The 30/30 Rule

[Callum]

So I was just reading an article on Autosport that mentioned the 30/30 rule for aerodynamic testing. The rule says that you're allowed 30 hours of wind tunnel time or 30 teraflops of computing data (or a combination of both) per week.

I've briefly delved into the black art of CFD but it was just doing simple flow problems on (good) desktop PCs. I was wondering how an hour of wind tunnel time really compared to a teraflop of CFD. They obviously both have their advantages and disadvantages (and are both not 'real') but is it an fair ratio?


Source link: http://www.autosport.com/news/report.php/id/113942

[tim|away]

FLOP = FLoating-point OPeration
FLOPS = FLoating-point Operations Per Second

1 TERAflop = 10^12 FLOP

30 terraflop would therefore mean 3* 10^13 floating point operations (per week) which quite frankly doesn't sound like a whole lot to me. I was just looking up how many floating point operatings per second (FLOPS) the high-end GPUs can get. Apparently, the Geforce Titan series can get anywhere from 4.5 up to over 8 teraflops.

I'm not sure if I'm missing something here, but it appears as if an off-the-shelf GPU could in theory calculate the weekly allowance of 30 teraFLOP in under 4 seconds.

[rjsa]

Code: Select all

2.6 The calculation used for the declaration of the 8 week Aerodynamic Testing Period shall be carried out as below. 
TotFLOPS = (MFPPC * CCF * NCU * NSS ) / (604,800 * 8 * 1000) 
Where : 
TotFLOPS = The total number of TeraFLOPs used per CFD solve run. 
MFPPC = Peak double precision floating point operations per cycle per core of the Processing Unit (excluding AVX if declared under 2.5(d) or using natural precision operations under 2.5(e) if the core is not double precision capable). 
CCF = Peak Processing Unit clock frequency in GigaHertz achieved during the CFD solver run. This will be the peak frequency theoretically achievable during the run based on one of the following : 
a) The standard clock frequency value from the Processing Unit Manufacturer's specification sheet (if overclocking or enhanced modes are not used in the run). 
b) The maximum “turbo”, “HPC” or other enhanced mode frequency value. 
c) The maximum overclocked frequency value. 
NCU = Number of Processing Unit cores used for the run. 
NSS = Number of solver wall clock seconds elapsed during the run. NB Message passing time during calculation should be included. 
All information required for auditing should be present in the output from the run including the CCF value. 
For the avoidance of doubt, any offload processing for example FPU, FPGA, GPU/GPGPU, VFP, softfp etc. should be included and calculated using the same method as above. 

Code: Select all

3.6 The Limit Line is defined as follows : 
WT <= WT_limit (1 – CFD/CFD_limit) 
Where : 
WT = Wind On Time 
WT_limit = 30 hours 
CFD = CFD Teraflops Usage 
CFD_limit = 30 Teraflops 

Your standard TeraFlop it seems, being the constants:

604,800 Seconds in a week
8 Weeks period
1000 Gigs in a Tera

[tim|away]

Thank you for that, rjsa.

It should therefore be noted that we are not dealing with a limit of 30 TeraFlop (floating-point operations), but 30 TeraFlops (floating point operations per second). A lot of journalists even seem to get this wrong and confuse things. Having said that, the teams may run their simulations 24/7 as long as the computing performance doesn't exceed 30 terra floating-point operations per second on average.

They have the choice to throw even more FLOPS at their CFD by trading windtunnel time for additional FLOPS. Either way, the CFD machines almost certainly run 24/7.

[Callum]

Thanks guys, that's pretty interesting.

So we can basically say that they could do ten simulations with 3 TeraFlops per second each... or one simulation consuming all 30 TeraFlops per second. I'd be interested to know what the teams have found most beneficial.

[turbof1]

Anyone an idea how the fia monitors that? This is intern cuisine; how is having a hold on the 30 terraflops over 11 teams with some of having departments in different countries?

@Callum: no no! Terraflops is a measurement of speed. Teams can use cfd at a speed of 30 terraflops. So they can either do 10 operations AT THE SAME MOMENT at 3 terraflops or 1 operation of 30 terraflops. The total amount of operations during 1 week is actually unlimited.

