You're assuming that they are all independent. ARM's role in this is to coordinate cross-licensing, so companies chasing different market segments can share R&D results and even portions of processor design.
With ARMv8, a lot of companies have this kind of license. I think there are six independent ARMv8 implementations that I'm aware of, but there may be more. Well, I say independent - they all had engineers from ARM consult on the design, but they're quite different in pipeline structure. This is the problem Intel is going to face over the next few years. They could compete with AMD by outspending them on R&D: Intel could afford to design 10 processors and only bring 3 to market depending on what customers wanted, AMD couldn't afford to throw away that much investment. This is how the Pentium M happened: they rushed to market one of their back-burner designs. Now, however, they're competing with half a dozen companies all of whom have ISA-compatible chips and all of whom are content to heavily optimise their designs for a particular market segment.
Most people get Android apps from an App Store though, and it can easily select the correct version. This happens already for MIPS-based devices, so there's no reason why it wouldn't work for x86, if it were worth the effort for developers to provide them.
Here's one that's easy: outright lying. Unless you're arguing that fraud shouldn't be illegal, because it's just an expression of free speech. Astroturfing is a form of fraud: you're trying to present views as coming from someone else. If the cable companies want to say 'net neutrality is bad because it will cost us money', then that's fine. If they lie and pretend to be a consumer group, then that's not.
So when the 'gambling addicted nutjobs' take their commission on every trade, that doesn't harm anyone by increasing costs for investors? When they cause a flash crash, that doesn't harm anyone by making it harder for companies to raise capital? When they lose all of their money, cry that it was an algorithmic error, and get the exchange to reverse the transactions so that they can keep gambling but other people take the risks, they don't harm anyone? Good to know, thanks for clearing that up.
As a business traveller who isn't going to type his credit card details into a random captive portal that claims to be affiliated with an airport-honest-really-not-a-scam, I welcome free WiFi at airports.
Public transport is already easy to do with conventional batteries. Between overhead power lines and induction charging at bus stops, there are ample opportunities to top up the batteries during the day.
This isn't meant to replace that, it's meant to augment it. Most of the time, the 260 mile range is fine. Sometimes, it isn't, and this gives you an emergency reserve. In normal use, you'll never use it, but if you're planning a long trip and don't manage to get to a charging station anywhere in the middle then you're not going to be stuck miles from civilisation with an empty battery.
I don't think I've ever seen a Phoronix report that showed error bars in their graphs and I've never done performance tests of the ones that they run without any jitter. They also have a history of making spectacularly bad decisions about what to benchmark. For example, they did benchmarks of FreeBSD against Ubuntu, but used a FreeBSD beta that had WITNESS and INVARIANTS options compiled into the kernel, both of which impose a fairly significant performance cost (but give better error reporting). Yesterday, they benchmarked GCC's OpenMP support against Intel's Clang-omp branch. Sounds easy, except that they didn't specify optimisation flags and GCC defaults to -O2, whereas Clang defaults to -O0. A sane evaluation would have shown the results for -O0 to -O3 for both compilers, along with error bars showing the standard deviation over ten or more runs.
A subscription for Phoronix? The site that yesterday published a multi-page article comparing Clang and GCC's OpenMP performance using -O0 for Clang and -O2 for GCC? I'm sure that's a good use of money...
If you are a subscriber, then you can use HTTPS, however your traffic then stands out to traffic analysis (it's not hard to tie the encrypted connection which sends more information than an HTTP GET with the sudden appearance of a new post in the next user's unencrypted request for the page). They're also, I believe, still using a certificate that was issued before Heartbleed was disclosed and so is almost certainly compromised.
I can't speak for Finnland, but in the UK traffic offences are handled a bit differently. You typically receive a letter in the post telling you to pay a fine or appear in court. If you pay the fine, it's cheaper (the same number of points are still added to your license in either case). If you're definitely guilty then it's the obvious thing to do. Before you get to court, there are several things you can do, including requesting the camera photos. These are typically the only evidence that they have and requesting them means that a human will look at them carefully. If there's obviously been an error, then they will usually drop the case then. Because most of the obviously guilty cases never make it to court, the magistrates expect that a significant number of people who actually make it that far will be not guilty, so there isn't quite the prejudice that the US system has.
