Probably the same way Intel takes it every time MS fucks them over, ie bending over and grabbing their ankles while praying that Linux takes off.
Seriously, there is no love between Intel and Microsoft. They both want control of the PC business. Microsoft has it, Intel doesn't, but despite how much Intel may dislike this, they know that they need MS more than MS needs them.
At one point in time MS may have needed to make deals to keep Intel happy, but now it's definitely the other way around.
Probably not much will happen othar than that it will run. Windows was designed to be ported to different platforms fairly easily. It has run on PowerPC before (and MIPS and Alpha), and it can be made to run on that platform again.
Hmm, that's odd, I've never had it recreate the registry key after I deleted it.
However, that being said, iTunes leaves far too much clutter around. Not only does it automatically start QuickTime on bootup, it also starts up an iPod executable (even though i don't own an iPod) as well as GEARSecurity CD-RW service (even though I don't currently have a CD-RW drive). If you stop the CD service then iTunes complains every time you boot it up.
All of these can be disabled if you know how, but they are things that shouldn't be installed in the first place.
I've been exploring a few of the music download services (iTunes, Napster2, MusicMatch and Puretracks), and I would definitely NOT say that iTunes did it "right". What they have done is a less shitty job than the rest of the pack, but it's still VERY weak.
First off, the selection is terrible as compared to what a 2-bit CD store in some dank corner has to offer. It was at best a 50/50 shot if bands I was looking for were there, and I was looking for fairly well known bands signed to major record deals. Even if they were there, usually there were only 1 or 2 albums, or in some cases just a few songs from their albums. FWIW MusicMatch was the service that I found which had the best selection (albeit not by too much), but their interface is absolutely terrible.
Second, it's only available in the US. Ok, that may be the biggest market and they are working to get into other markets, but until they do...
Their interface also needs a LOT of work. Half of the time it suffers from the Apple "I have a paw and not a hand" one-button mouse syndrom. If you right-click on most things, it brings up a menu that gives you the option of "Eject Disc" and nothing else. Great, real useful.
You are VERY restricted in how and where you can play music as compared to CDs. Sure, the restrictions are less than other services, but they are still pretty restrictive. For example, if I boot into Linux to do some work, about the only way I could play their songs is to burn them to CD and possibly re-rip them (thereby losing audio quality).
Their pricing scheme could use a bit of work. Sometimes you can find them charging the usual $9.99 for a full CD of 10 (or fewer?) songs, when buying the individual songs would be cheaper. They also charge the regular $0.99 for "songs" that are just a few seconds long intros.
Now I'm all for music downloads, I see great potential in the medium, but at the moment I see that potential being BADLY underutilized. I could even put up with most of the flaws if I could: a.) buy the music in the first place (I don't live in the US and therefore can't buy from most services yet) and b.) they had GOOD selection. In particular services like iTunes could potentially be a GREAT way to get rare CDs and imports which are often hard to track down in conventional stores. However, for right now, unless your looking for top-40 material, you probably won't find it.
As I said, iTunes did not do it "right". The potential is there, but they have a LONG way to go.
Red Storm, the machine by itself itself, uses 2.0MW
To quote page 6 of the Red Storm PDF:
"Less than 2 MW total power and cooling"
Sounds to me like they're counting more than just the machine, no?
As for the PPC970 chip itself, my understanding is that the 48W number is it's *typical* power consumption rather than it's maximum power consumption or thermal design power. These days there are a LOT of different methods of measuring power consumption being thrown around. Quite respectible for a chip of the PPC970's level of performance, but not the sort of night-and-day difference vs. some competing chips that many make it out to be (ie typical power consumption of a 3.2GHz P4 is only around 60W, much lower than the 82W TDP that is often quoted in comparison).
Err, I believe you're describing the EXACT thing I described in the initial message! This will reduce your memory errors, perhaps sufficiently that you can ignore them, but it won't make them disapear (think memory error before you get a chance to do the coding, than all your code does is ensure that you have the same bad data that you had earlier). It could be that the performance hit you take from this is not too sufficient and the level of protection you get is "good enough", but I haven't seen any remotely technical info about this.
As for my "fucking modem" (is that a Winmodem? I've always prefered full hardware, external modems:> ), it too will only correct errors if the data you start with is good. If you get a data error going over the serial cable, the error correction on the modem does dick-all to fix it, it'll just make sure that the same incorrect data arrives at the far end. However, the chance of getting a data error on the serial cable (or PCI bus for an internal modem) is sufficient low that this isn't really an issue, besides which you have end-to-end error correction at another level (eg PPP and TCP, or zmodem for the old-timers).
