You can get external SCSI ZIP drives with disks for under $10 on ebay, but the PB140 has that funky Apple high-density SCSI connector (so sayeth the AppleSpec page. Conversion cables used to run about $40, but that was perhaps 10 years ago, so they're probably dollar-bin items these days. The spec page also says you should have a floppy drive (regular 1.44MB one, even), which isn't much help these days, but if you've still got a PB140 kicking around, finding a newer mac with a floppy shouldn't be too hard...
I also flipped back and forth between like ABC and NBC, Pacific Time, and notice they were like 3 seconds or so different in their countdown clocks. What is up with that?
Who's going to be producing a 32GB SD card? You can barely buy 16GB CF cards these days.
Let's assume 16GB cards versus a 500GB hard drive. That's 30 cards, for a total volume of 232 cm^3 according to the spec. A 500GB 3.5" hard drive has a volume of 386 cm^3. However, that hard drive includes interface hardware, which 30 CF cards don't. So you've got 160cm^3 to attach 30 CF interfaces and enough circuitry to convert that into some sort of reasonable attachment standard, SATA for example, as well as power, and a reasonably tough enclosure to put it in. Since I presume we're safely in the realm of fantasy now, I'll let you argue that you're not beholden to the CF form factor and therefore don't need that many big pin blocks to plug into. Fair enough. But you do need an awful lot of routing, and a nice beefy front-end chip to run all this, since you can't just pack 500GB of memory on one controller - you'd end up with 10MB/sec transfer rates if you're lucky. We're trying to beat modern disks now, which means you've got to hit at least 70MB/sec. So realistically that's 8 flash controllers with 60GB of flash behind each one, and a SATA power and data connection.
Ok, I think you're right. It probably actually could be done, and since we're ignoring price for the moment I'll concede the point. Flash certainly doesn't win this one hands down, but it looks like you actually could make a device with roughly the same size and capacity as a modern disk drive.
I'm drawing my data from a variety of places, but primarily Pretec's product line. It looks like a Flash-based disk would sport similar non-operating shock tolerance, but would handily beat magnetic storage when spinning. The read error rates are identical, though there's no mention of longevity. As you mention, wear leveling algorithms would help quite a bit, hopefully without sacrificing too much performance for the extra read/write cycles.
Power's a hard one to get a handle on. Pretec's 8GB Flash drive consumes half a watt in use - the 60 of them required to reach about 500GB would draw double the power of a magnetic disk, though the power / GB ratio should improve quite a bit with larger devices as the power consumed by the flash controllers is amortized over a larger amount of storage. Pretec quotes 8MB/sec read/write speed, meaning you'd need 10 of the devices in some sort of array to reach the STR performance of a magnetic disk. The extra flash shouldn't actually consume any power - it's just a matter of the controller chips required. Since we'll need 10 of them to match a disk, plus a SATA controller and some sort of custom ASIC or FPGA or microcontroller or some such to handle wear leveling, striping, remapping of bad blocks, etc. I'd guess the power consumption would be about on par as well. At least for a flash device in operation. It should be able to idle down to less than 50% of full power, which is about what a disk can manage.
Perhaps I spoke a bit too soon. You probably could get 500GB of flash storage with comparable size, power, and performance as a disk drive. So I suppose the question isn't so much "why would you want to put up with the drawbacks" as "why would you want to use new, unproven tech for no real gains". Also, coming back to reality for a moment, there's no way you're going to get 4 trillion and change transistors for anything like $300 in the visible future. Or even 1 trillion, assuming fancy MLC flash cells.
Well, Microsoft is no doubt concerned about ISPs who include branded browsers as part of their install kit restricting or blocking access to the 'net from IE (which is 98% insecure). A wholesale switch to either Moz or Opera isn't the answer (but abandoning IE can't hurt), but both could use somewhat increased market share. A 3-way race with no eventual winner is probably the best possible outcome.
400GB of flash would be bigger, heavier, and probably slower than 400GB of magnetic storage. It would also be less reliable. You might be able to get decent performance in a lower-power, quieter device, but even with price parity, why would you want flash with all its drawbacks?
The winchester hard drive really deserves some sort of award. Second only to the microchip, the hard drive has been the most successful technology product of the past 20 years, I would say. Consider that its evolution in terms of capacity has far outstripped that of the CPU, while its price has remained low. The same basic principles have scaled from the largest several-hundred-pound devices of old to the 19 gram Seagate ST1, and from the early 1MB drives to current half-terabyte drives. These devices can be found in all but the smallest of consumer electronics and in the largest of mainframes. Only the integrated circuit has shown similar technical improvements and wider applicability, yet the hard drive gets little respect, even within the computer industry. Sad.
The lawsuit wasn't taking issue with not being able to get tons of movies. The lawsuit was objecting to Netflix increasing its turnaround time for specific customers based on high usage. If Netflix is working flat out to get people their movies, there's no problem. Once they start deliberately slowing their service down, they're in breach of contract (or false advertising, or however you want to phrase it).