OpenSSL took the worst possible route. They had FOO_{standard library function} and BAR_{standard library function}, and also just used the unadorned library function. The FOO_ variant had some special behaviour, the BAR_ version was sometimes the standard library version and sometimes their own (depending on both the platform and the function - in some cases they always wrote their own even when there's an adequate - or even better - version shipped with the platform, in some cases they made a per-platform decision about it).
Almost right, except that you stick #define printf GOOD_printf at the end of the #ifdef block and then always use printf(), don't force everyone reading the code to work out that GOOD_printf() means printf().
Closed-source for-profit software is typically backed by a company that is willing to (but not always competent at) fire bad programmers whose contributions end up adding more work for others than solving problems.
The problem with this idea is that it assumes that it's easy to attribute costs to a particular programmer. The Heartbleed bug added a new feature (yay! Marketing checkbox!) and sat undiscovered for a couple of years. Do you think that the developer would have been fired immediately after introducing it in a closed-source environment?
You mention 256GB of ram, which nothing needs to compile, but I wonder, are you using FreeBSD's equivalent of tmpfs (ramdisk, I don't know what it's called) for storage of the build tree? I'm sure that would result in massive speedup over compiling from a on disk/ssd filesystem.
That's not quite true. For one thing, if your buffer cache is big enough that the entire source tree stays there then it has a noticeable impact on compile time. I tend to put the obj tree in tmpfs when doing a universe build (I don't actually want it, I just want to check everything builds most of the time), and that consumes 50GB. You might also be surprised at how much RAM the compile takes.
Bullshit, Quartus II has had Multiprocessor support since at least version 7.2 (I can't find the exact release they added it), and has been improving P&R speeds every release since then. In the version I am using, 14.1 Web Edition, Multiprocessor support requires Talkback (give back performance metrics in exchange for parallel speedup) and works quite well. You should probably look at buying the subscription version and adding some design partitions if your synthesis is taking an hour and a half.
Hmm, is it something that you need to specifically enable? We're taking about an hour and a half with Quartus 13 on these machines. We do pay for the subscription version, although adding more design partitions is not a simple thing to do as our Verilog is generated from a higher-level HDL.
Can you define what you mean by 'a lot' of compiling? We have some machines where I regularly build LLVM and the FreeBSD base system. They're 32-core with 256GB of RAM, and on them most builds are nice and fast, although it still takes close to two hours to do a universe build (all kernel configs, userland for all supported architectures - the 'quick' sanity test that you haven't broken anything). They're similar specs to the 'beefy' machines in the FreeBSD cluster, which take around 24 hours to build a complete package set (all 24K ports for a single release / architecture).
I'm not sure what you got 6 years ago, but there are still things where the fastest you can buy it not as fast as you'd like. I'm really interested in better single-threaded performance, because some companies (*cough*Altera*cough*Xylinx*cough*) haven't yet worked out how to multithread their compilers and so it takes an hour and a half to do a synthesis run of our CPU.
As I said in my original post, the thing that killed it was their time to market. It was initially proposed as a device where you'd pay a slight premium, but get something with an open software stack and a nice form factor as a hacker toy. By the time it shipped, that premium had become 200-300% over comparable devices.
512MB made sense in 2008. It was the largest available Package-on-Package (PoP) RAM module you could buy. The OMAP series of SoCs are usually used in a stacked configuration, with the CPU, RAM, and Flash all vertically stacked. This reduces the complexity of the board a lot, because you don't need any traces to be run for RAM or Flash (only for external peripherals), but it also means that you're quite restricted in the memory chips you can use: they must be single-chip modules and must have a sufficiently low power and thermal envelope that they can sit happily on top of the SoC. 1GB PoP modules were released a few months after the Pandora team finished the design, but it was too late to replace it by then and they've never done a revision B.
I was really interested in the Pandora when it was announced, but eventually got a phone with similar specs for about £50, before most people got their Pandoras. I've now also got an Asus TransformerPad TF700, which is a very nice machine, but I'd really like something that could run FreeBSD out of the box. The Dragonbox Pyra specs look pretty good, but I note a lack of any physical dimensions on the web site (other than a 5" screen), which makes it a bit difficult to evaluate. I didn't realise hns was involved - hopefully this will go a bit better than the GTA handsets...
TI is notoriously bad at documenting their non-CPU cores. The c64x still needs a proprietary compiler (which supports only a weird dialect of C). Is there some commitment from them to open specs for the OMAP5 SoC? One of the nice things about the ChromeBook is that Google has the purchasing power to be able to demand this from their suppliers. I don't think the Pyra has the same influence, so may need to shop around a bit more...
