Agreed. You know what's more efficient for many tasks? SMP!
Instead of one cutting edge chip pushing really fast (and being expensive, hot, etc.), most systems would do fine with two or more slower CPUs. The slower (older) CPUs are cheap and stable because they bleeding edge moved away from them.
People need to think about absolute speed vs. throughput. For a game you probably want absolute speed. But for word processing, web browsing, playing MP3s, and photo editing all at the same time, a SMP system composed of slower processors is a very good solution.
Operating systems and apps will have to improve to make this work. Windows 95/98/ME don't do SMP, and that is where most home users are. Running many independent apps (mp3 player, netscape, etc) concurrently will work fine, but Photoshop is one of the few examples of an app that will partition its work and use multiple CPUs well.
First off, I'm sorry, but there probably isn't a cool selection of translucent cases to choose from. They mention using eight i-mac logic boards, not eight entire i-macs.
Details are a little short - the article says that it's all in the same box, which implies that there is a high-speed interconnect. I hope that isn't the ethernet port on each of the i-macs.
What's so special? They're doing parallel programming against a few (almost) machines? Parallel programming types have been doing this for years, and there are highly developed libraries to help synchronize and send data to independent machines across a network which are participating in a parallel computation.
This is going to start a craze of little kids gutting their i-macs to make a super computer with all of the other little kids in the neighborhood. Be the first on your block!
They are just trying to make sure that people who participate in infringing behavior are held liable, not the technology.
Peer-to-Peer is not illegal; it has been around for a while. Directories are not illegal either. Ripping people off is illegal.
Whether Napster is facilitating the rip-off or whether people are making "fair use" of the music is an entirely different issue. That is what these groups are trying to make clear.
If I write a Logo interpreter in one of the seven languages, and then display the output on the screen as well as send the commands to the Mindstorm, then I can re-create a classic kid's learning environment from 1985 ! Except now my turtle can bite.
[1] The RIAA had no problem selling me the Rolling Stones in record, cassette, and CD format. Why is it that every time the format changes, I get to pay the artist (and the RIAA) again? The was especially annoying when the media suffered from wear, like cassettes and vinyl records did. CDs are not immune either - give it a good scratch, and you get to pay again.
[2] What happened to singles? CD singles existed for a brief period. I hate paying $18 for a CD that I only want one or two songs from, and most of the CDs in my collection are like that. Seems like gouging to not make those songs available in single format, and to charge less for it.
[3] MP3s are generally 128bps, which is pretty low quality. It is slightly better than radio broadcast quality. Why get upset about MP3s available electronically when people can still tape off the air?
[4] How does one obtain (or even purchase) out-of-print music? Try to find "The Screaming Blue Messiahs" on CD - I dare you. (There is a record store with some of the out-of-print vinyl still available.) The major labels have chosen to make music unavailable... is that ethical? Is it ethical to complain when people start giving this music away, music that can't be bought?
[5] What about the people experimenting? I'm discovering Jazz and Swing for the first time... how else can I do this? My local library isn't a source.
[6] Speaking of which, are libraries the next target? They loan out CDs which can be copied.
[7] Napster would be a lot more defensible if they weren't a commercial enterprise. There are a lot of legitimate uses for it, but we know that most people are leaching. The RIAA can be evil, but some artists insist on being compensated and that is their right. And supposedly, the RIAA is defending that right. Too bad... I think the RIAA has been the bigger leach for the longer period of time.
It is a civic responsibility to disobey unjust laws. ("Civil disobedience.) Part of the reponsibility also involves doing the punishment. If you are willing to take the punishment, it makes more of a statement.
My question was on how to feed this thing, both with memory and I/O. You can add RAM to eliminate disk I/O, but even the RAM is slow compared to the processor core.
What does a Xeon have - 4MB of L2 cache running at processor core speed? What's that equivalent to - a 4 minute MP3 at a low bit rate? What about the player and the OS? It's getting kind of tight in the cache.