I'll bet Netflix is still making money on the people getting 20 movies a month. I assume they prioritize order fulfillment so infrequent users get the best service - keep the people happiest who're giving you the most profit. In that case, the heavy users are getting "excess" manpower and warehouse space, so the only price Netflix needs to beat is the postage costs. I think you're spot on at $0.60 round trip with presorted bulk mail discounts with the post office. The break-even point in that case is actually 30 movies a month with the $18 plan. As long as Netflix can efficiently prioritize their orders, they can probably make a profit on even extreme corner cases. They just have to make sure those users are getting DVDs which would otherwise be unrented, packed by people who would otherwise be chatting with their coworkers.
> And by the way, if you email support, they will send you an XP disc and a drivers disc (for your model) in the mail for free. It took about 3 days.
I think you need to re-read the review. It didn't quite work out that way. I'll grant that it should have, but I'm unsurprised by how tech support handled the issue.
These bacteria are way too big to be of any use in modern photolithography. Assuming each one's square, and there's 1 per pixel, each bacterium takes up an area of about 6.5 square microns (1 100-millionth of a square inch). For comparison, the smallest production SRAM cell I can find is.25 square microns, and contains 6 transistors. That makes these bacteria 150x as big as a transistor, and even larger when compared to the features that make up the transistors and connect them together.
Now, in situations where you want a physically large product, such as the circuitry to drive an LCD, biology holds huge promise.
isupply aren't some fly-by-night firm. So when isupply runs the numbers on a shipping product and computes today's cost, it's a fact. When Merril Lynch predicts the future for a just-released product and one that won't ship for months, it gets a question mark. I fail to see the problem here.
> Anyone out there who could shed some insight into why aluminum is preferred over well-designed plastic?
Because aluminum is more EXTREME.
The last thumb drive I read, which did a much better job, indicated that performance increased with block size even out to several MB blocks. If my thumb drive usage is any indication, we're mostly dealing with smaller sizes. I would have liked to have seen in the review some more scientific methodology regarding this issue.
Are you sure? That doesn't make sense (except perhaps for initial research). Each of the hundreds of sensor elements under each lens will be imaging a slightly different object, at a slightly different angle. Not taking this into account could explain the softness in even the in-focus parts of the images though.
That's not completely fair. If I understand you, what you're saying is that in fact you DO lose resolution, but the loss in resolution can be compensated for by higher-resolution sensors and because you don't have to increase the physical size of the sensor, the production costs won't go up too much. I don't know enough about CCDs and CMOS sensors to know what the probable increase in cost would be, but it sounds fairly minor. At least for CPUs, I know die size is a stronger indicator of manufacturing cost than transistor count, though the latter obviously plays a role.
The other problems that you've swept under the rug seem to me to be more important, at least in the near term. If you take a CCD and replace each of its sensor sites with a 12x12 array, as you suggest, you're talking over a 100-fold increase in the data to be processed. While I haven't read the technical papers on the subject, it seems like the processing is more complicated than the processing that goes on in a standard digicam, which probably means at least a 200x increase in processing requirements. If you wait for Moore's law to save you, that's 10 years. Budgeting for a more expensive image processor will shave maybe a year or two off that number, but it's still fairly long-term research.
You could reduce the processing needed in-camera by storing closer-to-raw data and doing the processing at a workstation later, but then you have the problem of a data stream that's ~100x as large. Even with very fast flash storage, that would take 30+ seconds to write a single image, and you could only fit a few onto a 1GB card. Also, you introduce the problem that the photographer doesn't get feedback as to what he or she actually shot, and unless you can also post-process to correct for motion blur, abberation in color, etc. you still need that functionality.
It all sounds interesting, and I applaud research into what useful things could be done with likely future technology, but (and maybe I'm misreading the situation) it sounds like the core research is being cast as a thing we could be doing RSN, which I highly doubt. As a technique to make use of sensor densities that would normally exceed the capabilities of the lens they're attached to in order to do something useful, this is interesting. As a technique to be applied to today's or near-future sensors and cameras, I find it less interesting.
The linked article comments that there's an effective loss of resolution, but goes no further.
Obviously taking a camera that's designed to record light intensity and modifying it to record light intensity and direction isn't free. In the worst case, you're decreasing your effective resolution by the number of new lenses, or by a factor of 90,000. I don't think that's quite what happens though, because many of these lenses will be recording essentially the same information, and while only one may be perfectly focussed on part of the frame, nearby lenses can probably contribute color and intensity information as well. If we assume a 2Mpixel image is "good", the article's comment that the student's using a 16Mpixel camera but that an 8Mpixel camera might be good enough seems to support a roughly 4x to 8x decrease in effective resolution. Can the poster who claims to have heard the actual discussion at Siggraph comment?
That's a high price to pay for not having to use the viewfinder. It's cool tech, and I'm sure there are practical uses for it somewhere, but I don't think consumer cameras are the place for it just yet.