Why do you think that contradicts his statement? Take a look at how much funding GCHQ receives from the NSA - it's a significant amount of their total budget and has led to some concerns that they act more in the interests of the US than UK.
The big problem with OpenPandora was their time to market. When they started (2008), the Cortex A8 was pretty new (released 2007, but not much silicon until 2008) and there were few devices you could get with one in, although a few evaluation boards were starting to appear. They promised something in a small mobile form factor running a completely open software stack, which sounded like a fun platform. By the time they actually shipped anything (2010), the Cortex A8 was starting to feel a bit dated and was available in cheap mobile phones running Android. By the time they shipped to more than a token number of people, the A8 was ancient and you could get a dual or quad-core A9 or similar for half the price.
Oh, and their UK operation effectively went bust after taking huge numbers of pre-orders. Both the UK and German companies had the pre-order money in their accounts for 2+ years earning interest before they shipped anything, but at least people who ordered things from the German company eventually got something, even if it was worth a fraction of what they paid for it, as a result of being two years - over an entire generation of the technology - old by the time they got it.
The even larger problem, as I see it, is that being hired because of a quota is the ultimate stigma: "Look at her, she only got this job because of her tits." No-one takes the quota employee seriously - even when they actually are the best.
This is a very serious problem and has actually led to an increase in biases in some parts of the world. If 90% of a population of doctors of a particular caste/ethnicity/gender are hired based on quotas, then there's a good chance that when you see a doctor who meets those criteria that they will be less competent than someone who doesn't. You then start assuming that everyone in that category is less competent, even if you're unaware of the quota. For the 10% who deserve to be there, this makes life difficult because no one knows that they are the competent ones. This leads to fewer competent applicants from the minority group, because they know that they'll be stigmatised even if they're at the very top of the profession. And how do you address the lack of applicants? More quotas!
You can't fix inequality at the hiring end, you need to find out why you aren't getting enough competent applicants from a minority group and work backwards. Is it because they aren't entering the right degree programs? Is that because of bias in admissions, problems with earlier education, cultural bias, or some other factors?
Slashdot Mirror
Oh yes, Islam was not supposed to replace Christianity
Muhammad originally tried to be a messiah to the Jews, he had little interaction with Christianity.
You're assuming that they are all independent. ARM's role in this is to coordinate cross-licensing, so companies chasing different market segments can share R&D results and even portions of processor design.
With ARMv8, a lot of companies have this kind of license. I think there are six independent ARMv8 implementations that I'm aware of, but there may be more. Well, I say independent - they all had engineers from ARM consult on the design, but they're quite different in pipeline structure. This is the problem Intel is going to face over the next few years. They could compete with AMD by outspending them on R&D: Intel could afford to design 10 processors and only bring 3 to market depending on what customers wanted, AMD couldn't afford to throw away that much investment. This is how the Pentium M happened: they rushed to market one of their back-burner designs. Now, however, they're competing with half a dozen companies all of whom have ISA-compatible chips and all of whom are content to heavily optimise their designs for a particular market segment.
Most people get Android apps from an App Store though, and it can easily select the correct version. This happens already for MIPS-based devices, so there's no reason why it wouldn't work for x86, if it were worth the effort for developers to provide them.
Here's one that's easy: outright lying. Unless you're arguing that fraud shouldn't be illegal, because it's just an expression of free speech. Astroturfing is a form of fraud: you're trying to present views as coming from someone else. If the cable companies want to say 'net neutrality is bad because it will cost us money', then that's fine. If they lie and pretend to be a consumer group, then that's not.
So when the 'gambling addicted nutjobs' take their commission on every trade, that doesn't harm anyone by increasing costs for investors? When they cause a flash crash, that doesn't harm anyone by making it harder for companies to raise capital? When they lose all of their money, cry that it was an algorithmic error, and get the exchange to reverse the transactions so that they can keep gambling but other people take the risks, they don't harm anyone? Good to know, thanks for clearing that up.
As a business traveller who isn't going to type his credit card details into a random captive portal that claims to be affiliated with an airport-honest-really-not-a-scam, I welcome free WiFi at airports.
Public transport is already easy to do with conventional batteries. Between overhead power lines and induction charging at bus stops, there are ample opportunities to top up the batteries during the day.
