News on China's scientific and technological development.

SamuraiBlue said:

There is a gimick in the fastest claim. The Sunway TaihuLight CPUs doesn't communicate with one another as fast as the Japan's K supercomputer. That is the reason why K is at top of

Please, Log in or Register to view URLs content!

ratings with less CPU cores while the Sunway TaihuLight that has 15 time more CPU can't out do K in Graph500.
It has to do as wiki explains,

Please, Log in or Register to view URLs content!

in which K is able to move more data at one time through parallel computing. This is required for real time weather prediction and/or process big data collected through all the sensors that is accumulating at an alarming rate.
At the end it all boils down to the architecture of the super computer's design.

Click to expand...

There is a reason that the LINPACK benchmark for evaluating supercomputers is considered the gold standard and the Graph500 is not. K-computer might communicate with other computers faster but it's going to need to work in parallel with about a dozen more K-computers in perfect communication to match the performance of a Taihulight. Basically, you just made a statement along the lines of "Your car might be faster, smoother, nimbler, and more reliable, but my car can call a tow at the touch of a button so don't think you've got the best machine around."

To be honest, Sam, your statement at first sounded insightful to me, since I'm not a computer expert and only get my knowledge from news articles, but because of your history of making claims that don't stand up to the challenge of experts (you still owe me answer as to how genetically-modified cancer-killing T-cells can be turned into a bio-weapon in a terrorist's kitchen, so do address that if you have the time), I found this statement of yours to be very likely full of hot air as well.

manqiangrexue said:

There is a reason that the LINPACK benchmark for evaluating supercomputers is considered the gold standard and the Graph500 is not. K-computer might communicate with other computers faster but it's going to need to work in parallel with about a dozen more K-computers in perfect communication to match the performance of a Taihulight. Basically, you just made a statement along the lines of "Your car might be faster, smoother, nimbler, and more reliable, but my car can call a tow at the touch of a button so don't think you've got the best machine around."

To be honest, Sam, your statement at first sounded insightful to me, since I'm not a computer expert and only get my knowledge from news articles, but because of your history of making claims that don't stand up to the challenge of experts (you still owe me answer as to how genetically-modified cancer-killing T-cells can be turned into a bio-weapon in a terrorist's kitchen, so do address that if you have the time), I found this statement of yours to be very likely full of hot air as well.

Click to expand...

It's not about communicating with other computers, it's about each CPUs within the same computer communicating with one another.
That is what Graph500 rating is about piling in various odd data into different equations to obtain information. The weather is the most simple example.
At each monitoring station you gather abstract data like humidity wind velocity,wind direction, temperature, sun radiation amount,etc. you punch in all those data from thousands of monitoring stations into the super computer to predict the weather pattern but all data correlate in a different way so they have to be calculated in different equations where then the amount will be used as a variable in another equation and so on and then fed back to the original equation to equate time lapse to see the big picture in how the weather will change within the next hour.
In doing so parallel computing is essential or it will take more then a hour to calculate making the information itself meaningless. Basically it's trying to calculate what will the out come be through the butterfly effect.

SamuraiBlue said:

There is a gimick in the fastest claim. The Sunway TaihuLight CPUs doesn't communicate with one another as fast as the Japan's K supercomputer. That is the reason why K is at top of

Please, Log in or Register to view URLs content!

ratings with less CPU cores while the Sunway TaihuLight that has 15 time more CPU can't out do K in Graph500.
It has to do as wiki explains,

Please, Log in or Register to view URLs content!

in which K is able to move more data at one time through parallel computing. This is required for real time weather prediction and/or process big data collected through all the sensors that is accumulating at an alarming rate.

Click to expand...

I can also call the K setup a gimmick.

SamuraiBlue said:

At the end it all boils down to the architecture of the super computer's design.

Click to expand...

Exactly. That is what enables the TaihuLight to take the top spot.

SamuraiBlue said:

It's not about communicating with other computers, it's about each CPUs within the same computer communicating with one another.
That is what Graph500 rating is about piling in various odd data into different equations to obtain information. The weather is the most simple example.
At each monitoring station you gather abstract data like humidity wind velocity,wind direction, temperature, sun radiation amount,etc. you punch in all those data from thousands of monitoring stations into the super computer to predict the weather pattern but all data correlate in a different way so they have to be calculated in different equations where then the amount will be used as a variable in another equation and so on and then fed back to the original equation to equate time lapse to see the big picture in how the weather will change within the next hour.
In doing so parallel computing is essential or it will take more then a hour to calculate making the information itself meaningless. Basically it's trying to calculate what will the out come be through the butterfly effect.

Click to expand...

Oh, good! So you do see my posts. Since you'd been ignoring my inquiries to explain yourself on other claims (such as mentioned in my last post), I thought perhaps it's because you somehow didn't see my posts. But you do... so you've been selectively ignoring me when you realize you're wrong instead of just admitting it?