Google needs much beefier CPUs than that. They're not running a SAN where pushing bytes is the only goal. They're running a massively parallel distributed computing project, that just happens to like ready access to a boatload of disk.
The petabox project has essentially one design goal: "What is the absolute minimum amount of hardware we can wrap hard drives in and still have a useful system?" And the answer, apparently, is "a 1U half-depth case with a tiny Via board". That they can get power consumption down to 40W/TB is incredible - that's just twice the power consumption of the disk they're building with. But it's not useful to Google.
Google is asking a completely different question, because they need not just a boatload of disk, but a lot of processing power to be constantly crunching that data, either running Gmail, web searches, data analysis, Google Maps, or whatever the AJAX app of the week is. These boxes will be doing a lot more than serving data, they'll be responding to queries that in aggregate will require thousands of MIPS and terabytes of RAM.
Ultimately, I suppose, it's the same quest. Least stuff-you-don't-care-about possible. Most value per {dollar, watt, square foot,...} But the approaches are different. Last I heard Google was using dual-proc 1U machines. I suspect they're still doing that. It's a cheap, standard size, and provides a good mix. You can get a pair of dual-core Opterons in a case that size, along with 2-4GB of RAM fairly cheaply, and you've got room for 2TB of disk too. Multiply that by a few thousand and you've got the numbers Cringeley's quoting.
I understand where you're coming from, but I don't think I buy that. $1M would be a rounder number, and probably closer to the truth. And for a company as big as Google, that's not a big investment in infrastructure, so I don't think the difference is meaningful in terms of making the story sell better. It's still fabulously cheap compared to what an end-user would pay.
I also don't believe that the hardware prices will be lower in the future. Every couple years people forget and someone posts a new big article about how scandalously cheap CPUs are to make. Google's not buying $600 chips, the ones whose prices are going down. They're buying $60 chips (or, say, $200 chips for $60), paying just over the manufacturer's costs, I'd wager. Those costs are going up, not down. If you're buying in sufficient volume that you're getting the lowest price the manufacturer can afford to sell at, your prices are only going to go up as raw materials, manpower, and technology get more expensive. Waiting 6 months still gets you more power for your money, but not a cheaper end result, and that 6 month lead on its competitors is worth way more to Google than a lousy few million dollars.
I'm not even addressing wholesale prices. The cost to the manufacturer to make this stuff sounds too high to hit Cringeley's US$500,000 price point. And while you can undercut that for high-profile uses that are small enough (it just comes out of AMD's ad budget to say that Google's new datacenter uses Opterons), that won't get you 300 peerage points * 5k CPUs each worth of below-cost chips.
Even with liquid cooling it's a hard problem. For one thing you need 5000 little hoses, and beefy pumps to get the water through there at a reasonable speed. Opterons are specced to run up to about 85C (depends slightly on model / family). Suppose you've got incoming water at 10C, and heat it up all the way to 85C. That's 75C difference, or 313.5J/g of water you're taking away. That works out to 5.75 million grams of water per hour, or just under 6000 liters per hour. You can't just dump it into a lake or river or you'll completely nuke the resident ecosystem. It's a manageable number from the point of view of getting it through the machines, but it's still an awful lot of energy to get rid of.
The sort of temperature-differential energy recovery you speak of is technically possible but isn't efficient enough to substantially reduce the cluster's power requirements, and thus its need to vent waste heat.
> what's the point of putting network latency between all those shipping containers?
To remove the network latency between them and you.
They're not being used "for computing" in the sense you're envisioning. For one thing, 5000 Opterons is enough to tackle pretty much any problem you'd care to throw at it, so there's no need to talk to anyone else. For another thing, they wouldn't be doing big computations, they'd be doing massive numbers of small ones. Think Gmail. 3.5PB is enough to store an awful lot of email, and a few thousand Opterons can run rather a lot of simultaneous HTTP connections from people accessing the mail. Add in a fast network link (for talking to all those many people accessing the mail, and for replicating everything offsite), and you're set.
Cringeley's penchant for sensationalism aside, it's pretty clear that Google's got the expertise and the mindset to deal with problems that start with "if we had 10,000 fast CPUs, 10,000 hard disks, and 10,000 GB of RAM...". Google's rapidly expanding, and has been ever since they started. Back when Google fit in a closet, a new server constituted a big expansion. I'm not surprised that these days their unit of expansion is a tractor trailer with a few dozen racks in it. And if you've got something that packages up that nicely, it only makes sense to pepper the globe with capacity.
Actually, I'm curious how Cringeley thinks Google can get the hardware at the prices he quotes in the article. I'm sure he's given it some thought, but unless they're getting hardware at below-cost prices, I don't see how it can be done. The CPUs cost about $50 each to make, so that's $250k for chips. Then you need a few petabytes of disk. I don't know what the manufacturing cost is for disks, but I'd guess about $50 there too. Say $50 for a 500GB drive. That's a few thousand drives to reach the several petabytes, and there goes the rest of his half-million dollars. You still need motherboards, RAM, power supplies, chassis, racks, switches, etc.