This isn't meant to replace that, it's meant to augment it. Most of the time, the 260 mile range is fine. Sometimes, it isn't, and this gives you an emergency reserve. In normal use, you'll never use it, but if you're planning a long trip and don't manage to get to a charging station anywhere in the middle then you're not going to be stuck miles from civilisation with an empty battery.
I don't think I've ever seen a Phoronix report that showed error bars in their graphs and I've never done performance tests of the ones that they run without any jitter. They also have a history of making spectacularly bad decisions about what to benchmark. For example, they did benchmarks of FreeBSD against Ubuntu, but used a FreeBSD beta that had WITNESS and INVARIANTS options compiled into the kernel, both of which impose a fairly significant performance cost (but give better error reporting). Yesterday, they benchmarked GCC's OpenMP support against Intel's Clang-omp branch. Sounds easy, except that they didn't specify optimisation flags and GCC defaults to -O2, whereas Clang defaults to -O0. A sane evaluation would have shown the results for -O0 to -O3 for both compilers, along with error bars showing the standard deviation over ten or more runs.
A subscription for Phoronix? The site that yesterday published a multi-page article comparing Clang and GCC's OpenMP performance using -O0 for Clang and -O2 for GCC? I'm sure that's a good use of money...
If you are a subscriber, then you can use HTTPS, however your traffic then stands out to traffic analysis (it's not hard to tie the encrypted connection which sends more information than an HTTP GET with the sudden appearance of a new post in the next user's unencrypted request for the page). They're also, I believe, still using a certificate that was issued before Heartbleed was disclosed and so is almost certainly compromised.
Shooting politicians will almost certainly effect change, but it most likely won't be the change that you're looking for.
I can't speak for Finnland, but in the UK traffic offences are handled a bit differently. You typically receive a letter in the post telling you to pay a fine or appear in court. If you pay the fine, it's cheaper (the same number of points are still added to your license in either case). If you're definitely guilty then it's the obvious thing to do. Before you get to court, there are several things you can do, including requesting the camera photos. These are typically the only evidence that they have and requesting them means that a human will look at them carefully. If there's obviously been an error, then they will usually drop the case then. Because most of the obviously guilty cases never make it to court, the magistrates expect that a significant number of people who actually make it that far will be not guilty, so there isn't quite the prejudice that the US system has.
OpenSSL took the worst possible route. They had FOO_{standard library function} and BAR_{standard library function}, and also just used the unadorned library function. The FOO_ variant had some special behaviour, the BAR_ version was sometimes the standard library version and sometimes their own (depending on both the platform and the function - in some cases they always wrote their own even when there's an adequate - or even better - version shipped with the platform, in some cases they made a per-platform decision about it).
Almost right, except that you stick #define printf GOOD_printf at the end of the #ifdef block and then always use printf(), don't force everyone reading the code to work out that GOOD_printf() means printf().
Closed-source for-profit software is typically backed by a company that is willing to (but not always competent at) fire bad programmers whose contributions end up adding more work for others than solving problems.
The problem with this idea is that it assumes that it's easy to attribute costs to a particular programmer. The Heartbleed bug added a new feature (yay! Marketing checkbox!) and sat undiscovered for a couple of years. Do you think that the developer would have been fired immediately after introducing it in a closed-source environment?
You mention 256GB of ram, which nothing needs to compile, but I wonder, are you using FreeBSD's equivalent of tmpfs (ramdisk, I don't know what it's called) for storage of the build tree? I'm sure that would result in massive speedup over compiling from a on disk/ssd filesystem.
That's not quite true. For one thing, if your buffer cache is big enough that the entire source tree stays there then it has a noticeable impact on compile time. I tend to put the obj tree in tmpfs when doing a universe build (I don't actually want it, I just want to check everything builds most of the time), and that consumes 50GB. You might also be surprised at how much RAM the compile takes.
Bullshit, Quartus II has had Multiprocessor support since at least version 7.2 (I can't find the exact release they added it), and has been improving P&R speeds every release since then. In the version I am using, 14.1 Web Edition, Multiprocessor support requires Talkback (give back performance metrics in exchange for parallel speedup) and works quite well. You should probably look at buying the subscription version and adding some design partitions if your synthesis is taking an hour and a half.
Hmm, is it something that you need to specifically enable? We're taking about an hour and a half with Quartus 13 on these machines. We do pay for the subscription version, although adding more design partitions is not a simple thing to do as our Verilog is generated from a higher-level HDL.