SamuraiBlue said:

It's not about communicating with other computers, it's about each CPUs within the same computer communicating with one another.
That is what Graph500 rating is about piling in various odd data into different equations to obtain information. The weather is the most simple example.
At each monitoring station you gather abstract data like humidity wind velocity,wind direction, temperature, sun radiation amount,etc. you punch in all those data from thousands of monitoring stations into the super computer to predict the weather pattern but all data correlate in a different way so they have to be calculated in different equations where then the amount will be used as a variable in another equation and so on and then fed back to the original equation to equate time lapse to see the big picture in how the weather will change within the next hour.
In doing so parallel computing is essential or it will take more then a hour to calculate making the information itself meaningless. Basically it's trying to calculate what will the out come be through the butterfly effect.

Click to expand...

Let's be frank here. The way I see it and the way I think everyone sees it, this entire discussion from you has basically been an indirect of you trying to say that the TOP500's linpack benchmarks for judging supercomputers is poor and thus the "best" supercomputer shouldn't be the Taihulight, and that the top computer shoulld really be K computer with all supercomputers measured instead using Graph500 rating.

I don't necessarily have a problem with such a suggestion, but you really need to try and come out with a strong argument for why you think graph500 is supposedly a better way of measuring supercomputers rather than linpack that top500 uses... and which is still considered the most authoritative measure of supercomputer performance in the world.


So I'm open for you to argue your case, but the position you're holding is quite a difficult one to defend and would require a lot more proof and evidence and much stronger premises than I think you would be able to produce -- because essentially you will have to be trying to disprove the entire basis of the last 20+ years of measuring how supercomputers compare to each other that is the norm for well... virtually all of humanity. That is quite a daunting task for you to try and defend.

OTOH, if that position isn't what you're trying to argue for, then frankly I'm a little bit confused as to what exactly the purpose of your last few posts are and what your point is.

Linpack benchmarks is basically an industrial benchmark seeing how fast a computer can go but figuratively only in a straight line like a drag race doing quarter mile sprint but in real life cars are rarely used that way.
In the past computers were used to calculate the trajectory of projectiles for this TaihuLight will undoubtedly be at top but in real life computers are seldomly used only in those purposes in which these days computers are used more to solve problems utilizing big data, like predicting choke points in traffic or predicting peak amount of electricity usage during the summer months,etc.
For those purposes K computer comes top.

You can argue all you want on which is more important since I am not really that much interested in that but stating Linpack is the only benchmark that is relevant is not constructive when you are talking about computer science.

SamuraiBlue said:

Linpack benchmarks is basically an industrial benchmark seeing how fast a computer can go but figuratively only in a straight line like a drag race doing quarter mile sprint but in real life cars are rarely used that way.
In the past computers were used to calculate the trajectory of projectiles for this TaihuLight will undoubtedly be at top but in real life computers are seldomly used only in those purposes in which these days computers are used more to solve problems utilizing big data, like predicting choke points in traffic or predicting peak amount of electricity usage during the summer months,etc.
For those purposes K computer comes top.

You can argue all you want on which is more important since I am not really that much interested in that but stating Linpack is the only benchmark that is relevant is not constructive when you are talking about computer science.

Click to expand...

Okay, so essentially you are saying you think that linpack and the entire top500 list is a gimmick, yes?
In fact that is exactly what you are saying in your original reply in post #2930.

I hope you can understand why such an ambitious suggestion might not be taken seriously without a much larger arsenal of premises to defend that argument.

Bltizo said:

Okay, so essentially you are saying you think that linpack and the entire top500 list is a gimmick, yes?
In fact that is exactly what you are saying in your original reply in post #2930.

I hope you can understand why such an ambitious suggestion might not be taken seriously without a much larger arsenal of premises to defend that argument.

Click to expand...

If you don't like the word gimmick then use industrial benchmark for all I care.
Basically the fastest doesn't mean the most useful as I have given various examples of.
You are just arguing for the sake of arguing which is a waste of time and bandwidth.

SamuraiBlue said:

If you don't like the word gimmick then use industrial benchmark for all I care.
Basically the fastest doesn't mean the most useful as I have given various examples of.
You are just arguing for the sake of arguing which is a waste of time and bandwidth.

Click to expand...

Right okay, so essentially you have abdicated your original overly ambitious position then?

And come on, don't be like that. You're the one who made the ridiculous claim to begin with, if other people call you out on it and want you to back it up then you don't really have grounds to say that they are the ones arguing for the sake of arguing.

Bltizo said:

Right okay, so essentially you have abdicated your original overly ambitious position then?

And come on, don't be like that. You're the one who made the ridiculous claim to begin with, if other people call you out on it and want you to back it up then you don't really have grounds to say that they are the ones arguing for the sake of arguing.

Click to expand...

Ambitious?
Oh, boy. Look, the fact is with computer science and super computers being utilized in more broader functions just measuring how fast it can compute in a straight line is becoming less and less important through the various examples I had elaborated.
Like a car race in which there is the quarter mile dash, the endurance race and sprint race in the circuit. Not any of the cars will win all form of races.
The TaihuLight claims the title for the quarter mile dash but the K computer at the moment is claiming the circuit sprint title, that is all.