I'm not saying he's wrong, but I'd be curious to hear where I've gone astray in my figuring.
Not to mention, of course, the enormous electrical requirements this thing would have, as you've commented. If we round the CPU's power consumption up to account for all the support machinery, and figure 100W per CPU, this neat little semi-load is going to want half a megawatt, plus cooling. Just the disk array will chew through 50kW or so. Even from a power plant's perspective, that's a pretty hefty chunk of juice.
Why do you care about itunes? As far as I can tell, it's a sub-par audio player (huge, slow, nothing to recommend it over winamp), a sub-par music organizer (choking on a lousy 20,000 songs), a nearly feature-less ripper/encoder (though effective), absolute crap for tagging (easytag's where it's at). I can come up with only 2 things to recommend it. One, live playlists or whatever they call them are pretty cool. Two, iTMS. Which I don't care about because the content they provide isn't worth the prices they charge, but I can see it being a selling point for some people.
I know this reads as rather a slam, but I'm serious - what's the attraction to itunes?
But, to your actual question. I use Vorbis. Mostly for audiobooks. Sure, it doesn't work on the ipod, but audiobooks don't really work on the ipod anyway, and vorbis support, while not exactly common, is hardly a rare feature. I can think of 3 big-name players that support it. I'm probably going to switch soon to flac because space really isn't that big an issue. Tripling my audio space requirements would use another $50 or so of disk, and then I don't have to think about it anymore.
128kb MP3s don't sound bad, but they're readily distinguishable from the CD source. 192kb MP3s sound quite good (with VBR and good LAME options), but are still distinguishable from the CD source on certain tracks, and my audio hardware isn't all that high-spec. If I'm going much higher than 192, I might as well go straight to flac and put an end to it. The only reason I can see for not using flac is DAPs, where selection is more limited, and the space actually is an issue.
Actually, the CD logo isn't the guarantee it should be. Though if there's a problem, you can always threaten to sue under truth in advertising laws, which usually gets you a return, since the stores don't actually HAVE a real audio CD of the content.
Articles about this topic seem inevitably to get bogged down in details. Why only 512MB? Why only 100 songs, regardless of length? Why so big, why so ugly, why....?
That's not the point. Ultimately the technical issues are what they are. The central point is that Apple could have done better, and did not. The market was not clamoring for a mp3 phone, so Apple's decision was not borne out of rush-to-market, and must therefore have been a conscious decision to make a sub-par product. THAT is the question that people should be concerning themselves with: Why did Apple deliberately release a crap product?
I think the only reasonable answer is to manipulate the market. Whether you call that sobotage or not depends more on your view of Apple than it does on their actions in this case, but I think it's pretty clear that market manipulation was the goal. Certainly Apple's not making much money on this deal. They're hardly polishing their image with it. It's not increasing iTMS market penetration or visibility. I think Apple's sending a message, and that message is "music phones are not a viable product yet". Apple 'tries' and fails. The market, seeing this, reasons that it can't be done because Apple's the music god right now and if they can't do it, nobody can. And the VCs aren't going to send out money to people intending to try, because they've got the best evidence in the world that it can't be done well.
Pirates, man! I mean, a real actual honest-to-$deity story. On Slashdot. About pirates! I mean, PIRATES. The ones with hats. And parrots. I never thought I'd see the day.
Well, yeah, sort of. For "opened" read "opened its doors for press and qualification". AMD's press release about the event says production shipments will begin in 1Q06, with the production ramp and 65nm conversion continuing through 2007. There's future capacity coming online which AMD could easily earmark a part of for Apple. My apologies for the lack of clarity.
You're right about Intel though. I could have sworn they're having shortages with the low- to mid- range chipsets these days as well, but I can't find a corroborating link right now. I blame the clock ticking past 3am here.
Volume is an often-quoted reason for going with Intel over AMD, but I don't think I buy it. It's true that Intel has much more manufacturing capacity than AMD, but they also have greater outstanding contracts. AMD is constantly expanding their manufacturing capacity, and are in the process of building another huge fab in Dresden - I think AMD could easily meet Apple's projected demands, particularly since they've got a year or so to do it. Bear in mind that 1M CPUs / quarter is less than 10% of Fab 30's capacity, let alone AMD's total chip-manufacturing capacity.
I suspect Intel just cut Apple a better deal than AMD was willing to, particularly since Apple wants desktop, mobile, consumer electronics, and flash chips. Dealing with only one supplier for all 4 segments may be the cheapest and easiest option. Cringely has an interesting take on the deal, speculating that Apple went to Intel not for technology reasons, or for financial reasons, but for business reasons. Apple and Intel each have something the other wants, and that's better than a straight cash-for-CPUs deal that Apple would get with AMD.