Can you define what you mean by 'a lot' of compiling? We have some machines where I regularly build LLVM and the FreeBSD base system. They're 32-core with 256GB of RAM, and on them most builds are nice and fast, although it still takes close to two hours to do a universe build (all kernel configs, userland for all supported architectures - the 'quick' sanity test that you haven't broken anything). They're similar specs to the 'beefy' machines in the FreeBSD cluster, which take around 24 hours to build a complete package set (all 24K ports for a single release / architecture).
I'm not sure what you got 6 years ago, but there are still things where the fastest you can buy it not as fast as you'd like. I'm really interested in better single-threaded performance, because some companies (*cough*Altera*cough*Xylinx*cough*) haven't yet worked out how to multithread their compilers and so it takes an hour and a half to do a synthesis run of our CPU.
As I said in my original post, the thing that killed it was their time to market. It was initially proposed as a device where you'd pay a slight premium, but get something with an open software stack and a nice form factor as a hacker toy. By the time it shipped, that premium had become 200-300% over comparable devices.
512MB made sense in 2008. It was the largest available Package-on-Package (PoP) RAM module you could buy. The OMAP series of SoCs are usually used in a stacked configuration, with the CPU, RAM, and Flash all vertically stacked. This reduces the complexity of the board a lot, because you don't need any traces to be run for RAM or Flash (only for external peripherals), but it also means that you're quite restricted in the memory chips you can use: they must be single-chip modules and must have a sufficiently low power and thermal envelope that they can sit happily on top of the SoC. 1GB PoP modules were released a few months after the Pandora team finished the design, but it was too late to replace it by then and they've never done a revision B.
I was really interested in the Pandora when it was announced, but eventually got a phone with similar specs for about £50, before most people got their Pandoras. I've now also got an Asus TransformerPad TF700, which is a very nice machine, but I'd really like something that could run FreeBSD out of the box. The Dragonbox Pyra specs look pretty good, but I note a lack of any physical dimensions on the web site (other than a 5" screen), which makes it a bit difficult to evaluate. I didn't realise hns was involved - hopefully this will go a bit better than the GTA handsets...
TI is notoriously bad at documenting their non-CPU cores. The c64x still needs a proprietary compiler (which supports only a weird dialect of C). Is there some commitment from them to open specs for the OMAP5 SoC? One of the nice things about the ChromeBook is that Google has the purchasing power to be able to demand this from their suppliers. I don't think the Pyra has the same influence, so may need to shop around a bit more...
Why do you think that contradicts his statement? Take a look at how much funding GCHQ receives from the NSA - it's a significant amount of their total budget and has led to some concerns that they act more in the interests of the US than UK.
The big problem with OpenPandora was their time to market. When they started (2008), the Cortex A8 was pretty new (released 2007, but not much silicon until 2008) and there were few devices you could get with one in, although a few evaluation boards were starting to appear. They promised something in a small mobile form factor running a completely open software stack, which sounded like a fun platform. By the time they actually shipped anything (2010), the Cortex A8 was starting to feel a bit dated and was available in cheap mobile phones running Android. By the time they shipped to more than a token number of people, the A8 was ancient and you could get a dual or quad-core A9 or similar for half the price.
Oh, and their UK operation effectively went bust after taking huge numbers of pre-orders. Both the UK and German companies had the pre-order money in their accounts for 2+ years earning interest before they shipped anything, but at least people who ordered things from the German company eventually got something, even if it was worth a fraction of what they paid for it, as a result of being two years - over an entire generation of the technology - old by the time they got it.
The even larger problem, as I see it, is that being hired because of a quota is the ultimate stigma: "Look at her, she only got this job because of her tits." No-one takes the quota employee seriously - even when they actually are the best.
This is a very serious problem and has actually led to an increase in biases in some parts of the world. If 90% of a population of doctors of a particular caste/ethnicity/gender are hired based on quotas, then there's a good chance that when you see a doctor who meets those criteria that they will be less competent than someone who doesn't. You then start assuming that everyone in that category is less competent, even if you're unaware of the quota. For the 10% who deserve to be there, this makes life difficult because no one knows that they are the competent ones. This leads to fewer competent applicants from the minority group, because they know that they'll be stigmatised even if they're at the very top of the profession. And how do you address the lack of applicants? More quotas!
You can't fix inequality at the hiring end, you need to find out why you aren't getting enough competent applicants from a minority group and work backwards. Is it because they aren't entering the right degree programs? Is that because of bias in admissions, problems with earlier education, cultural bias, or some other factors?