Web sites should put a message rectangle - like the current idiotic EU cookie banner -, that says:
"You are licensed to continue using this web site, if either you view ads and do not use any ad blocker, or if you pay our modest subscription fee, which is 2 cent/day."
It is even better if it is displayed by the advertising agency, so if a somebody regularly use an ad blocker and do not pay on many sites, than he accumulates enough dept that it worths pursuing him by some debt collection agencies.
it depends on how the scripting engine based on the back end works, and if it's injectable. I've seen code snippets for Java code handed back to the back end, with the code handed back being in the front end web page. An attack on that, even for a serialized object, is as simple as writing a transcoding proxy to substitute your serialized objects for the intended serialized objects, thereby compromising the back end JVM. I'd also point out that a shop running Java back ends is much more likely to also run Java front ends, and depend on the security model.
I develop web applications in Java as a full time job for about 12 years, but I have no idea what are you talking about. Serialization deals with object data, it does not save or retrieve code. And what kind of scripting engine do you talk about? There are scripting engines written in Java, but Java is not a scripting engine. Of course everybody can write unsafe code in any language, but Java makes it hard. I never heard a single high profile exploited vulnerability in Java server side applications. There were serious vulnerabilities recently in the Java browser plugin and in sandboxing of downloaded applet code. It is quite sad that Oracle does not put more effort into securing the plugin. However, these issues has absolutely nothing to do with web applications, such as a forum.
Whether your metric is "performance per transistor" or "performance per mm^2 of chip area" or "performance per watt"
On the other hand AMD is quite good in "Performance per dollar".
Steamroller apparently only shares the MMX/FMAC unit, while standard floating point operations are per-core.
Uh, no. Steamroller retains the shared-FPU arrangement. In fact, you effectively said as much: "FMAC" operations are standard floating point operations. It doesn't get much more fundamental than multiplication and addition.
Standard floating point is already per-core even in Bulldozer. The shared 256 bit floating point unit can be split into two 128 bit unit. Standard floating point is less than 128 bits, so both cores perform these operations simultaneously.
The only case when one core of the unit waits for the other core to execute a standard floating point operation is when the other core is executing not a usual floating point operation, but a 256 bit AVX extended instruction. (MMX is only 128 bit).
The power at full load means a high load which can easily occur in real life.
The maximum power usage is exactly that, the wattage used in the theoretically worst case. It is practically impossible to reach the maximum power usage with any real life load.
the i3-2120 offers far better performance in the area's that count for business/enterprise users. The CPU has enough performance that even with the crappy Intel IGP, they still do what's needed quite well while offering a much lower TDP and even that's becoming important to everyone.
Do you talk about business desktops? They are idling all the time, you have to look at the idle power, not TDP.
Would I buy any current AMD processors for a server farm? Probably not.
The predecessors of this series, the Opteron 6200 is used in quite a few supercomputers. Actually I counted 21 Opteron bases systems in the last supercomputer top 100 list.
Although that is not on a driver level, but macro programmable keyboards do exist. The advantage is that you do not have install anything on any OS. The disadvantage is that keyboards are not really mobile.
... and get a Kinesis Advantage.. One of it was in my drawer for two years, when I finally brought it out, and to my biggest surprise I was able to touch type English letters within two minutes. It is ergonomic; narrow enough to reach the mouse easily; still your arms can keep a longer, natural distance; there are many thumb keys so there is no need for keys which are very far from your fingers; etc.
But its most important property, which is not even listed in its description, is that the keys are in a diagonal layout. The fingers move straight up and down, this means that finding the keys without looking at the keyboard is almost natural. After trying for ten years without any success, it took me only two days to get back my original (non touch typing) performance, and two months to get proficient in touch typing. And I can tell you that it is a huge difference.
I also used a keyboard with a layout similar to Microsoft Natural keyboard for several years. I do not know which Kinesis you referred to, but now I am using the Kinesis Advantage, and it was a pleasant surprise. The Kinesis was in a drawer for two years, because I was afraid that it took months to learn it. Quite the opposite, I was never able to learn touch typing with my old keyboards, but after I brought the Kinesis out I touch typed English letters within 10 minutes. It took me about 2 months to get proficient in touch typing, but I think I reached my original performance within a few days.
You just proved my point. You ran a test with the most ideal case (pure integer code) and it ran slower when on the same module.
No, no, this was not the ideal case. Actually this was one of the worst test case for Bulldozer. The test is a simple integer addition in a cycle. The emphasis is on the cycle, not on the integer addition!