Slashdot Mirror
Download? No.
You can get external SCSI ZIP drives with disks for under $10 on ebay, but the PB140 has that funky Apple high-density SCSI connector (so sayeth the AppleSpec page. Conversion cables used to run about $40, but that was perhaps 10 years ago, so they're probably dollar-bin items these days. The spec page also says you should have a floppy drive (regular 1.44MB one, even), which isn't much help these days, but if you've still got a PB140 kicking around, finding a newer mac with a floppy shouldn't be too hard...
Differing amounts of broadcast time delay?
Who's going to be producing a 32GB SD card? You can barely buy 16GB CF cards these days.
Let's assume 16GB cards versus a 500GB hard drive. That's 30 cards, for a total volume of 232 cm^3 according to the spec. A 500GB 3.5" hard drive has a volume of 386 cm^3. However, that hard drive includes interface hardware, which 30 CF cards don't. So you've got 160cm^3 to attach 30 CF interfaces and enough circuitry to convert that into some sort of reasonable attachment standard, SATA for example, as well as power, and a reasonably tough enclosure to put it in. Since I presume we're safely in the realm of fantasy now, I'll let you argue that you're not beholden to the CF form factor and therefore don't need that many big pin blocks to plug into. Fair enough. But you do need an awful lot of routing, and a nice beefy front-end chip to run all this, since you can't just pack 500GB of memory on one controller - you'd end up with 10MB/sec transfer rates if you're lucky. We're trying to beat modern disks now, which means you've got to hit at least 70MB/sec. So realistically that's 8 flash controllers with 60GB of flash behind each one, and a SATA power and data connection.
Ok, I think you're right. It probably actually could be done, and since we're ignoring price for the moment I'll concede the point. Flash certainly doesn't win this one hands down, but it looks like you actually could make a device with roughly the same size and capacity as a modern disk drive.
I'm drawing my data from a variety of places, but primarily Pretec's product line. It looks like a Flash-based disk would sport similar non-operating shock tolerance, but would handily beat magnetic storage when spinning. The read error rates are identical, though there's no mention of longevity. As you mention, wear leveling algorithms would help quite a bit, hopefully without sacrificing too much performance for the extra read/write cycles.
Power's a hard one to get a handle on. Pretec's 8GB Flash drive consumes half a watt in use - the 60 of them required to reach about 500GB would draw double the power of a magnetic disk, though the power / GB ratio should improve quite a bit with larger devices as the power consumed by the flash controllers is amortized over a larger amount of storage. Pretec quotes 8MB/sec read/write speed, meaning you'd need 10 of the devices in some sort of array to reach the STR performance of a magnetic disk. The extra flash shouldn't actually consume any power - it's just a matter of the controller chips required. Since we'll need 10 of them to match a disk, plus a SATA controller and some sort of custom ASIC or FPGA or microcontroller or some such to handle wear leveling, striping, remapping of bad blocks, etc. I'd guess the power consumption would be about on par as well. At least for a flash device in operation. It should be able to idle down to less than 50% of full power, which is about what a disk can manage.
Perhaps I spoke a bit too soon. You probably could get 500GB of flash storage with comparable size, power, and performance as a disk drive. So I suppose the question isn't so much "why would you want to put up with the drawbacks" as "why would you want to use new, unproven tech for no real gains". Also, coming back to reality for a moment, there's no way you're going to get 4 trillion and change transistors for anything like $300 in the visible future. Or even 1 trillion, assuming fancy MLC flash cells.
Well, Microsoft is no doubt concerned about ISPs who include branded browsers as part of their install kit restricting or blocking access to the 'net from IE (which is 98% insecure). A wholesale switch to either Moz or Opera isn't the answer (but abandoning IE can't hurt), but both could use somewhat increased market share. A 3-way race with no eventual winner is probably the best possible outcome.
Why?
400GB of flash would be bigger, heavier, and probably slower than 400GB of magnetic storage. It would also be less reliable. You might be able to get decent performance in a lower-power, quieter device, but even with price parity, why would you want flash with all its drawbacks?
The winchester hard drive really deserves some sort of award. Second only to the microchip, the hard drive has been the most successful technology product of the past 20 years, I would say. Consider that its evolution in terms of capacity has far outstripped that of the CPU, while its price has remained low. The same basic principles have scaled from the largest several-hundred-pound devices of old to the 19 gram Seagate ST1, and from the early 1MB drives to current half-terabyte drives. These devices can be found in all but the smallest of consumer electronics and in the largest of mainframes. Only the integrated circuit has shown similar technical improvements and wider applicability, yet the hard drive gets little respect, even within the computer industry. Sad.
The lawsuit wasn't taking issue with not being able to get tons of movies. The lawsuit was objecting to Netflix increasing its turnaround time for specific customers based on high usage. If Netflix is working flat out to get people their movies, there's no problem. Once they start deliberately slowing their service down, they're in breach of contract (or false advertising, or however you want to phrase it).