The only way I'm aware of to be really sure of not getting errors is to double up on the calculations, which cuts your performance in half. There are some tricks you can do that will cut the rate of errors down, perhaps sufficiently that you can ignore the possibility (after all, ECC is by no means 100% infalliable either) with less of a performance hit, but I certainly haven't seen any technical details of it.
If they're doing that, why not just use ECC and double (or better) their performance? Supercomputing applications are often limited by memory bandwidth much more than by CPU speed, so by reading every bit of data three times, you're cutting your performance in half at best, and possibly much worse (especially if you end up taking a latency hit on top of the bandwidth hit).
According to the other post, they simply aren't worrying about memory errors at this time, ie they are "fixing" the problem by ignoring it. I guess that's what solution:>
The efficiency is quite poor for this machine, at least as far as efficiency is termed for supercomputers. The cluster has a theoretical peak of 17.6TFlops/s if I did my math right (8GFlops/s per processor), but they are only turning in an actual score of 9.56TFlops/s, for an efficiency of only 54%. Even if they boost performance by 10%, they'll still only be ~60% efficient.
For comparison, ASCI Q (#2 on Top500) reaches 68% efficiency, MCR Linux Cluster (currently #3, but to be pushed by by this new Mac cluster) reaches 69% efficiency, and the #1 spot, Earth Simulator, reaches a quite impressive 88% efficinecy.
Of course, there are other ways to measure efficinecy. When it comes to performance/price, this Mac cluster does very well, even if you do take into account the real costs (ie MUCH more than just the $5.2 million up front cost). For cost/power consumption it seems reasonable, but not outstanding. 10TFlops/1.5MW of power is ok, and not too far off the Earth Simulator's 35TFlops/3.5MW of power, but it's certainly nothing to write home about. Cray's next big cluster, Red Storm, is likely to get over 30TFlops when it's released, but will consume only 2.0MW of power.
For anyone interesting in learning a bit more about what some of the issues are when creating a super-computer, you might want to have a look at the following:
The article is talking about Cray/Sandia's new Red Storm machine, a supercomputer using over 10,000 AMD Opteron processors that is expected to be competitive with the Earth Simulator for the #1 spot on the Top500 list. It does, however, talk about a lot more than just the specifics of this cluster, describing what some of the bottlenecks in supercomputers are and how to avoid/work around them.
I keep seeing reference to some sort of software that will defeat hardware memory errors.
How, pray tell, are they planning on detecting these errors? I can understand how you could reduce the frequency of errors with only a slight loss in performance, ie take some sort of checksum of your data after every x number of cycles, but that doesn't eliminate the errors, only reduces their frequency. Maybe it reduces the frequency by enough that you don't need to worry about it, especially if 'x' is a sufficiently small number, but it still seems like a pretty risky prospect to me.
Anyone seen any actual TECHNICAL details on this point, ie not just some Mac fan yelling "Deja Vu, DEJA VU!!!"?
Just FWIW, they are claiming power usage of 1.5MW for this cluster of 2200 processors. Cray just released the numbers for their upcoming Red Storm cluster with over 10,000 AMD Opteron processors, just slightly less than 2.0MW.
Long story short, this Big Mac cluster consumes a LOT of power. To be fair though, the Earth Simulator apparently uses around 3.5MW of power, so on a power/performance comparison, the Big Mac and Earth Simulator are roughly on-par with one another.
Wouldn't that mean that SCO would then be in direct violation of said license, since they have (and continue to) distributed code that is covered by the GPL. Therefore they would not longer be able to distribute Unixware with ANY applications covered by the GPL.
Yet again, SCO wants to have their cake and eat it to.
Note that the only reason why it's listed as worse than the largest US SUV's is because the largest US SUV's don't list their fuel economy. You'll notice that neither the Ford Excursion or the large-engine Chevy Suburbans are listed. The Excusion definitely gets worse fuel economy, roughly 11/13mpg. Also, the Cadilac Escalades is tied with the Range Rover for worst fuel economy of those listed, and the Dodge Durango is only one notch better (12/17mpg).
A bit more info on differences in octane rating. My understanding is that in North America, we tend to use an "Anti-Knock Index", which is made up of the RON+MON/2, where:
RON = Research Octane Number, a kind of best-case number MON = Motor Octane Number, a tougher test measured at high engine speed and temps.
In Europe (and Japan), my understanding is that they just use the RON number. The end result is that the numbers are not directly comperable, though US gas with an 87 AKI Octane index is roughly the same as European gas with a 91 or 92 RON octane rating.
You are quite correct though that European cars do tend to use higher octane levels still though. Most cars in Europe use a 93 or 95 RON octane level, which is going to be roughly equivilent to using "Premium" (usually 89 AKI octane) or "Super" (usually 91 AKI octane) gas in North America. In addition to the higher compression ratios, European cars also tend to burn gas less cleanly since, until the past year or two, Europe was quite a ways behind North America in terms of emission standards. The European Union has just started phasing in new emissions regulations that bring it up to (and in a few cases beyond) North American standards.