If I run the same cycle but with 8 integer addition instead of only one, then the difference between running two threads in one Bulldozer module vs two Bulldozer modules is 47 s vs 53 s. That is only 13%.
If I run the cycle with one floating point addition then the differrence is 17.0 s vs 17.6 s. That is only 3%. And how ironic this is! The most frequent complaint against the Bulldozer architecture is that two cores share a single floating point unit. AMD should tell one million times that yes, they share a single floating point unit, but that is a 256 bit wide unit, which can be split into two 128 bit parts. And what is the size of the usual floating point number? Not 256 bit, not 128 bit, but only 64.
If AMD had been truthful then the FX CPUs would perform to their fullest potential on any OS, and Linux and Win 8 would not have had to be modified to work around the issue.
I think you underestimate the complexity of scheduling on modern processors. Consider the turbo mode, which boosts the frequency of a module, but cannot boost all modules at the same time, the new Opteron 6200 family which is essentially two processors with two memory controllers in a single socket (NUMA), and multiple sockets on most server motherboards. This is no longer SMP, and not a simple hyper-threading either.
The processor should inform the OS about its capabilities the most precisely. If a Bulldozer tells the OS that it is hyper-threaded, then I consider that a dirty hack. That is the same when Chrome, Opera, etc. says that it is Internet Explorer. It might be practical, because they wouldn't need to beg the Microsoft guys to update their scheduler in Windows 7. By the way, it is strange, that the Linux developers were able to update the kernel quickly, is not it?
While the 51% in the second application is indeed disappointing, it must be noted that we are comparing a $160 processor to a $290 processor. Even in this particular workload the performance / price of the overall AMD system was much better (but not the performance / watt), according to the person who posted the result. In the PassMark score this is even more obvious.
I have good experience with both Asus AMD and Asus Intel motherboards. I (let's say) supervised about 30 pc's over the years, 70% of them was AMD, and none of them failed early. There were about 4 motherboard failures, roughly proportional to the processor manufacturer, and all of them after several years of usage. They usually become obsolete before they have a chance to broke.
OK, I did a small benchmark. I run a test program with two threads. In each test run I assigned two cores to the program. Both threads add integers in a cycle, nothing else. The test program completes in 14 seconds if it is running on two cores of the same Bulldozer unit, while it takes 11 seconds on 1-1 cores of two different Bulldozer units.
AMD A6-5400K, for 70$, yes, it is not 50$, but it is also faster, it has far better graphics and it is 65W too.
I hope you also recognize, that office computers most of the time are idling. So if anything, the idle power is the interesting, not the TDP. And AMD is quite good in that.
A unit is is not a pair of full cores, but almost. It is not in the same leauge as hyperthreading. Almost all parts are duplicated, or has enough performance to serve both cores without slowdown. As I see there are only special situations, which cause the two cores to collide (like the issue with shared libraries on randomized addresses, occassionally causing frequent cache collisions - already solved in Linux kernel), but in 99% of the workloads the pairs do not slow down each other.
For example, contrary to the common oppinion, the two cores in the module can do floating point operations parallel, the only restriction is that they cannot do 256 bit operations. On the other hand if only one core in a pair does 128 bit operations it can grab the other half of the floating point hardware of the unit. I think it is a very clever design, but the OS and compilers do need time to adopt.
I am not sure how accurate these tests. Bulldozer is a new architecture, and the compilers must catch up (and Windows too, on the other hand the new Linux kernels are good). AMD measured a 20% floating point performance increase just by recompiling the same test program with the newest gcc. 70% when they they used their own optimized compiler.
You can get Phenom II with 65W. But there are more than a few AMD desktop processors with 18, 25, 35, 45 watts.
May I ask you something? When AMD lead the performance/watt charts by a huge margin, did you buy AMD processors?
By the way, TDP is misleading. I do not know what is the current situation but a few years ago Intel and AMD calculated TDP differently. The AMD 100W TDP meant much lower consumption than the Intel 100W. It is also misleading because it is a theoretical limit, which is quite hard to actually reach. I have a single processor 12 core Opteron server, processor rated at 115W. Plus 4 drives and 48G ram. Even if I stress test the processor with 100% load on all cores, I cannot reach more than 160W for the whole server. And that also includes the ~20% loss in the power supply.
Actually AMD CPUs are bad at HPC, due to terrible floating point processing.
More than 10 percent of the world's top supercomputers are based on Opterons. And you say that they are bad at HPC:) Not to mention that HPC does not equal with floating performance. Actually it is very far from it, moving the data between processors is the harder issue. Anyway, care to provide a citation about the "terrible floating point performance"? As far as I know AMD Opterons has much better multicore floating point performance / dollar than Intels.