I'll bet Netflix is still making money on the people getting 20 movies a month. I assume they prioritize order fulfillment so infrequent users get the best service - keep the people happiest who're giving you the most profit. In that case, the heavy users are getting "excess" manpower and warehouse space, so the only price Netflix needs to beat is the postage costs. I think you're spot on at $0.60 round trip with presorted bulk mail discounts with the post office. The break-even point in that case is actually 30 movies a month with the $18 plan. As long as Netflix can efficiently prioritize their orders, they can probably make a profit on even extreme corner cases. They just have to make sure those users are getting DVDs which would otherwise be unrented, packed by people who would otherwise be chatting with their coworkers.
> And by the way, if you email support, they will send you an XP disc and a drivers disc (for your model) in the mail for free. It took about 3 days.
I think you need to re-read the review. It didn't quite work out that way. I'll grant that it should have, but I'm unsurprised by how tech support handled the issue.
These bacteria are way too big to be of any use in modern photolithography. Assuming each one's square, and there's 1 per pixel, each bacterium takes up an area of about 6.5 square microns (1 100-millionth of a square inch). For comparison, the smallest production SRAM cell I can find is .25 square microns, and contains 6 transistors. That makes these bacteria 150x as big as a transistor, and even larger when compared to the features that make up the transistors and connect them together.
Now, in situations where you want a physically large product, such as the circuitry to drive an LCD, biology holds huge promise.
isupply aren't some fly-by-night firm. So when isupply runs the numbers on a shipping product and computes today's cost, it's a fact. When Merril Lynch predicts the future for a just-released product and one that won't ship for months, it gets a question mark. I fail to see the problem here.
> Anyone out there who could shed some insight into why aluminum is preferred over well-designed plastic?
Because aluminum is more EXTREME.
The last thumb drive I read, which did a much better job, indicated that performance increased with block size even out to several MB blocks. If my thumb drive usage is any indication, we're mostly dealing with smaller sizes. I would have liked to have seen in the review some more scientific methodology regarding this issue.
Are you sure? That doesn't make sense (except perhaps for initial research). Each of the hundreds of sensor elements under each lens will be imaging a slightly different object, at a slightly different angle. Not taking this into account could explain the softness in even the in-focus parts of the images though.
That's not completely fair. If I understand you, what you're saying is that in fact you DO lose resolution, but the loss in resolution can be compensated for by higher-resolution sensors and because you don't have to increase the physical size of the sensor, the production costs won't go up too much. I don't know enough about CCDs and CMOS sensors to know what the probable increase in cost would be, but it sounds fairly minor. At least for CPUs, I know die size is a stronger indicator of manufacturing cost than transistor count, though the latter obviously plays a role.
The other problems that you've swept under the rug seem to me to be more important, at least in the near term. If you take a CCD and replace each of its sensor sites with a 12x12 array, as you suggest, you're talking over a 100-fold increase in the data to be processed. While I haven't read the technical papers on the subject, it seems like the processing is more complicated than the processing that goes on in a standard digicam, which probably means at least a 200x increase in processing requirements. If you wait for Moore's law to save you, that's 10 years. Budgeting for a more expensive image processor will shave maybe a year or two off that number, but it's still fairly long-term research.
You could reduce the processing needed in-camera by storing closer-to-raw data and doing the processing at a workstation later, but then you have the problem of a data stream that's ~100x as large. Even with very fast flash storage, that would take 30+ seconds to write a single image, and you could only fit a few onto a 1GB card. Also, you introduce the problem that the photographer doesn't get feedback as to what he or she actually shot, and unless you can also post-process to correct for motion blur, abberation in color, etc. you still need that functionality.
It all sounds interesting, and I applaud research into what useful things could be done with likely future technology, but (and maybe I'm misreading the situation) it sounds like the core research is being cast as a thing we could be doing RSN, which I highly doubt. As a technique to make use of sensor densities that would normally exceed the capabilities of the lens they're attached to in order to do something useful, this is interesting. As a technique to be applied to today's or near-future sensors and cameras, I find it less interesting.
The linked article comments that there's an effective loss of resolution, but goes no further.
Obviously taking a camera that's designed to record light intensity and modifying it to record light intensity and direction isn't free. In the worst case, you're decreasing your effective resolution by the number of new lenses, or by a factor of 90,000. I don't think that's quite what happens though, because many of these lenses will be recording essentially the same information, and while only one may be perfectly focussed on part of the frame, nearby lenses can probably contribute color and intensity information as well. If we assume a 2Mpixel image is "good", the article's comment that the student's using a 16Mpixel camera but that an 8Mpixel camera might be good enough seems to support a roughly 4x to 8x decrease in effective resolution. Can the poster who claims to have heard the actual discussion at Siggraph comment?
That's a high price to pay for not having to use the viewfinder. It's cool tech, and I'm sure there are practical uses for it somewhere, but I don't think consumer cameras are the place for it just yet.