When you get right down to it though, you can easily see that the cost of gas has a very direct impact on the average fuel economy of the cars that tend to be sold in the country. That is why in Canada we have always tended to drive more fuel efficient cars than the US, even though models available are essentially identical between the two countries (gas in Canada is more expensive than the US, but not as much as in Europe). On the same front, European cars tend to be more fuel efficient still.
Actually, your car is only doing 26.3 miles to the US gallon (which, presumably was the type of "gallon" that they were talking about). Not at all far off the 25mpg figure quoted. Also, European cars tend to have better fuel economy but burn gas less cleanly than North American cars. The EU is only just now catching up with North America in terms of emissions requirements.
I doubt that Sun has any long-term contracts that extend beyond producing the UltraSparc IV, if even that. If this Sun/Fujitsu partenership does pan out, it probably won't be for another couple of years at least. Sun is still fairly early in their design of the Ultrasparc V, so perhaps it will be around that time frame.
Besides, TI hasn't exactly been a stelar partener for Sun. They are definitely NOT one of the big three manufacturer's of high-end chips (Big-three being Intel, AMD and IBM, ie the only three companies that are really manufacturing top-end performance processors). They were late to the 130nm node, they'll probably be late to the 90nm node and they aren't ramping clock speeds at nearly the pace of the others. At best TI is just keeping pace with the low-cost manufacturers like TSMC and UMC (if they're even managing that), and I'd wager a guess that Sun is paying a fair bit more to get their chips made by TI than what the Taiwanese would charge.
In any case, this merger of technologies makes GREAT sense. Neither Sun or Fujitsu make nearly enough money on hardware sales to justify continous development of new, high-end processors. Developing new, top-end performance processors costs a LOT of money, which is why there are now only 5 companies left doing so including Sun and Fujitsu (again, Intel, IBM and AMD are the other three). By combining together, they might just have enough sales to make it worthwhile though. If not, Sun is going to have to majorly change the way they do business.
Err, Apple's G5 and the AMD Opteron don't have an even remotely related memory setup. The G5 looks a lot more like the AthlonXP and AthlonMP setups. The Opteron has an integrated 128-bit wide DDR memory controller, connects multiple CPUs directly through cache-coherent Hyptertransport links, and uses additional 32-bit, 1600MT/s HT links (3.2GB/s in each direction) to connect the CPU directly to the I/O chips.
The Powermac G5 uses up to 1GT/s, 64-bit wide version of IBM's Elastic I/O bus to connect each processor to a memory controller chip, which in turn has a pair of 64-bit wide DDR memory controllers. These buses are also shared for the processors I/O needs, which are passed over a 800MT/s, 16-bit wide hypertransport link to the PCI-X controller.
As for the width and speed of the Hypertransport links, Apple is very confusing on this front. In the document you linked they say "two bidirectional 16-bit, 800MHz HyperTransport interconnects for a maximum throughput of 3.2GB per second." In their PowerMac G5 Tech Specs PDF they say "two bidirectional 800MHz HyperTransport interconnects for a maximum throughput of 1.6 GBps." So which is it? And just what bandwidth are they measuring?
The PowerMac does indeed have two separate bi-directional Hypertransport links, the first connects the memory and processor controller chip to the PCI-X controller, and the second goes from the PCI-X controller to the extra I/O chips. It seems to me like the page you quoted is ADDING the bandwidth of the two daisy-chained hypertransport links, which would be TOTALLY incorrect.
My numbers came from the fact that a 16-bit (8-bits per direction) 800MT/s hypertransport link gets you only 800MB/s in each direction. Of course, it could really indeed be a "800MHz" hypertransport link, ie a 1600MT/s link since Hypertransport is a DDR protocol, but I highly doubt that since every other specification they mention just doubles the "MHz" number anytime they encounter a DDR bus (not that Apple is the only one to do this, Intel's "800MHz" bus runs at either 200MHz or 400MHz, depending on which clock you look at).
There are a handful of AMD Athlon clusters in the Top500 list. The fastest that I'm aware of uses 512 Athlon (T-Birds I believe) clocked at 1.4GHz. It comes in at 825GFlops, or roughly 1/10th of what this Mac cluster is managing. This system is sitting at #80 in the Top500 list for July. In total there are 11 AMD Athlon clusters in the Top500 list, putting it just behind the Intel Itanium (13 systems) but ahead of Sun Ultrasparcs (9 systems).