On a big machine maybe I can have 8MB or 16MB of L2 cache. Ok, now that is enough to fit a good size scanned picture. That pales in comparison to the index of a good sized DB. If a server can have a main memory of 40GB or more, then what's a 16MB cache? A drop...
MS Outlook shouldn't need 1 Ghz processor. Win2K shouldn't either. Enough said.
Scientific problems where the data sets are small. I can't think of any. Most problems worth solving have huge datasets.
Image processing - a reasonable static image can overwhelm a modern cache easily. Video is even more rediculous. At least with scientific programming and image processing you can try to predict the data you'll need next and prefetch.
This is like dropping a great engine in a crappy car. It will idle really nicely, but the suspension and the steering won't let you go fast around corners. I don't care what the redline of the engine is - I want to know how fast the vehicle goes. Processors are far ahead of memory and I/O technology - you'd have a much more useful system if this was balanced better.
We're not even going to touch on how crappy the state of software is compared to the hardware...
I was thinking more along the lines of using a small portion of the cache repeatedly, so as to not pollute vast portions of the cache.
Does [1] fit the bill? Does "anywhere" mean to any level of the cache hierarchy, or does it give you a true way to limit cache pollution? Lots of architectures can do prefetches. I guess you'd have to work with a very small portion of the address range to keep from filling the whole cache.
[2] Lots of architectures can turn caching off on a page by page basis. This is usable.
I congratulate you on your tuning, but my point remains the same - this thing is hard to feed.
Also, the value of pre-fetches diminishes as the latency to memory becomes greater. Quick example - if a cache miss results in a reference to main memory taking 100 to 150 cycles to complete. Do you know 100 cycles in advance what data you are going to touch? We are not far away from being in this predicament.
Too many people forget that all CPUs wait at the same speed. A 1Ghz anything is a waste considering the state of I/O and memory technology.
This CPU is going to spend a lot of time waiting for memory, even with a generous cache. How many programmers design their data structures to be cache friendly?
With all of the processing of multi-media data types (music, video, and pictures), there isn't a cache big enough to contain the data. Also, the temporal and spatial locality of these data types stink - you process a few pixels, and move on. You don't get to revisit a certain pixel very often. Yet it is wasting space in the cache.
Intel and other manufactures would do much better to add some architectural improvements designed to help multi-media, which is much of what people do with these chips now. How about a section of "streaming cache" for data that will pass through, but only once? That way you don't have to fill the entire cache with useless bulk data.
Or how about I/O model improvements - split the bulk data from the signal and control data so that the bulk data doesn't have to go through the memory hierarchy and the processor at all? If I'm playing a video file, why should the cache and processor be deluged with data being routed to the sound card and the video card? Put the signal and control data out of band from the bulk data so that the processor doesn't have to sift through the bulk data.
8086: 16 bit processor (debatable), 16 bit path. 8088: 16 bit processor (debatable), crippled 8 bit path for IBM. 80286: Real mode enhancements to 8088. PCAT 80386: 32 Bit. Virtual Processor enhancements to 80286. First used by Compaq. 80386sx: 32 bit, 16 bit data path. Cheap. 80486: Cache. 80486sx: Found a use for chips with bad floating point units. Pentium: Dual Pipes. Lots of heat. Math errors. Pentium Pro: RISC core, X86 instructions Pentium II: Consumer Pentium Pro; cheaper package Celery: Crippled Pentium II Celery A series: Crippled Pentium II with cache Pentium III: Since when does adding a few instructions make a new product? Pentium 4: Since when does changing the manufacturing technique make a new product?
Intel has never been able to name a processor. The only trend here is that the next generation offers less and less. Pentium III is a Pentium II with some new instructions - big deal. Pentium 4 is even less of a big deal.
Forget the annoucements and the color schemes! Let's see some innovation!
Slashdot Mirror
Agreed. You know what's more efficient for many tasks? SMP!