Google needs much beefier CPUs than that. They're not running a SAN where pushing bytes is the only goal. They're running a massively parallel distributed computing project, that just happens to like ready access to a boatload of disk.
...} But the approaches are different. Last I heard Google was using dual-proc 1U machines. I suspect they're still doing that. It's a cheap, standard size, and provides a good mix. You can get a pair of dual-core Opterons in a case that size, along with 2-4GB of RAM fairly cheaply, and you've got room for 2TB of disk too. Multiply that by a few thousand and you've got the numbers Cringeley's quoting.
The petabox project has essentially one design goal: "What is the absolute minimum amount of hardware we can wrap hard drives in and still have a useful system?" And the answer, apparently, is "a 1U half-depth case with a tiny Via board". That they can get power consumption down to 40W/TB is incredible - that's just twice the power consumption of the disk they're building with. But it's not useful to Google.
Google is asking a completely different question, because they need not just a boatload of disk, but a lot of processing power to be constantly crunching that data, either running Gmail, web searches, data analysis, Google Maps, or whatever the AJAX app of the week is. These boxes will be doing a lot more than serving data, they'll be responding to queries that in aggregate will require thousands of MIPS and terabytes of RAM.
Ultimately, I suppose, it's the same quest. Least stuff-you-don't-care-about possible. Most value per {dollar, watt, square foot,
I understand where you're coming from, but I don't think I buy that. $1M would be a rounder number, and probably closer to the truth. And for a company as big as Google, that's not a big investment in infrastructure, so I don't think the difference is meaningful in terms of making the story sell better. It's still fabulously cheap compared to what an end-user would pay.
I also don't believe that the hardware prices will be lower in the future. Every couple years people forget and someone posts a new big article about how scandalously cheap CPUs are to make. Google's not buying $600 chips, the ones whose prices are going down. They're buying $60 chips (or, say, $200 chips for $60), paying just over the manufacturer's costs, I'd wager. Those costs are going up, not down. If you're buying in sufficient volume that you're getting the lowest price the manufacturer can afford to sell at, your prices are only going to go up as raw materials, manpower, and technology get more expensive. Waiting 6 months still gets you more power for your money, but not a cheaper end result, and that 6 month lead on its competitors is worth way more to Google than a lousy few million dollars.
I'm not even addressing wholesale prices. The cost to the manufacturer to make this stuff sounds too high to hit Cringeley's US$500,000 price point. And while you can undercut that for high-profile uses that are small enough (it just comes out of AMD's ad budget to say that Google's new datacenter uses Opterons), that won't get you 300 peerage points * 5k CPUs each worth of below-cost chips.
Even with liquid cooling it's a hard problem. For one thing you need 5000 little hoses, and beefy pumps to get the water through there at a reasonable speed. Opterons are specced to run up to about 85C (depends slightly on model / family). Suppose you've got incoming water at 10C, and heat it up all the way to 85C. That's 75C difference, or 313.5J/g of water you're taking away. That works out to 5.75 million grams of water per hour, or just under 6000 liters per hour. You can't just dump it into a lake or river or you'll completely nuke the resident ecosystem. It's a manageable number from the point of view of getting it through the machines, but it's still an awful lot of energy to get rid of.
The sort of temperature-differential energy recovery you speak of is technically possible but isn't efficient enough to substantially reduce the cluster's power requirements, and thus its need to vent waste heat.
> what's the point of putting network latency between all those shipping containers?
To remove the network latency between them and you.
They're not being used "for computing" in the sense you're envisioning. For one thing, 5000 Opterons is enough to tackle pretty much any problem you'd care to throw at it, so there's no need to talk to anyone else. For another thing, they wouldn't be doing big computations, they'd be doing massive numbers of small ones. Think Gmail. 3.5PB is enough to store an awful lot of email, and a few thousand Opterons can run rather a lot of simultaneous HTTP connections from people accessing the mail. Add in a fast network link (for talking to all those many people accessing the mail, and for replicating everything offsite), and you're set.
Cringeley's penchant for sensationalism aside, it's pretty clear that Google's got the expertise and the mindset to deal with problems that start with "if we had 10,000 fast CPUs, 10,000 hard disks, and 10,000 GB of RAM...". Google's rapidly expanding, and has been ever since they started. Back when Google fit in a closet, a new server constituted a big expansion. I'm not surprised that these days their unit of expansion is a tractor trailer with a few dozen racks in it. And if you've got something that packages up that nicely, it only makes sense to pepper the globe with capacity.
Actually, I'm curious how Cringeley thinks Google can get the hardware at the prices he quotes in the article. I'm sure he's given it some thought, but unless they're getting hardware at below-cost prices, I don't see how it can be done. The CPUs cost about $50 each to make, so that's $250k for chips. Then you need a few petabytes of disk. I don't know what the manufacturing cost is for disks, but I'd guess about $50 there too. Say $50 for a 500GB drive. That's a few thousand drives to reach the several petabytes, and there goes the rest of his half-million dollars. You still need motherboards, RAM, power supplies, chassis, racks, switches, etc.