There are also some new Opteron clusters in the works, including one that is expected to match the performance of the current #1 (Japan's Earth Simulator) at 40TFlops using ~12,000 Opterons. Another 3 Opteron clusters in the works would likely make it into the top 10 of the current list if they were ready today.
You can't reliably detect memory errors through software (short of doing every calculation twice). VT is going to have to live with not being able to trust their data. At best they might be able to reduce the frequency of software errors biting them sufficiently that such an occurance is fairly rare (once a month? once a year?). Certainly better than what they would get out of the box, but it will hurt performance.
I certainly hope that IBM has some plans for HPC clusters with the PowerPC970. The chip itself is showing some GREAT potential for this market. The only problem at the moment is that it's being sadled by a piss-poor design for supercomputers. Now, I don't mean any offense against the PowerMac G5 as a desktop computer, but come on, only 2 processors in a 51x21x48cm enclosure?! That's ridiculous! Add in the lack of ECC memory and the relatively low bandwidth and potentially high latency connection between the CPU and the networking hardware (going through two controller chips and an 800MB/s (each direction) hypertransport link), and you've got a design that doesn't really cut it for a supercomputer. That's without mentioning the fact that the systems come with DVD writers and high-end gaming video cards, both of which are REALLY useless for the application and just wasting power and space.
Now, when IBM releases their PowerPC 970 servers, I would expect that they do things right. At the very least they should support ECC memory, which IMO is a MAJOR oversight on this Big Mac cluster (given the 4.4TB of memory in this cluster, I would expect at least one soft memory error every day or two).
Actually I don't think so, at least not for most supercomputer tasks. Most things done by a supercomputer are easily vectorized, and here the G5's Altivec unit should make it faster than the Opteron using SSE2. The IBM PowerPC 970 itself is a very nice little processor for supercomputers, though the PowerMac G5 desktop computers are a rather poor choice in systems to utilize that processor. If/when IBM releases their PPC970 servers, they could make for some EXCELLENT HPC clusters.
However, the Opterons could have a major advantage over the Macs and Intel chips in it's I/O, and that is rather critical for this sort of clustered supercomputer. The Opteron has a high-bandwidth, low-latency chip to chip interconnect RIGHT on the processor die. With Hypertransport, you could theoretically hang a high-bandwidth netwrok connection right off the processor without having to worry about going through any other controllers. This gives you much lower latency than essentially ANY other processor in the world now, and higher bandwidth than just about anything else out there as well (though bandwidth is always easier to solve than latency, just throw more money at it and you can get more bandwidth).
Given this fact, it's not altogether surprising that the Opteron IS being used to make some HPC cluster. In fact, it's being used for at least 5 major new clusters, at least 4 of which would make the current Top500 list if they were available now. The largest Opteron cluster, being built by Cray for Sandie will feature ~12,000 Opteron processors and will roughly match the performance of the current #1, Japan's Earth Simulator.
No they could not bond NICs, because they're using Infiniband and not ethernet. Besides, I think that they are being limited more by latency than bandwidth, so therefore adding bandwidth isn't going to help much. What's worse, their bandwidth limit is being reached inside the computer, with their chip to chip interconnect having less bandwidth than their computer to computer interconnect.
This is not altogether surprising, given that they are using a desktop computer and trying to shoehorn it into a supercomputer role. They are bound to run into some limitations.
Communication latencies are definitely going to be their major problem, it always is in a super-computer cluster. However, that being said, I'm not sure that they've got the bandwidth problem licked either. Yes,infiniband provides tons of bandwidth, but I'm not sure that the system has the internal bandwidth to make use of it.
They are connecting the Inifiband cards using 64-bit PCI-X, which allows for up to 2.1GB/s of bandwidth. However, the PCI-X controller is only connected to the memory and processor controller chip through an 800MB/s hypertransport link (err,two links, 800MB/s in either direction). This is actually less than the 10Gb/s (1.25GB/s) of bandwidth that infiniband provides. What's even worse, that 800MB/s hypertransport link is shared with basically all I/O other than memory and graphics, and it's also a peak figure, with real-world bandwidth being somewhat lower.
In short, internal system bandwidth could be a problem here. Likely they are also running into some internal system latency issues as well. To get to their infiniband cards, they first have to go from the processor though the Elastic I/O channel to the memory+processor controller chip, than through Hypertransport to the PCI-X, than over the PCI-X bus to the Infiniband card. Each step of the way is adding a bit of latency.
They would have been much better off with something like how the Opteron handles things, with the PCI-X controller hanging right off the processor itself, connected via a 3.2GB/s (each direction) hypertransport link. More bandwidth and fewer chips and buses to go through.
Slashdot Mirror
I don't think that played even a remote part of the decision.
However, keeping software pirates out probably did play a part in this.