Instead of one cutting edge chip pushing really fast (and being expensive, hot, etc.), most systems would do fine with two or more slower CPUs. The slower (older) CPUs are cheap and stable because they bleeding edge moved away from them.
People need to think about absolute speed vs. throughput. For a game you probably want absolute speed. But for word processing, web browsing, playing MP3s, and photo editing all at the same time, a SMP system composed of slower processors is a very good solution.
Operating systems and apps will have to improve to make this work. Windows 95/98/ME don't do SMP, and that is where most home users are. Running many independent apps (mp3 player, netscape, etc) concurrently will work fine, but Photoshop is one of the few examples of an app that will partition its work and use multiple CPUs well.
First off, I'm sorry, but there probably isn't a cool selection of translucent cases to choose from. They mention using eight i-mac logic boards, not eight entire i-macs.
Details are a little short - the article says that it's all in the same box, which implies that there is a high-speed interconnect. I hope that isn't the ethernet port on each of the i-macs.
What's so special? They're doing parallel programming against a few (almost) machines? Parallel programming types have been doing this for years, and there are highly developed libraries to help synchronize and send data to independent machines across a network which are participating in a parallel computation.
This is going to start a craze of little kids gutting their i-macs to make a super computer with all of the other little kids in the neighborhood. Be the first on your block!
Who the he** cares?!?
They are just trying to make sure that people who participate in infringing behavior are held liable, not the technology.
Peer-to-Peer is not illegal; it has been around for a while. Directories are not illegal either. Ripping people off is illegal.
Whether Napster is facilitating the rip-off or whether people are making "fair use" of the music is an entirely different issue. That is what these groups are trying to make clear.
If I write a Logo interpreter in one of the seven languages, and then display the output on the screen as well as send the commands to the Mindstorm, then I can re-create a classic kid's learning environment from 1985 ! Except now my turtle can bite.
Some not-so-random thoughts:
... is that ethical? Is it ethical to complain when people start giving this music away, music that can't be bought?
... how else can I do this? My local library isn't a source.
... I think the RIAA has been the bigger leach for the longer period of time.
[1] The RIAA had no problem selling me the Rolling Stones in record, cassette, and CD format. Why is it that every time the format changes, I get to pay the artist (and the RIAA) again? The was especially annoying when the media suffered from wear, like cassettes and vinyl records did. CDs are not immune either - give it a good scratch, and you get to pay again.
[2] What happened to singles? CD singles existed for a brief period. I hate paying $18 for a CD that I only want one or two songs from, and most of the CDs in my collection are like that. Seems like gouging to not make those songs available in single format, and to charge less for it.
[3] MP3s are generally 128bps, which is pretty low quality. It is slightly better than radio broadcast quality. Why get upset about MP3s available electronically when people can still tape off the air?
[4] How does one obtain (or even purchase) out-of-print music? Try to find "The Screaming Blue Messiahs" on CD - I dare you. (There is a record store with some of the out-of-print vinyl still available.) The major labels have chosen to make music unavailable
[5] What about the people experimenting? I'm discovering Jazz and Swing for the first time
[6] Speaking of which, are libraries the next target? They loan out CDs which can be copied.
[7] Napster would be a lot more defensible if they weren't a commercial enterprise. There are a lot of legitimate uses for it, but we know that most people are leaching. The RIAA can be evil, but some artists insist on being compensated and that is their right. And supposedly, the RIAA is defending that right. Too bad
>> or sitting at the front of the bus.
Or refusing to obey an unjust order.
It is a civic responsibility to disobey unjust laws. ("Civil disobedience.) Part of the reponsibility also involves doing the punishment. If you are willing to take the punishment, it makes more of a statement.
Hi. I'm the nimrod. My name is Mike.
...
...
My question was on how to feed this thing, both with memory and I/O. You can add RAM to eliminate disk I/O, but even the RAM is slow compared to the processor core.
What does a Xeon have - 4MB of L2 cache running at processor core speed? What's that equivalent to - a 4 minute MP3 at a low bit rate? What about the player and the OS? It's getting kind of tight in the cache.