[beelsebob]

The funny thing about the phrasing above is that 30TFLOPS per week means your computing power can actually accelerate. If you use 30TFLOPS in the first week, you can use a twice as powerful computer next week, at 60TFLOPS

[beelsebob]

Thanks guys, that's pretty interesting.

So we can basically say that they could do ten simulations with 3 TeraFlops per second each... or one simulation consuming all 30 TeraFlops per second. I'd be interested to know what the teams have found most beneficial.

No, it means that your computer can at most run at 30TFLOPS, continuously. A FLOPS is a unit of measurement of computer speed.

Re the GeForce Titan. It is only capable of 1.3 double precision TFLOPS, note, double precision, not single precision. That means you would need 23 GeForce titans running continuously to hit the allowed computational power.

[Kiril Varbanov]

Anyone an idea how the fia monitors that? This is intern cuisine; how is having a hold on the 30 terraflops over 11 teams with some of having departments in different countries?

There's onsite monitoring from FIA technicians sometimes, but most of these are based on a gentlemen agreements after live technical directives meetings where things are explained in detail. In reality, if you want to cheat, that's fairly easy. Same goes for wind tunnel hours.

[beelsebob]

Also, to bring the "it's only 23 graphics cards" down to earth with a bit more of a bump. Graphics cards are very special purpose hardware. They are great at multiplying matrices together, and even better at scanning across triangles generating fragments, but they are terrible at general purpose work. CFD can be poorly approximated on them, but can't be done well. Because of that, the teams will actually be using CPUs, not GPUs. A Core i7 4770 (or it's Xeon equivalent) will only achieve 64 GFLOPS across all 4 cores, so to hit the limits with those, you would need a 468 CPU, or 1872 core computer.

[bhall]

The funny thing about the phrasing above is that 30TFLOPS per week means your computing power can actually accelerate. If you use 30TFLOPS in the first week, you can use a twice as powerful computer next week, at 60TFLOPS

How does that work? I know f-all about computers.

[Kiril Varbanov]

Also, to bring the "it's only 23 graphics cards" down to earth with a bit more of a bump. Graphics cards are very special purpose hardware. They are great at multiplying matrices together, and even better at scanning across triangles generating fragments, but they are terrible at general purpose work. CFD can be poorly approximated on them, but can't be done well. Because of that, the teams will actually be using CPUs, not GPUs. A Core i7 4770 (or it's Xeon equivalent) will only achieve 64 GFLOPS across all 4 cores, so to hit the limits with those, you would need a 468 CPU, or 1872 core computer.

Sauber are the team that has been most generous in revealing details (2009):

Now, BMW Sauber F1 Team has launched the next step by extending the existing system. A further 384 nodes, equipped with Intel Xeon E5472 quadcore processors (four cores per processor) and related Intel technology where added to the existing system so that the new supercomputer, Albert3, now has 4224 cores. The main memory grew to 8448 GBytes and the peak compute power is now at 57,7 TFlops, that's 50.700.000.000.000 arithmetic operations per second. In order to achieve the same compute power, all residents of the two cities Munich and Berlin (4.7 million) would have to multiply two eight digit numbers every three seconds during one whole year. The new performance corresponds to rank 45 on the worldwide top500 list of currently running systems and rank 3 of all systems in private industry.

The new supercomputer which, like its predecessor, was developed by DALCO and which runs CFD-software from Ansys-Fluent, has a total weight of 38 tons and covers an area of only 24 square meters.

[beelsebob]

The funny thing about the phrasing above is that 30TFLOPS per week means your computing power can actually accelerate. If you use 30TFLOPS in the first week, you can use a twice as powerful computer next week, at 60TFLOPS

How does that work? I know f-all about computers.

It doesn't - it was just incorrectly phrased above. 30 trillion floating point operations per second per week was not what they intended to say.

[bhall]

Gotcha. TFLOPs, not TFLOPS.

[beelsebob]

Gotcha. TFLOPs, not TFLOPS.

No, TFLOPS is what matters – but it's TFLOPS, not TFLOPS per week.

That is, they may own a computer capable of performing 30TFLOPS, and use it as much or as little as they like, but they're obviously dumb if they don't use it as much as they can.

Alternatively, they can use the wind tunnel for 30 hours each week.

If they choose to use the wind tunnel for 10 hours each week, they can only own a computer that's capable of performing 20 TFLOPS, etc.