Probably the same way Intel takes it every time MS fucks them over, ie bending over and grabbing their ankles while praying that Linux takes off.
Seriously, there is no love between Intel and Microsoft. They both want control of the PC business. Microsoft has it, Intel doesn't, but despite how much Intel may dislike this, they know that they need MS more than MS needs them.
At one point in time MS may have needed to make deals to keep Intel happy, but now it's definitely the other way around.
Probably not much will happen othar than that it will run. Windows was designed to be ported to different platforms fairly easily. It has run on PowerPC before (and MIPS and Alpha), and it can be made to run on that platform again.
Hmm, that's odd, I've never had it recreate the registry key after I deleted it.
However, that being said, iTunes leaves far too much clutter around. Not only does it automatically start QuickTime on bootup, it also starts up an iPod executable (even though i don't own an iPod) as well as GEARSecurity CD-RW service (even though I don't currently have a CD-RW drive). If you stop the CD service then iTunes complains every time you boot it up.
All of these can be disabled if you know how, but they are things that shouldn't be installed in the first place.
I've been exploring a few of the music download services (iTunes, Napster2, MusicMatch and Puretracks), and I would definitely NOT say that iTunes did it "right". What they have done is a less shitty job than the rest of the pack, but it's still VERY weak.
First off, the selection is terrible as compared to what a 2-bit CD store in some dank corner has to offer. It was at best a 50/50 shot if bands I was looking for were there, and I was looking for fairly well known bands signed to major record deals. Even if they were there, usually there were only 1 or 2 albums, or in some cases just a few songs from their albums. FWIW MusicMatch was the service that I found which had the best selection (albeit not by too much), but their interface is absolutely terrible.
Second, it's only available in the US. Ok, that may be the biggest market and they are working to get into other markets, but until they do...
Their interface also needs a LOT of work. Half of the time it suffers from the Apple "I have a paw and not a hand" one-button mouse syndrom. If you right-click on most things, it brings up a menu that gives you the option of "Eject Disc" and nothing else. Great, real useful.
You are VERY restricted in how and where you can play music as compared to CDs. Sure, the restrictions are less than other services, but they are still pretty restrictive. For example, if I boot into Linux to do some work, about the only way I could play their songs is to burn them to CD and possibly re-rip them (thereby losing audio quality).
Their pricing scheme could use a bit of work. Sometimes you can find them charging the usual $9.99 for a full CD of 10 (or fewer?) songs, when buying the individual songs would be cheaper. They also charge the regular $0.99 for "songs" that are just a few seconds long intros.
Now I'm all for music downloads, I see great potential in the medium, but at the moment I see that potential being BADLY underutilized. I could even put up with most of the flaws if I could: a.) buy the music in the first place (I don't live in the US and therefore can't buy from most services yet) and b.) they had GOOD selection. In particular services like iTunes could potentially be a GREAT way to get rare CDs and imports which are often hard to track down in conventional stores. However, for right now, unless your looking for top-40 material, you probably won't find it.
As I said, iTunes did not do it "right". The potential is there, but they have a LONG way to go.
Red Storm, the machine by itself itself, uses 2.0MW
To quote page 6 of the Red Storm PDF:
"Less than 2 MW total power and cooling"
Sounds to me like they're counting more than just the machine, no?
As for the PPC970 chip itself, my understanding is that the 48W number is it's *typical* power consumption rather than it's maximum power consumption or thermal design power. These days there are a LOT of different methods of measuring power consumption being thrown around. Quite respectible for a chip of the PPC970's level of performance, but not the sort of night-and-day difference vs. some competing chips that many make it out to be (ie typical power consumption of a 3.2GHz P4 is only around 60W, much lower than the 82W TDP that is often quoted in comparison).
Err, I believe you're describing the EXACT thing I described in the initial message! This will reduce your memory errors, perhaps sufficiently that you can ignore them, but it won't make them disapear (think memory error before you get a chance to do the coding, than all your code does is ensure that you have the same bad data that you had earlier). It could be that the performance hit you take from this is not too sufficient and the level of protection you get is "good enough", but I haven't seen any remotely technical info about this.
:> ), it too will only correct errors if the data you start with is good. If you get a data error going over the serial cable, the error correction on the modem does dick-all to fix it, it'll just make sure that the same incorrect data arrives at the far end. However, the chance of getting a data error on the serial cable (or PCI bus for an internal modem) is sufficient low that this isn't really an issue, besides which you have end-to-end error correction at another level (eg PPP and TCP, or zmodem for the old-timers).