I'm not saying he's wrong, but I'd be curious to hear where I've gone astray in my figuring.
Not to mention, of course, the enormous electrical requirements this thing would have, as you've commented. If we round the CPU's power consumption up to account for all the support machinery, and figure 100W per CPU, this neat little semi-load is going to want half a megawatt, plus cooling. Just the disk array will chew through 50kW or so. Even from a power plant's perspective, that's a pretty hefty chunk of juice.
Why do you care about itunes? As far as I can tell, it's a sub-par audio player (huge, slow, nothing to recommend it over winamp), a sub-par music organizer (choking on a lousy 20,000 songs), a nearly feature-less ripper/encoder (though effective), absolute crap for tagging (easytag's where it's at). I can come up with only 2 things to recommend it. One, live playlists or whatever they call them are pretty cool. Two, iTMS. Which I don't care about because the content they provide isn't worth the prices they charge, but I can see it being a selling point for some people.
I know this reads as rather a slam, but I'm serious - what's the attraction to itunes?
But, to your actual question. I use Vorbis. Mostly for audiobooks. Sure, it doesn't work on the ipod, but audiobooks don't really work on the ipod anyway, and vorbis support, while not exactly common, is hardly a rare feature. I can think of 3 big-name players that support it. I'm probably going to switch soon to flac because space really isn't that big an issue. Tripling my audio space requirements would use another $50 or so of disk, and then I don't have to think about it anymore.
128kb MP3s don't sound bad, but they're readily distinguishable from the CD source. 192kb MP3s sound quite good (with VBR and good LAME options), but are still distinguishable from the CD source on certain tracks, and my audio hardware isn't all that high-spec. If I'm going much higher than 192, I might as well go straight to flac and put an end to it. The only reason I can see for not using flac is DAPs, where selection is more limited, and the space actually is an issue.
Actually, the CD logo isn't the guarantee it should be. Though if there's a problem, you can always threaten to sue under truth in advertising laws, which usually gets you a return, since the stores don't actually HAVE a real audio CD of the content.
Articles about this topic seem inevitably to get bogged down in details. Why only 512MB? Why only 100 songs, regardless of length? Why so big, why so ugly, why ....?
That's not the point. Ultimately the technical issues are what they are. The central point is that Apple could have done better, and did not. The market was not clamoring for a mp3 phone, so Apple's decision was not borne out of rush-to-market, and must therefore have been a conscious decision to make a sub-par product. THAT is the question that people should be concerning themselves with: Why did Apple deliberately release a crap product?
I think the only reasonable answer is to manipulate the market. Whether you call that sobotage or not depends more on your view of Apple than it does on their actions in this case, but I think it's pretty clear that market manipulation was the goal. Certainly Apple's not making much money on this deal. They're hardly polishing their image with it. It's not increasing iTMS market penetration or visibility. I think Apple's sending a message, and that message is "music phones are not a viable product yet". Apple 'tries' and fails. The market, seeing this, reasons that it can't be done because Apple's the music god right now and if they can't do it, nobody can. And the VCs aren't going to send out money to people intending to try, because they've got the best evidence in the world that it can't be done well.
Pirates, man! I mean, a real actual honest-to-$deity story. On Slashdot. About pirates! I mean, PIRATES. The ones with hats. And parrots. I never thought I'd see the day.
Well, yeah, sort of. For "opened" read "opened its doors for press and qualification". AMD's press release about the event says production shipments will begin in 1Q06, with the production ramp and 65nm conversion continuing through 2007. There's future capacity coming online which AMD could easily earmark a part of for Apple. My apologies for the lack of clarity.
You're right about Intel though. I could have sworn they're having shortages with the low- to mid- range chipsets these days as well, but I can't find a corroborating link right now. I blame the clock ticking past 3am here.
Volume is an often-quoted reason for going with Intel over AMD, but I don't think I buy it. It's true that Intel has much more manufacturing capacity than AMD, but they also have greater outstanding contracts. AMD is constantly expanding their manufacturing capacity, and are in the process of building another huge fab in Dresden - I think AMD could easily meet Apple's projected demands, particularly since they've got a year or so to do it. Bear in mind that 1M CPUs / quarter is less than 10% of Fab 30's capacity, let alone AMD's total chip-manufacturing capacity.
I suspect Intel just cut Apple a better deal than AMD was willing to, particularly since Apple wants desktop, mobile, consumer electronics, and flash chips. Dealing with only one supplier for all 4 segments may be the cheapest and easiest option. Cringely has an interesting take on the deal, speculating that Apple went to Intel not for technology reasons, or for financial reasons, but for business reasons. Apple and Intel each have something the other wants, and that's better than a straight cash-for-CPUs deal that Apple would get with AMD.