nightly and milestone releases using gcc 2.8 and 2.95 several times, but it seems I need sgi's cc to bulid nspr.
I did find a page detailing how to build mozilla under Irix (didn't work), but the person who wrote that page has since left sgi, and after I was able to get in contact with him, he said that he couldn't really help.
Could anyone point me in a direction to an Irix build? Or even better maybe some pointers on why nspr doesn't build with Irix/gcc (but it does with Linux/gcc or Irix/cc)?
Thanks!
Took a look at SGI's 2800 Rack specs. I/O bandwidth goes is 160GB/s *sustained* (199.68GB/s peak). That'll handle a whole slew of 1GB/s devices.
Of course, by the time this disc technology pans out, at the possible 1GB/s speed, most new desktops will be able to handle 1GB/s probably.
I'm sure there are technologies like this for other platforms, but . . .
It's name is XIO. The interconnect system used on SGI's Origin, Onyx2, and Octane systems. With a slower cpu (250MHz R10k) you get around 720MB/s peak. That means, with a Origin2100 (the lowest-end of SGI's "supercomputers", others can do much better, maybe even by an order of magnitude or so) you would get around 5.0GB/s sustained (or 6.24GB/s peak). As you speed up the cpu (the 300+ MHz R12k), the bandwidth is increased.
This of course means that even a desktop Octane is able to do 1.6GB/s/port (six ports of course). With only two cpus the Octane would begin to run out of steam after a little while, but that still blows the pants off of standard pcs.
Of course, this has been around since 1996; sgi should be releasing a new line to replace the current in the coming months, so it will be exciting to see what *it* will be capable of.
Forgot one last thing...
cray does sell the te3-1200e of course (http://www.sgi.com/t3e/tech_info.html), with a 122Gb/sec max bisection bandwidth and up to 2.4TFLOPS peak cpu performance. And up to 2048 (when liquid-cooled) processors.
Bad pun there, but couldn't resist.
/each/ capable of 4.88GFLOPS).
The crays they compare to are pretty old beasts, and they only tested with a few processors (Cray's SV1 for example can take advantage of over 1200 cpus!).
Drop by http://www.sgi.com/sv1/tech_info.html
(or http://www.cray.com/) to see info on the SV1 if you're interested.
Now, don't get me wrong; this is a very nice cluster, but them seem to unfairly compare it to a cray (the t3e-900 is not even a recent machine!). I'm sure someone else will explain where computers such as crays and sgis come into real use (high-throughoutput work), but for distributed systems requiring less than gigatnic amounts of communication bandwidth, beowulfs do handle many kinds of tasks very well (and cheaply!).
Just didn't want eveyone to think a 16-node g4/g3 cluster was faster than a cray (actually, the sv1 can use cpus
Sorry about posting "KDE smokes X...", that was a bit of sarcasm on my part in response to the previous poster. I went on to explain the various levels of the X windowing system in hopes of clearing this topic up for a few people.
I'll try to either not use sarcasm in my next post, or be sure to include a winking face or some such.