As for my "fucking modem" (is that a Winmodem? I've always prefered full hardware, external modems
The only way I'm aware of to be really sure of not getting errors is to double up on the calculations, which cuts your performance in half. There are some tricks you can do that will cut the rate of errors down, perhaps sufficiently that you can ignore the possibility (after all, ECC is by no means 100% infalliable either) with less of a performance hit, but I certainly haven't seen any technical details of it.
If they're doing that, why not just use ECC and double (or better) their performance? Supercomputing applications are often limited by memory bandwidth much more than by CPU speed, so by reading every bit of data three times, you're cutting your performance in half at best, and possibly much worse (especially if you end up taking a latency hit on top of the bandwidth hit).
:>
According to the other post, they simply aren't worrying about memory errors at this time, ie they are "fixing" the problem by ignoring it. I guess that's what solution
The efficiency is quite poor for this machine, at least as far as efficiency is termed for supercomputers. The cluster has a theoretical peak of 17.6TFlops/s if I did my math right (8GFlops/s per processor), but they are only turning in an actual score of 9.56TFlops/s, for an efficiency of only 54%. Even if they boost performance by 10%, they'll still only be ~60% efficient.
For comparison, ASCI Q (#2 on Top500) reaches 68% efficiency, MCR Linux Cluster (currently #3, but to be pushed by by this new Mac cluster) reaches 69% efficiency, and the #1 spot, Earth Simulator, reaches a quite impressive 88% efficinecy.
Of course, there are other ways to measure efficinecy. When it comes to performance/price, this Mac cluster does very well, even if you do take into account the real costs (ie MUCH more than just the $5.2 million up front cost). For cost/power consumption it seems reasonable, but not outstanding. 10TFlops/1.5MW of power is ok, and not too far off the Earth Simulator's 35TFlops/3.5MW of power, but it's certainly nothing to write home about. Cray's next big cluster, Red Storm, is likely to get over 30TFlops when it's released, but will consume only 2.0MW of power.
For anyone interesting in learning a bit more about what some of the issues are when creating a super-computer, you might want to have a look at the following:
Red Storm PDF
The article is talking about Cray/Sandia's new Red Storm machine, a supercomputer using over 10,000 AMD Opteron processors that is expected to be competitive with the Earth Simulator for the #1 spot on the Top500 list. It does, however, talk about a lot more than just the specifics of this cluster, describing what some of the bottlenecks in supercomputers are and how to avoid/work around them.
I keep seeing reference to some sort of software that will defeat hardware memory errors.
How, pray tell, are they planning on detecting these errors? I can understand how you could reduce the frequency of errors with only a slight loss in performance, ie take some sort of checksum of your data after every x number of cycles, but that doesn't eliminate the errors, only reduces their frequency. Maybe it reduces the frequency by enough that you don't need to worry about it, especially if 'x' is a sufficiently small number, but it still seems like a pretty risky prospect to me.
Anyone seen any actual TECHNICAL details on this point, ie not just some Mac fan yelling "Deja Vu, DEJA VU!!!"?
Just FWIW, they are claiming power usage of 1.5MW for this cluster of 2200 processors. Cray just released the numbers for their upcoming Red Storm cluster with over 10,000 AMD Opteron processors, just slightly less than 2.0MW.
Long story short, this Big Mac cluster consumes a LOT of power. To be fair though, the Earth Simulator apparently uses around 3.5MW of power, so on a power/performance comparison, the Big Mac and Earth Simulator are roughly on-par with one another.
Wouldn't that mean that SCO would then be in direct violation of said license, since they have (and continue to) distributed code that is covered by the GPL. Therefore they would not longer be able to distribute Unixware with ANY applications covered by the GPL.
Yet again, SCO wants to have their cake and eat it to.
Note that the only reason why it's listed as worse than the largest US SUV's is because the largest US SUV's don't list their fuel economy. You'll notice that neither the Ford Excursion or the large-engine Chevy Suburbans are listed. The Excusion definitely gets worse fuel economy, roughly 11/13mpg. Also, the Cadilac Escalades is tied with the Range Rover for worst fuel economy of those listed, and the Dodge Durango is only one notch better (12/17mpg).
A bit more info on differences in octane rating. My understanding is that in North America, we tend to use an "Anti-Knock Index", which is made up of the RON+MON/2, where:
RON = Research Octane Number, a kind of best-case number
MON = Motor Octane Number, a tougher test measured at high engine speed and temps.
In Europe (and Japan), my understanding is that they just use the RON number. The end result is that the numbers are not directly comperable, though US gas with an 87 AKI Octane index is roughly the same as European gas with a 91 or 92 RON octane rating.