On a big machine maybe I can have 8MB or 16MB of L2 cache. Ok, now that is enough to fit a good size scanned picture. That pales in comparison to the index of a good sized DB. If a server can have a main memory of 40GB or more, then what's a 16MB cache? A drop
MS Outlook shouldn't need 1 Ghz processor. Win2K shouldn't either. Enough said.
Scientific problems where the data sets are small. I can't think of any. Most problems worth solving have huge datasets.
Image processing - a reasonable static image can overwhelm a modern cache easily. Video is even more rediculous. At least with scientific programming and image processing you can try to predict the data you'll need next and prefetch.
This is like dropping a great engine in a crappy car. It will idle really nicely, but the suspension and the steering won't let you go fast around corners. I don't care what the redline of the engine is - I want to know how fast the vehicle goes. Processors are far ahead of memory and I/O technology - you'd have a much more useful system if this was balanced better.
We're not even going to touch on how crappy the state of software is compared to the hardware
I was thinking more along the lines of using a small portion of the cache repeatedly, so as to not pollute vast portions of the cache.
Does [1] fit the bill? Does "anywhere" mean to any level of the cache hierarchy, or does it give you a true way to limit cache pollution? Lots of architectures can do prefetches. I guess you'd have to work with a very small portion of the address range to keep from filling the whole cache.
[2] Lots of architectures can turn caching off on a page by page basis. This is usable.
I congratulate you on your tuning, but my point remains the same - this thing is hard to feed.
Also, the value of pre-fetches diminishes as the latency to memory becomes greater. Quick example - if a cache miss results in a reference to main memory taking 100 to 150 cycles to complete. Do you know 100 cycles in advance what data you are going to touch? We are not far away from being in this predicament.
These things need work.
Too many people forget that all CPUs wait at the same speed. A 1Ghz anything is a waste considering the state of I/O and memory technology.
This CPU is going to spend a lot of time waiting for memory, even with a generous cache. How many programmers design their data structures to be cache friendly?
With all of the processing of multi-media data types (music, video, and pictures), there isn't a cache big enough to contain the data. Also, the temporal and spatial locality of these data types stink - you process a few pixels, and move on. You don't get to revisit a certain pixel very often. Yet it is wasting space in the cache.
Intel and other manufactures would do much better to add some architectural improvements designed to help multi-media, which is much of what people do with these chips now. How about a section of "streaming cache" for data that will pass through, but only once? That way you don't have to fill the entire cache with useless bulk data.
Or how about I/O model improvements - split the bulk data from the signal and control data so that the bulk data doesn't have to go through the memory hierarchy and the processor at all? If I'm playing a video file, why should the cache and processor be deluged with data being routed to the sound card and the video card? Put the signal and control data out of band from the bulk data so that the processor doesn't have to sift through the bulk data.
I'm getting too old ...
8086: 16 bit processor (debatable), 16 bit path.
8088: 16 bit processor (debatable), crippled 8 bit path for IBM.
80286: Real mode enhancements to 8088. PCAT
80386: 32 Bit. Virtual Processor enhancements to 80286. First used by Compaq.
80386sx: 32 bit, 16 bit data path. Cheap.
80486: Cache.
80486sx: Found a use for chips with bad floating point units.
Pentium: Dual Pipes. Lots of heat. Math errors.
Pentium Pro: RISC core, X86 instructions
Pentium II: Consumer Pentium Pro; cheaper package
Celery: Crippled Pentium II
Celery A series: Crippled Pentium II with cache
Pentium III: Since when does adding a few instructions make a new product?
Pentium 4: Since when does changing the manufacturing technique make a new product?
Intel has never been able to name a processor. The only trend here is that the next generation offers less and less. Pentium III is a Pentium II with some new instructions - big deal. Pentium 4 is even less of a big deal.
Forget the annoucements and the color schemes! Let's see some innovation!