And you get those wonderful bottlenecks when every once in a while you have to aggregate all those parallel computed values, which requires a single core...
No, this is simply wrong as a general statement. First, aggregation is likely insignificant relative to the computation (otherwise it would be useless to parallelize the computation), moreover if aggregation in itself is still a significant task then it may be parallelized too and performed on multiple levels by aggregating subsets first.
so it all slows down to the AMD's single core processing power, which is prett much as bad as an atom cpu
Single core performance becomes less and less important. If you have a huge computing task, it is still a huge task on a single core even if that single core is the best of the world and it has double throughput than a usual one. For a huge task you have to use a huge number of cores if you want results within reasonable times.
You always bought Intel processors* even when they were far inferior compared to AMD. Now stop that, and just buy AMD processors even if they may perform a bit lower in some measurements in benchmarks (and better in a few others). For example people does not like Bulldozer, because it is not much better on the desktop front than its predecessor, but it is a great server architecture, especially for virtualization and high performance computing. There are quite a few AMD systems in the top 500 supercomputer list. Maybe it will be good enough for you too... The last time I checked an AMD Opteron offered the best performance / price ratio, not an Intel.
Not doing so will bring back the years when processor technology stopped improving and the prices remained high.
(*) Except those who build their own computers, among them AMD is much stronger, there were years when AMD had more than 50 percent.
Last time I checked most traffic came from Netflix and YouTube. Btw. ads are hosted by the ad companies - at least for now, until there are only a few percent of visitors who block ads. If ad-blockers become a significant loss for web sites they will have to start mixing ads with their own content.
Is it just me who's thought it f'in hilarious to be on a friends computer hit a website and get porn based ads & pop-ups?:)
It is even more hilarious that you look pron on your friends' laptops. The average site explicitly disables - actually does not enable - pron ads. Pron ad is enabled only on other pron and similar, non-family-friendly sites.
Slashdot Mirror
For example the Google banner displayed for me here on Slashdot is about a cloud backup solution.
You do not need to block ads if you have some self-respect. You have to visit ad-free and payment free pages. Good luck for that.
Web sites should put a message rectangle - like the current idiotic EU cookie banner -, that says:
"You are licensed to continue using this web site, if either you view ads and do not use any ad blocker, or if you pay our modest subscription fee, which is 2 cent/day." It is even better if it is displayed by the advertising agency, so if a somebody regularly use an ad blocker and do not pay on many sites, than he accumulates enough dept that it worths pursuing him by some debt collection agencies.
it depends on how the scripting engine based on the back end works, and if it's injectable. I've seen code snippets for Java code handed back to the back end, with the code handed back being in the front end web page. An attack on that, even for a serialized object, is as simple as writing a transcoding proxy to substitute your serialized objects for the intended serialized objects, thereby compromising the back end JVM. I'd also point out that a shop running Java back ends is much more likely to also run Java front ends, and depend on the security model.
I develop web applications in Java as a full time job for about 12 years, but I have no idea what are you talking about. Serialization deals with object data, it does not save or retrieve code. And what kind of scripting engine do you talk about? There are scripting engines written in Java, but Java is not a scripting engine. Of course everybody can write unsafe code in any language, but Java makes it hard. I never heard a single high profile exploited vulnerability in Java server side applications. There were serious vulnerabilities recently in the Java browser plugin and in sandboxing of downloaded applet code. It is quite sad that Oracle does not put more effort into securing the plugin. However, these issues has absolutely nothing to do with web applications, such as a forum.
Whether your metric is "performance per transistor" or "performance per mm^2 of chip area" or "performance per watt"
On the other hand AMD is quite good in "Performance per dollar".
Steamroller apparently only shares the MMX/FMAC unit, while standard floating point operations are per-core.
Uh, no. Steamroller retains the shared-FPU arrangement. In fact, you effectively said as much: "FMAC" operations are standard floating point operations. It doesn't get much more fundamental than multiplication and addition.
Standard floating point is already per-core even in Bulldozer. The shared 256 bit floating point unit can be split into two 128 bit unit. Standard floating point is less than 128 bits, so both cores perform these operations simultaneously.
The only case when one core of the unit waits for the other core to execute a standard floating point operation is when the other core is executing not a usual floating point operation, but a 256 bit AVX extended instruction. (MMX is only 128 bit).
The power at full load means a high load which can easily occur in real life.