You are quite correct though that European cars do tend to use higher octane levels still though. Most cars in Europe use a 93 or 95 RON octane level, which is going to be roughly equivilent to using "Premium" (usually 89 AKI octane) or "Super" (usually 91 AKI octane) gas in North America. In addition to the higher compression ratios, European cars also tend to burn gas less cleanly since, until the past year or two, Europe was quite a ways behind North America in terms of emission standards. The European Union has just started phasing in new emissions regulations that bring it up to (and in a few cases beyond) North American standards.
When you get right down to it though, you can easily see that the cost of gas has a very direct impact on the average fuel economy of the cars that tend to be sold in the country. That is why in Canada we have always tended to drive more fuel efficient cars than the US, even though models available are essentially identical between the two countries (gas in Canada is more expensive than the US, but not as much as in Europe). On the same front, European cars tend to be more fuel efficient still.
Actually, your car is only doing 26.3 miles to the US gallon (which, presumably was the type of "gallon" that they were talking about). Not at all far off the 25mpg figure quoted. Also, European cars tend to have better fuel economy but burn gas less cleanly than North American cars. The EU is only just now catching up with North America in terms of emissions requirements.
Hey, Certicom is a Canadian company, so maybe the song isn't out of place here?
I doubt that Sun has any long-term contracts that extend beyond producing the UltraSparc IV, if even that. If this Sun/Fujitsu partenership does pan out, it probably won't be for another couple of years at least. Sun is still fairly early in their design of the Ultrasparc V, so perhaps it will be around that time frame.
Besides, TI hasn't exactly been a stelar partener for Sun. They are definitely NOT one of the big three manufacturer's of high-end chips (Big-three being Intel, AMD and IBM, ie the only three companies that are really manufacturing top-end performance processors). They were late to the 130nm node, they'll probably be late to the 90nm node and they aren't ramping clock speeds at nearly the pace of the others. At best TI is just keeping pace with the low-cost manufacturers like TSMC and UMC (if they're even managing that), and I'd wager a guess that Sun is paying a fair bit more to get their chips made by TI than what the Taiwanese would charge.
In any case, this merger of technologies makes GREAT sense. Neither Sun or Fujitsu make nearly enough money on hardware sales to justify continous development of new, high-end processors. Developing new, top-end performance processors costs a LOT of money, which is why there are now only 5 companies left doing so including Sun and Fujitsu (again, Intel, IBM and AMD are the other three). By combining together, they might just have enough sales to make it worthwhile though. If not, Sun is going to have to majorly change the way they do business.
Err, Apple's G5 and the AMD Opteron don't have an even remotely related memory setup. The G5 looks a lot more like the AthlonXP and AthlonMP setups. The Opteron has an integrated 128-bit wide DDR memory controller, connects multiple CPUs directly through cache-coherent Hyptertransport links, and uses additional 32-bit, 1600MT/s HT links (3.2GB/s in each direction) to connect the CPU directly to the I/O chips.
The Powermac G5 uses up to 1GT/s, 64-bit wide version of IBM's Elastic I/O bus to connect each processor to a memory controller chip, which in turn has a pair of 64-bit wide DDR memory controllers. These buses are also shared for the processors I/O needs, which are passed over a 800MT/s, 16-bit wide hypertransport link to the PCI-X controller.
As for the width and speed of the Hypertransport links, Apple is very confusing on this front. In the document you linked they say "two bidirectional 16-bit, 800MHz HyperTransport interconnects for a maximum throughput of 3.2GB per second." In their PowerMac G5 Tech Specs PDF they say "two bidirectional 800MHz HyperTransport interconnects for a maximum throughput of 1.6 GBps." So which is it? And just what bandwidth are they measuring?
The PowerMac does indeed have two separate bi-directional Hypertransport links, the first connects the memory and processor controller chip to the PCI-X controller, and the second goes from the PCI-X controller to the extra I/O chips. It seems to me like the page you quoted is ADDING the bandwidth of the two daisy-chained hypertransport links, which would be TOTALLY incorrect.
My numbers came from the fact that a 16-bit (8-bits per direction) 800MT/s hypertransport link gets you only 800MB/s in each direction. Of course, it could really indeed be a "800MHz" hypertransport link, ie a 1600MT/s link since Hypertransport is a DDR protocol, but I highly doubt that since every other specification they mention just doubles the "MHz" number anytime they encounter a DDR bus (not that Apple is the only one to do this, Intel's "800MHz" bus runs at either 200MHz or 400MHz, depending on which clock you look at).
There are a handful of AMD Athlon clusters in the Top500 list. The fastest that I'm aware of uses 512 Athlon (T-Birds I believe) clocked at 1.4GHz. It comes in at 825GFlops, or roughly 1/10th of what this Mac cluster is managing. This system is sitting at #80 in the Top500 list for July. In total there are 11 AMD Athlon clusters in the Top500 list, putting it just behind the Intel Itanium (13 systems) but ahead of Sun Ultrasparcs (9 systems).