The maximum power usage is exactly that, the wattage used in the theoretically worst case. It is practically impossible to reach the maximum power usage with any real life load.
the i3-2120 offers far better performance in the area's that count for business/enterprise users. The CPU has enough performance that even with the crappy Intel IGP, they still do what's needed quite well while offering a much lower TDP and even that's becoming important to everyone.
Do you talk about business desktops? They are idling all the time, you have to look at the idle power, not TDP.
Would I buy any current AMD processors for a server farm? Probably not.
The predecessors of this series, the Opteron 6200 is used in quite a few supercomputers. Actually I counted 21 Opteron bases systems in the last supercomputer top 100 list.
Although that is not on a driver level, but macro programmable keyboards do exist. The advantage is that you do not have install anything on any OS. The disadvantage is that keyboards are not really mobile.
... and get a Kinesis Advantage.. One of it was in my drawer for two years, when I finally brought it out, and to my biggest surprise I was able to touch type English letters within two minutes. It is ergonomic; narrow enough to reach the mouse easily; still your arms can keep a longer, natural distance; there are many thumb keys so there is no need for keys which are very far from your fingers; etc.
But its most important property, which is not even listed in its description, is that the keys are in a diagonal layout. The fingers move straight up and down, this means that finding the keys without looking at the keyboard is almost natural. After trying for ten years without any success, it took me only two days to get back my original (non touch typing) performance, and two months to get proficient in touch typing. And I can tell you that it is a huge difference.
I also used a keyboard with a layout similar to Microsoft Natural keyboard for several years. I do not know which Kinesis you referred to, but now I am using the Kinesis Advantage, and it was a pleasant surprise. The Kinesis was in a drawer for two years, because I was afraid that it took months to learn it. Quite the opposite, I was never able to learn touch typing with my old keyboards, but after I brought the Kinesis out I touch typed English letters within 10 minutes. It took me about 2 months to get proficient in touch typing, but I think I reached my original performance within a few days.
I like the idea that the numbers are accessible via the SHIFT key, and in place of them there are frequently used symbols.
Colo means co-location, in which customers rent rack space, and they move their own hardware into the data center.
You just proved my point. You ran a test with the most ideal case (pure integer code) and it ran slower when on the same module.
No, no, this was not the ideal case. Actually this was one of the worst test case for Bulldozer. The test is a simple integer addition in a cycle. The emphasis is on the cycle, not on the integer addition!
If I run the same cycle but with 8 integer addition instead of only one, then the difference between running two threads in one Bulldozer module vs two Bulldozer modules is 47 s vs 53 s. That is only 13%.
If I run the cycle with one floating point addition then the differrence is 17.0 s vs 17.6 s. That is only 3%. And how ironic this is! The most frequent complaint against the Bulldozer architecture is that two cores share a single floating point unit. AMD should tell one million times that yes, they share a single floating point unit, but that is a 256 bit wide unit, which can be split into two 128 bit parts. And what is the size of the usual floating point number? Not 256 bit, not 128 bit, but only 64.
If AMD had been truthful then the FX CPUs would perform to their fullest potential on any OS, and Linux and Win 8 would not have had to be modified to work around the issue.
I think you underestimate the complexity of scheduling on modern processors. Consider the turbo mode, which boosts the frequency of a module, but cannot boost all modules at the same time, the new Opteron 6200 family which is essentially two processors with two memory controllers in a single socket (NUMA), and multiple sockets on most server motherboards. This is no longer SMP, and not a simple hyper-threading either.
The processor should inform the OS about its capabilities the most precisely. If a Bulldozer tells the OS that it is hyper-threaded, then I consider that a dirty hack. That is the same when Chrome, Opera, etc. says that it is Internet Explorer. It might be practical, because they wouldn't need to beg the Microsoft guys to update their scheduler in Windows 7. By the way, it is strange, that the Linux developers were able to update the kernel quickly, is not it?
While the 51% in the second application is indeed disappointing, it must be noted that we are comparing a $160 processor to a $290 processor. Even in this particular workload the performance / price of the overall AMD system was much better (but not the performance / watt), according to the person who posted the result. In the PassMark score this is even more obvious.
I have good experience with both Asus AMD and Asus Intel motherboards. I (let's say) supervised about 30 pc's over the years, 70% of them was AMD, and none of them failed early. There were about 4 motherboard failures, roughly proportional to the processor manufacturer, and all of them after several years of usage. They usually become obsolete before they have a chance to broke.
OK, I did a small benchmark. I run a test program with two threads. In each test run I assigned two cores to the program. Both threads add integers in a cycle, nothing else. The test program completes in 14 seconds if it is running on two cores of the same Bulldozer unit, while it takes 11 seconds on 1-1 cores of two different Bulldozer units.
AMD A6-5400K, for 70$, yes, it is not 50$, but it is also faster, it has far better graphics and it is 65W too.
I hope you also recognize, that office computers most of the time are idling. So if anything, the idle power is the interesting, not the TDP. And AMD is quite good in that.
A unit is is not a pair of full cores, but almost. It is not in the same leauge as hyperthreading. Almost all parts are duplicated, or has enough performance to serve both cores without slowdown. As I see there are only special situations, which cause the two cores to collide (like the issue with shared libraries on randomized addresses, occassionally causing frequent cache collisions - already solved in Linux kernel), but in 99% of the workloads the pairs do not slow down each other.
For example, contrary to the common oppinion, the two cores in the module can do floating point operations parallel, the only restriction is that they cannot do 256 bit operations. On the other hand if only one core in a pair does 128 bit operations it can grab the other half of the floating point hardware of the unit. I think it is a very clever design, but the OS and compilers do need time to adopt.
See this article for more information: AMD Bulldozer - What's a Module, what's a Core?
I am not sure how accurate these tests. Bulldozer is a new architecture, and the compilers must catch up (and Windows too, on the other hand the new Linux kernels are good). AMD measured a 20% floating point performance increase just by recompiling the same test program with the newest gcc. 70% when they they used their own optimized compiler.
You can get Phenom II with 65W. But there are more than a few AMD desktop processors with 18, 25, 35, 45 watts.
May I ask you something? When AMD lead the performance/watt charts by a huge margin, did you buy AMD processors?
By the way, TDP is misleading. I do not know what is the current situation but a few years ago Intel and AMD calculated TDP differently. The AMD 100W TDP meant much lower consumption than the Intel 100W. It is also misleading because it is a theoretical limit, which is quite hard to actually reach. I have a single processor 12 core Opteron server, processor rated at 115W. Plus 4 drives and 48G ram. Even if I stress test the processor with 100% load on all cores, I cannot reach more than 160W for the whole server. And that also includes the ~20% loss in the power supply.
Actually AMD CPUs are bad at HPC, due to terrible floating point processing.
More than 10 percent of the world's top supercomputers are based on Opterons. And you say that they are bad at HPC :) Not to mention that HPC does not equal with floating performance. Actually it is very far from it, moving the data between processors is the harder issue. Anyway, care to provide a citation about the "terrible floating point performance"? As far as I know AMD Opterons has much better multicore floating point performance / dollar than Intels.
And you get those wonderful bottlenecks when every once in a while you have to aggregate all those parallel computed values, which requires a single core...
No, this is simply wrong as a general statement. First, aggregation is likely insignificant relative to the computation (otherwise it would be useless to parallelize the computation), moreover if aggregation in itself is still a significant task then it may be parallelized too and performed on multiple levels by aggregating subsets first.
so it all slows down to the AMD's single core processing power, which is prett much as bad as an atom cpu
Single core performance becomes less and less important. If you have a huge computing task, it is still a huge task on a single core even if that single core is the best of the world and it has double throughput than a usual one. For a huge task you have to use a huge number of cores if you want results within reasonable times.
You always bought Intel processors* even when they were far inferior compared to AMD. Now stop that, and just buy AMD processors even if they may perform a bit lower in some measurements in benchmarks (and better in a few others). For example people does not like Bulldozer, because it is not much better on the desktop front than its predecessor, but it is a great server architecture, especially for virtualization and high performance computing. There are quite a few AMD systems in the top 500 supercomputer list. Maybe it will be good enough for you too... The last time I checked an AMD Opteron offered the best performance / price ratio, not an Intel.
Not doing so will bring back the years when processor technology stopped improving and the prices remained high.
(*) Except those who build their own computers, among them AMD is much stronger, there were years when AMD had more than 50 percent.
Last time I checked most traffic came from Netflix and YouTube. Btw. ads are hosted by the ad companies - at least for now, until there are only a few percent of visitors who block ads. If ad-blockers become a significant loss for web sites they will have to start mixing ads with their own content.
Is it just me who's thought it f'in hilarious to be on a friends computer hit a website and get porn based ads & pop-ups? :)
It is even more hilarious that you look pron on your friends' laptops. The average site explicitly disables - actually does not enable - pron ads. Pron ad is enabled only on other pron and similar, non-family-friendly sites.