There are also some new Opteron clusters in the works, including one that is expected to match the performance of the current #1 (Japan's Earth Simulator) at 40TFlops using ~12,000 Opterons. Another 3 Opteron clusters in the works would likely make it into the top 10 of the current list if they were ready today.
You can't reliably detect memory errors through software (short of doing every calculation twice). VT is going to have to live with not being able to trust their data. At best they might be able to reduce the frequency of software errors biting them sufficiently that such an occurance is fairly rare (once a month? once a year?). Certainly better than what they would get out of the box, but it will hurt performance.
I certainly hope that IBM has some plans for HPC clusters with the PowerPC970. The chip itself is showing some GREAT potential for this market. The only problem at the moment is that it's being sadled by a piss-poor design for supercomputers. Now, I don't mean any offense against the PowerMac G5 as a desktop computer, but come on, only 2 processors in a 51x21x48cm enclosure?! That's ridiculous! Add in the lack of ECC memory and the relatively low bandwidth and potentially high latency connection between the CPU and the networking hardware (going through two controller chips and an 800MB/s (each direction) hypertransport link), and you've got a design that doesn't really cut it for a supercomputer. That's without mentioning the fact that the systems come with DVD writers and high-end gaming video cards, both of which are REALLY useless for the application and just wasting power and space.
Now, when IBM releases their PowerPC 970 servers, I would expect that they do things right. At the very least they should support ECC memory, which IMO is a MAJOR oversight on this Big Mac cluster (given the 4.4TB of memory in this cluster, I would expect at least one soft memory error every day or two).
Actually I don't think so, at least not for most supercomputer tasks. Most things done by a supercomputer are easily vectorized, and here the G5's Altivec unit should make it faster than the Opteron using SSE2. The IBM PowerPC 970 itself is a very nice little processor for supercomputers, though the PowerMac G5 desktop computers are a rather poor choice in systems to utilize that processor. If/when IBM releases their PPC970 servers, they could make for some EXCELLENT HPC clusters.
However, the Opterons could have a major advantage over the Macs and Intel chips in it's I/O, and that is rather critical for this sort of clustered supercomputer. The Opteron has a high-bandwidth, low-latency chip to chip interconnect RIGHT on the processor die. With Hypertransport, you could theoretically hang a high-bandwidth netwrok connection right off the processor without having to worry about going through any other controllers. This gives you much lower latency than essentially ANY other processor in the world now, and higher bandwidth than just about anything else out there as well (though bandwidth is always easier to solve than latency, just throw more money at it and you can get more bandwidth).
Given this fact, it's not altogether surprising that the Opteron IS being used to make some HPC cluster. In fact, it's being used for at least 5 major new clusters, at least 4 of which would make the current Top500 list if they were available now. The largest Opteron cluster, being built by Cray for Sandie will feature ~12,000 Opteron processors and will roughly match the performance of the current #1, Japan's Earth Simulator.
No they could not bond NICs, because they're using Infiniband and not ethernet. Besides, I think that they are being limited more by latency than bandwidth, so therefore adding bandwidth isn't going to help much. What's worse, their bandwidth limit is being reached inside the computer, with their chip to chip interconnect having less bandwidth than their computer to computer interconnect.
This is not altogether surprising, given that they are using a desktop computer and trying to shoehorn it into a supercomputer role. They are bound to run into some limitations.
Communication latencies are definitely going to be their major problem, it always is in a super-computer cluster. However, that being said, I'm not sure that they've got the bandwidth problem licked either. Yes,infiniband provides tons of bandwidth, but I'm not sure that the system has the internal bandwidth to make use of it.
They are connecting the Inifiband cards using 64-bit PCI-X, which allows for up to 2.1GB/s of bandwidth. However, the PCI-X controller is only connected to the memory and processor controller chip through an 800MB/s hypertransport link (err,two links, 800MB/s in either direction). This is actually less than the 10Gb/s (1.25GB/s) of bandwidth that infiniband provides. What's even worse, that 800MB/s hypertransport link is shared with basically all I/O other than memory and graphics, and it's also a peak figure, with real-world bandwidth being somewhat lower.
In short, internal system bandwidth could be a problem here. Likely they are also running into some internal system latency issues as well. To get to their infiniband cards, they first have to go from the processor though the Elastic I/O channel to the memory+processor controller chip, than through Hypertransport to the PCI-X, than over the PCI-X bus to the Infiniband card. Each step of the way is adding a bit of latency.
They would have been much better off with something like how the Opteron handles things, with the PCI-X controller hanging right off the processor itself, connected via a 3.2GB/s (each direction) hypertransport link. More bandwidth and fewer chips and buses to go through.